teammrag-parser-moreai / mdr_pdf_parser.py
rodrigomasini's picture
Update mdr_pdf_parser.py
4d042fb verified
# -*- coding: utf-8 -*-
# /=====================================================================\ #
# | MagicDataReadiness - MAGIC PDF Parser | #
# |---------------------------------------------------------------------| #
# | Description: | #
# | Extracts structured content (text, tables, figures, formulas) | #
# | from PDF documents using layout analysis and OCR. | #
# | Combines logic from various internal components. | #
# |---------------------------------------------------------------------| #
# | Dependencies: | #
# | - Python 3.11+ | #
# | - External Libraries (See imports below and installation notes) | #
# | - Pre-trained CV Models (Downloaded automatically to model dir) | #
# |---------------------------------------------------------------------| #
# | Usage: | #
# | See the __main__ block at the end of the script for an example. | #
# \=====================================================================/ #
# --- External Library Imports ---
import os
import re
import io
import copy
import fitz # PyMuPDF
from fitz import Document as FitzDocument, Page as FitzPage, Matrix as FitzMatrix
import numpy as np
import cv2 # OpenCV
import requests # For downloading models
from pathlib import Path
from enum import auto, Enum
from dataclasses import dataclass, field
from typing import Iterable, Generator, Sequence, Callable, TypeAlias, List, Dict, Any, Optional
from typing import Literal
from collections import defaultdict
from math import pi, ceil, sin, cos, sqrt, atan2
from PIL.Image import Image, frombytes, new as new_image, Resampling as PILResampling, Transform as PILTransform, fromarray as pil_fromarray
from PIL.ImageOps import expand as pil_expand
from PIL import ImageDraw
from PIL.ImageFont import load_default, FreeTypeFont
from shapely.geometry import Polygon
import pyclipper
from unicodedata import category
from alphabet_detector import AlphabetDetector
from munch import Munch
from transformers import LayoutLMv3ForTokenClassification
import onnxruntime
# --- HUGGING FACE HUB IMPORT ONLY BECAUSE RUNNING IN SPACES NOT NECESSARY IN PROD ---
from huggingface_hub import hf_hub_download
from huggingface_hub.errors import HfHubHTTPError
import time # Added for example usage timing
# --- External Dependencies ---
try:
from doclayout_yolo import YOLOv10
except ImportError:
print("Warning: Could not import YOLOv10 from doclayout_yolo. Layout detection will fail.")
YOLOv10 = None
try:
from pix2tex.cli import LatexOCR
except ImportError:
print("Warning: Could not import LatexOCR from pix2tex.cli. LaTeX extraction will fail.")
LatexOCR = None
try:
pass # from struct_eqtable import build_model # Keep commented as per original
except ImportError:
print("Warning: Could not import build_model from struct_eqtable. Table parsing might fail.")
import torch
if not hasattr(torch, "get_default_device"):
torch.get_default_device = lambda: torch.device("cuda" if torch.cuda.is_available() else "cpu")
# --- MagicDataReadiness Core Components ---
# --- MDR Utilities ---
def mdr_download_model(url: str, file_path: Path):
"""Downloads a model file from a URL to a local path."""
try:
response = requests.get(url, stream=True, timeout=120) # Increased timeout
response.raise_for_status()
file_path.parent.mkdir(parents=True, exist_ok=True)
with open(file_path, "wb") as file:
for chunk in response.iter_content(chunk_size=8192):
file.write(chunk)
print(f"Successfully downloaded {file_path.name}")
except requests.exceptions.RequestException as e:
print(f"ERROR: Failed to download {url}: {e}")
if file_path.exists(): os.remove(file_path)
raise FileNotFoundError(f"Failed to download model from {url}") from e
except Exception as e:
print(f"ERROR: Failed writing file {file_path}: {e}")
if file_path.exists(): os.remove(file_path)
raise e
def mdr_ensure_directory(path: str) -> str:
"""Ensures a directory exists, creating it if necessary."""
path = os.path.abspath(path)
os.makedirs(path, exist_ok=True)
return path
def mdr_is_whitespace(text: str) -> bool:
"""Checks if a string contains only whitespace."""
return bool(re.match(r"^\s*$", text)) if text else True
def mdr_expand_image(image: Image, percent: float) -> Image:
"""Expands an image with a white border."""
if percent <= 0: return image.copy()
w, h = image.size
bw, bh = ceil(w * percent), ceil(h * percent)
fill: tuple[int, ...] | int
if image.mode == "RGBA":
fill = (255, 255, 255, 255)
elif image.mode in ("LA", "L"):
fill = 255
else:
fill = (255, 255, 255)
return pil_expand(image=image, border=(bw, bh), fill=fill)
# --- MDR Geometry ---
MDRPoint: TypeAlias = tuple[float, float]
@dataclass
class MDRRectangle:
"""Represents a geometric rectangle defined by four corner points."""
lt: MDRPoint;
rt: MDRPoint;
lb: MDRPoint;
rb: MDRPoint
def __iter__(self) -> Generator[MDRPoint, None, None]:
yield self.lt; yield self.lb; yield self.rb; yield self.rt
@property
def is_valid(self) -> bool:
try:
return Polygon(self).is_valid
except:
return False
@property
def segments(self) -> Generator[tuple[MDRPoint, MDRPoint], None, None]:
yield (self.lt, self.lb); yield (self.lb, self.rb); yield (self.rb, self.rt); yield (self.rt, self.lt)
@property
def area(self) -> float:
try:
return Polygon(self).area
except:
return 0.0
@property
def size(self) -> tuple[float, float]:
widths, heights = [], []
for i, (p1, p2) in enumerate(self.segments):
dx, dy = p1[0] - p2[0], p1[1] - p2[1]
dist = sqrt(dx * dx + dy * dy)
if i % 2 == 0:
heights.append(dist)
else:
widths.append(dist)
avg_w = sum(widths) / len(widths) if widths else 0.0
avg_h = sum(heights) / len(heights) if heights else 0.0
return avg_w, avg_h
@property
def wrapper(self) -> tuple[float, float, float, float]:
x1, y1, x2, y2 = float("inf"), float("inf"), float("-inf"), float("-inf")
for x, y in self:
x1, y1, x2, y2 = min(x1, x), min(y1, y), max(x2, x), max(y2, y)
return x1, y1, x2, y2
def mdr_intersection_area(rect1: MDRRectangle, rect2: MDRRectangle) -> float:
"""Calculates intersection area between two MDRRectangles."""
try:
p1 = Polygon(rect1)
p2 = Polygon(rect2)
if not p1.is_valid or not p2.is_valid:
return 0.0
return p1.intersection(p2).area
except:
return 0.0
# --- MDR Data Structures ---
@dataclass
class MDROcrFragment:
"""Represents a fragment of text identified by OCR."""
order: int;
text: str;
rank: float;
rect: MDRRectangle
class MDRLayoutClass(Enum):
"""Enumeration of different layout types identified."""
TITLE = 0;
PLAIN_TEXT = 1;
ABANDON = 2;
FIGURE = 3;
FIGURE_CAPTION = 4;
TABLE = 5;
TABLE_CAPTION = 6;
TABLE_FOOTNOTE = 7;
ISOLATE_FORMULA = 8;
FORMULA_CAPTION = 9
class MDRTableLayoutParsedFormat(Enum):
"""Enumeration for formats of parsed table content."""
LATEX = auto();
MARKDOWN = auto();
HTML = auto()
@dataclass(eq=False)
class MDRBaseLayoutElement:
"""Base class for layout elements found on a page."""
rect: MDRRectangle;
fragments: list[MDROcrFragment]
def __eq__(self, other):
return self is other
def __hash__(self):
return id(self)
@dataclass
class MDRPlainLayoutElement(MDRBaseLayoutElement):
"""Layout element for plain text, titles, captions, figures, etc."""
# MODIFIED: Replaced Literal[...] with the Enum class name
cls: MDRLayoutClass # The type hint is now the Enum class itself
@dataclass
class MDRTableLayoutElement(MDRBaseLayoutElement):
"""Layout element specifically for tables."""
parsed: tuple[str, MDRTableLayoutParsedFormat] | None
# MODIFIED: Replaced Literal[EnumMember] with the Enum class name
cls: MDRLayoutClass = MDRLayoutClass.TABLE # Hint with Enum, assign default member
@dataclass
class MDRFormulaLayoutElement(MDRBaseLayoutElement):
"""Layout element specifically for formulas."""
latex: str | None
# MODIFIED: Replaced Literal[EnumMember] with the Enum class name
cls: MDRLayoutClass = MDRLayoutClass.ISOLATE_FORMULA # Hint with Enum, assign default member
MDRLayoutElement = MDRPlainLayoutElement | MDRTableLayoutElement | MDRFormulaLayoutElement # Type alias
@dataclass
class MDRExtractionResult:
"""Holds the complete result of extracting from a single page image."""
rotation: float;
layouts: list[MDRLayoutElement];
extracted_image: Image;
adjusted_image: Image | None
# --- MDR Data Structures ---
MDRProgressReportCallback: TypeAlias = Callable[[int, int], None]
class MDROcrLevel(Enum): Once = auto(); OncePerLayout = auto()
class MDRExtractedTableFormat(Enum): LATEX = auto(); MARKDOWN = auto(); HTML = auto(); DISABLE = auto()
class MDRTextKind(Enum): TITLE = 0; PLAIN_TEXT = 1; ABANDON = 2
@dataclass
class MDRTextSpan:
"""Represents a span of text content within a block."""
content: str;
rank: float;
rect: MDRRectangle
@dataclass
class MDRBasicBlock:
"""Base class for structured blocks extracted from the document."""
rect: MDRRectangle
texts: list[MDRTextSpan]
font_size: float # Relative font size (0-1)
@dataclass
class MDRTextBlock(MDRBasicBlock):
"""A structured block containing text content."""
kind: MDRTextKind
has_paragraph_indentation: bool = False
last_line_touch_end: bool = False
class MDRTableFormat(Enum):
LATEX = auto()
MARKDOWN = auto()
HTML = auto()
UNRECOGNIZABLE = auto()
@dataclass
class MDRTableBlock(MDRBasicBlock):
"""A structured block representing a table."""
content: str
format: MDRTableFormat
image: Image # Image clip of the table
@dataclass
class MDRFormulaBlock(MDRBasicBlock):
"""A structured block representing a formula."""
content: str | None
image: Image # Image clip of the formula
@dataclass
class MDRFigureBlock(MDRBasicBlock):
"""A structured block representing a figure/image."""
image: Image # Image clip of the figure
MDRAssetBlock = MDRTableBlock | MDRFormulaBlock | MDRFigureBlock # Type alias
MDRStructuredBlock = MDRTextBlock | MDRAssetBlock # Type alias
# --- MDR Utilities ---
def mdr_similarity_ratio(v1: float, v2: float) -> float:
"""Calculates the ratio of the smaller value to the larger value (0-1)."""
if v1 == 0 and v2 == 0:
return 1.0
if v1 < 0 or v2 < 0:
return 0.0
v1, v2 = (v2, v1) if v1 > v2 else (v1, v2)
return 1.0 if v2 == 0 else v1 / v2
def mdr_intersection_bounds_size(r1: MDRRectangle, r2: MDRRectangle) -> tuple[float, float]:
"""Calculates width/height of the intersection bounding box."""
try:
p1 = Polygon(r1)
p2 = Polygon(r2)
if not p1.is_valid or not p2.is_valid:
return 0.0, 0.0
inter = p1.intersection(p2)
if inter.is_empty:
return 0.0, 0.0
minx, miny, maxx, maxy = inter.bounds
return maxx - minx, maxy - miny
except:
return 0.0, 0.0
_MDR_CJKA_PATTERN = re.compile(r"[\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7a3\u0600-\u06ff]")
def mdr_contains_cjka(text: str):
"""Checks if text contains Chinese, Japanese, Korean, or Arabic chars."""
return bool(_MDR_CJKA_PATTERN.search(text)) if text else False
# --- MDR Text Processing ---
class _MDR_TokenPhase(Enum):
Init = 0
Letter = 1
Character = 2
Number = 3
Space = 4
_mdr_alphabet_detector = AlphabetDetector()
def _mdr_is_letter(char: str):
if not category(char).startswith("L"):
return False
try:
return _mdr_alphabet_detector.is_latin(char) or _mdr_alphabet_detector.is_cyrillic(
char) or _mdr_alphabet_detector.is_greek(char) or _mdr_alphabet_detector.is_hebrew(char)
except:
return False
def mdr_split_into_words(text: str):
"""Splits text into words, numbers, and individual non-alphanumeric chars."""
if not text: return
sp = re.compile(r"\s")
np = re.compile(r"\d")
nsp = re.compile(r"[\.,']")
buf = io.StringIO()
phase = _MDR_TokenPhase.Init
for char in text:
is_l = _mdr_is_letter(char)
is_d = np.match(char)
is_s = sp.match(char)
is_ns = nsp.match(char)
if is_l:
if phase in (_MDR_TokenPhase.Number, _MDR_TokenPhase.Character):
w = buf.getvalue()
yield w if w else None
buf = io.StringIO()
buf.write(char)
phase = _MDR_TokenPhase.Letter
elif is_d:
if phase in (_MDR_TokenPhase.Letter, _MDR_TokenPhase.Character):
w = buf.getvalue()
yield w if w else None
buf = io.StringIO()
buf.write(char)
phase = _MDR_TokenPhase.Number
elif phase == _MDR_TokenPhase.Number and is_ns:
buf.write(char)
else:
if phase in (_MDR_TokenPhase.Letter, _MDR_TokenPhase.Number):
w = buf.getvalue()
yield w if w else None
buf = io.StringIO()
if is_s:
phase = _MDR_TokenPhase.Space
else:
yield char
phase = _MDR_TokenPhase.Character
if phase in (_MDR_TokenPhase.Letter, _MDR_TokenPhase.Number):
w = buf.getvalue()
yield w if w else None
def mdr_check_text_similarity(t1: str, t2: str) -> tuple[float, int]:
"""Calculates word-based similarity between two texts."""
w1 = list(mdr_split_into_words(t1))
w2 = list(mdr_split_into_words(t2))
l1 = len(w1)
l2 = len(w2)
if l1 == 0 and l2 == 0:
return 1.0, 0
if l1 == 0 or l2 == 0:
return 0.0, max(l1, l2)
if l1 > l2:
w1, w2, l1, l2 = w2, w1, l2, l1
taken = [False] * l2
matches = 0
for word1 in w1:
for i, word2 in enumerate(w2):
if not taken[i] and word1 == word2:
taken[i] = True
matches += 1
break
mismatches = l2 - matches
return 1.0 - (mismatches / l2), l2
# --- MDR Geometry Processing ---
class MDRRotationAdjuster:
"""Adjusts point coordinates based on image rotation."""
def __init__(self, origin_size: tuple[int, int], new_size: tuple[int, int], rotation: float,
to_origin_coordinate: bool):
fs, ts = (new_size, origin_size) if to_origin_coordinate else (origin_size, new_size)
self._rot = rotation if to_origin_coordinate else -rotation
self._c_off = (fs[0] / 2.0, fs[1] / 2.0)
self._n_off = (ts[0] / 2.0, ts[1] / 2.0)
def adjust(self, point: MDRPoint) -> MDRPoint:
x = point[0] - self._c_off[0]
y = point[1] - self._c_off[1]
if x != 0 or y != 0:
cos_r = cos(self._rot)
sin_r = sin(self._rot)
x, y = x * cos_r - y * sin_r, x * sin_r + y * cos_r
return x + self._n_off[0], y + self._n_off[1]
def mdr_normalize_vertical_rotation(rot: float) -> float:
while rot >= pi:
rot -= pi
while rot < 0:
rot += pi
return rot
def _mdr_get_rectangle_angles(rect: MDRRectangle) -> tuple[list[float], list[float]] | None:
h_angs, v_angs = [], []
for i, (p1, p2) in enumerate(rect.segments):
dx = p2[0] - p1[0]
dy = p2[1] - p1[1]
if abs(dx) < 1e-6 and abs(dy) < 1e-6:
continue
ang = atan2(dy, dx)
if ang < 0:
ang += pi
if ang < pi * 0.25 or ang >= pi * 0.75:
h_angs.append(ang - pi if ang >= pi * 0.75 else ang)
else:
v_angs.append(ang)
if not h_angs or not v_angs:
return None
return h_angs, v_angs
def _mdr_normalize_horizontal_angles(rots: list[float]) -> list[float]: return rots
def _mdr_find_median(data: list[float]) -> float:
if not data:
return 0.0
s_data = sorted(data)
n = len(s_data)
return s_data[n // 2] if n % 2 == 1 else (s_data[n // 2 - 1] + s_data[n // 2]) / 2.0
def _mdr_find_mean(data: list[float]) -> float: return sum(data) / len(data) if data else 0.0
def mdr_calculate_image_rotation(frags: list[MDROcrFragment]) -> float:
all_h, all_v = [], []
for f in frags:
res = _mdr_get_rectangle_angles(f.rect)
if res:
h, v = res
all_h.extend(h)
all_v.extend(v)
if not all_h or not all_v:
return 0.0
all_h = _mdr_normalize_horizontal_angles(all_h)
all_v = [mdr_normalize_vertical_rotation(a) for a in all_v]
med_h = _mdr_find_median(all_h)
med_v = _mdr_find_median(all_v)
rot_est = ((pi / 2 - med_v) - med_h) / 2.0
while rot_est >= pi / 2:
rot_est -= pi
while rot_est < -pi / 2:
rot_est += pi
return rot_est
def mdr_calculate_rectangle_rotation(rect: MDRRectangle) -> tuple[float, float]:
res = _mdr_get_rectangle_angles(rect);
if res is None: return 0.0, pi / 2.0;
h_rots, v_rots = res;
h_rots = _mdr_normalize_horizontal_angles(h_rots);
v_rots = [mdr_normalize_vertical_rotation(a) for a in v_rots]
return _mdr_find_mean(h_rots), _mdr_find_mean(v_rots)
# --- MDR ONNX OCR Internals ---
class _MDR_PredictBase:
"""Base class for ONNX model prediction components."""
def get_onnx_session(self, model_path: str, use_gpu: bool):
try:
sess_opts = onnxruntime.SessionOptions()
sess_opts.log_severity_level = 3
providers = ['CUDAExecutionProvider',
'CPUExecutionProvider'] if use_gpu and 'CUDAExecutionProvider' in onnxruntime.get_available_providers() else [
'CPUExecutionProvider']
session = onnxruntime.InferenceSession(model_path, sess_options=sess_opts, providers=providers)
print(f" ONNX session loaded: {Path(model_path).name} ({session.get_providers()})")
return session
except Exception as e:
print(f" ERROR loading ONNX session {Path(model_path).name}: {e}")
if use_gpu and 'CUDAExecutionProvider' not in onnxruntime.get_available_providers():
print(" CUDAExecutionProvider not available. Check ONNXRuntime-GPU installation and CUDA setup.")
raise e
def get_output_name(self, sess: onnxruntime.InferenceSession) -> List[str]:
return [n.name for n in sess.get_outputs()]
def get_input_name(self, sess: onnxruntime.InferenceSession) -> List[str]:
return [n.name for n in sess.get_inputs()]
def get_input_feed(self, names: List[str], img_np: np.ndarray) -> Dict[str, np.ndarray]:
return {name: img_np for name in names}
# --- MDR ONNX OCR Internals ---
class _MDR_NormalizeImage:
def __init__(self, scale=None, mean=None, std=None, order='chw', **kwargs):
self.scale = np.float32(
eval(scale) if isinstance(scale, str) else (scale if scale is not None else 1.0 / 255.0))
mean = mean if mean is not None else [0.485, 0.456, 0.406]
std = std if std is not None else [0.229, 0.224, 0.225]
shape = (3, 1, 1) if order == 'chw' else (1, 1, 3)
self.mean = np.array(mean).reshape(shape).astype('float32')
self.std = np.array(std).reshape(shape).astype('float32')
def __call__(self, data):
img = data['image']
img = np.array(img) if isinstance(img, Image) else img
data['image'] = (img.astype('float32') * self.scale - self.mean) / self.std
return data
class _MDR_DetResizeForTest:
def __init__(self, **kwargs):
self.resize_type = 0
self.keep_ratio = False
if 'image_shape' in kwargs:
self.image_shape = kwargs['image_shape']
self.resize_type = 1
self.keep_ratio = kwargs.get('keep_ratio', False)
elif 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min')
elif 'resize_long' in kwargs:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
else:
self.limit_side_len = 736
self.limit_type = 'min'
def __call__(self, data):
img = data['image']
src_h, src_w, _ = img.shape
if src_h + src_w < 64:
img = self._pad(img)
if self.resize_type == 0:
img, ratios = self._resize0(img)
elif self.resize_type == 2:
img, ratios = self._resize2(img)
else:
img, ratios = self._resize1(img)
if img is None:
return None
data['image'] = img
data['shape'] = np.array([src_h, src_w, ratios[0], ratios[1]])
return data
def _pad(self, im, v=0):
h, w, c = im.shape
p = np.zeros((max(32, h), max(32, w), c), np.uint8) + v
p[:h, :w, :] = im
return p
def _resize1(self, img):
rh, rw = self.image_shape
oh, ow = img.shape[:2]
if self.keep_ratio:
# Calculate new width based on aspect ratio
rw = ow * rh / oh
# Ensure width is a multiple of 32
N = ceil(rw / 32)
rw = N * 32
# Calculate resize ratios
r_h = float(rh) / oh
r_w = float(rw) / ow
# Resize image
img = cv2.resize(img, (int(rw), int(rh)))
return img, [r_h, r_w]
def _resize0(self, img):
lsl = self.limit_side_len
h, w, _ = img.shape
r = 1.0
if self.limit_type == 'max':
r = float(lsl) / max(h, w) if max(h, w) > lsl else 1.0
elif self.limit_type == 'min':
r = float(lsl) / min(h, w) if min(h, w) < lsl else 1.0
elif self.limit_type == 'resize_long':
r = float(lsl) / max(h, w)
else:
raise Exception('Unsupported limit_type')
rh = int(h * r)
rw = int(w * r)
rh = max(int(round(rh / 32) * 32), 32)
rw = max(int(round(rw / 32) * 32), 32)
if int(rw) <= 0 or int(rh) <= 0:
return None, (None, None)
img = cv2.resize(img, (int(rw), int(rh)))
r_h = rh / float(h)
r_w = rw / float(w)
return img, [r_h, r_w]
def _resize2(self, img):
h, w, _ = img.shape
rl = self.resize_long
r = float(rl) / max(h, w)
rh = int(h * r)
rw = int(w * r)
ms = 128
rh = (rh + ms - 1) // ms * ms
rw = (rw + ms - 1) // ms * ms
img = cv2.resize(img, (int(rw), int(rh)))
r_h = rh / float(h)
r_w = rw / float(w)
return img, [r_h, r_w]
class _MDR_ToCHWImage:
def __call__(self, data):
img = data['image']
img = np.array(img) if isinstance(img, Image) else img
data['image'] = img.transpose((2, 0, 1))
return data
class _MDR_KeepKeys:
def __init__(self, keep_keys, **kwargs): self.keep_keys = keep_keys
def __call__(self, data): return [data[key] for key in self.keep_keys]
def mdr_ocr_transform(
data: Any,
ops: Optional[List[Callable[[Any], Optional[Any]]]] = None
) -> Optional[Any]:
"""
Applies a sequence of transformation operations to the input data.
This function iterates through a list of operations (callables) and
applies each one sequentially to the data. If any operation
returns None, the processing stops immediately, and None is returned.
Args:
data: The initial data to be transformed. Can be of any type
compatible with the operations.
ops: An optional list of callable operations. Each operation
should accept the current state of the data and return
the transformed data or None to signal an early exit.
If None or an empty list is provided, the original data
is returned unchanged.
Returns:
The transformed data after applying all operations successfully,
or None if any operation in the sequence returned None.
"""
# Use an empty list if ops is None to avoid errors when iterating
# and to represent "no operations" gracefully.
if ops is None:
operations_to_apply = []
else:
operations_to_apply = ops
current_data = data # Use a separate variable to track the evolving data
# Sequentially apply each operation
for op in operations_to_apply:
current_data = op(current_data) # Apply the operation
# Check if the operation signaled failure or requested early exit
# by returning None.
if current_data is None:
return None # Short-circuit the pipeline
# If the loop completes without returning None, all operations succeeded.
return current_data
def mdr_ocr_create_operators(op_param_list, global_config=None):
ops = []
for operator in op_param_list:
assert isinstance(operator, dict) and len(operator) == 1, "Op config error";
op_name = list(operator)[0]
param = {} if operator[op_name] is None else operator[op_name];
if global_config: param.update(global_config)
op_class_name = f"_MDR_{op_name}" # Map to internal prefixed names
if op_class_name in globals() and isinstance(globals()[op_class_name], type):
ops.append(globals()[op_class_name](**param))
else:
raise ValueError(f"Operator class '{op_class_name}' not found.")
return ops
class _MDR_DBPostProcess:
def __init__(self, thresh=0.3, box_thresh=0.7, max_candidates=1000, unclip_ratio=1.5, use_dilation=False,
score_mode="fast", box_type='quad', **kwargs):
self.thresh = thresh
self.box_thresh = box_thresh
self.max_cand = max_candidates
self.unclip_r = unclip_ratio
self.min_sz = 3
self.score_m = score_mode
self.box_t = box_type
assert score_mode in ["slow", "fast"]
self.dila_k = np.array([[1, 1], [1, 1]], dtype=np.uint8) if use_dilation else None
def _polygons_from_bitmap(self, pred, bmp, dw, dh):
h, w = bmp.shape
boxes, scores = [], []
contours, _ = cv2.findContours((bmp * 255).astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
for contour in contours[:self.max_cand]:
eps = 0.002 * cv2.arcLength(contour, True)
approx = cv2.approxPolyDP(contour, eps, True)
pts = approx.reshape((-1, 2))
if pts.shape[0] < 4:
continue
score = self._box_score_fast(pred, pts.reshape(-1, 2))
if self.box_thresh > score:
continue
try:
box = self._unclip(pts, self.unclip_r)
except:
continue
if len(box) > 1:
continue
box = box.reshape(-1, 2)
_, sside = self._get_mini_boxes(box.reshape((-1, 1, 2)))
if sside < self.min_sz + 2:
continue
box = np.array(box)
box[:, 0] = np.clip(np.round(box[:, 0] / w * dw), 0, dw)
box[:, 1] = np.clip(np.round(box[:, 1] / h * dh), 0, dh)
boxes.append(box.tolist())
scores.append(score)
return boxes, scores
# In class _MDR_DBPostProcess:
def _boxes_from_bitmap(self, pred, bmp, dw, dh): # pred is the probability map, bmp is the binarized map
h, w = bmp.shape
# ADDED: More detailed logging
print(
f" DEBUG OCR: _boxes_from_bitmap: Processing bitmap of shape {h}x{w} for original dimensions {dw:.1f}x{dh:.1f}.")
contours, _ = cv2.findContours((bmp * 255).astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
num_contours_found = len(contours)
print(f" DEBUG OCR: _boxes_from_bitmap: Found {num_contours_found} raw contours.")
num_contours_to_process = min(num_contours_found, self.max_cand)
if num_contours_found > self.max_cand:
print(
f" DEBUG OCR: _boxes_from_bitmap: Processing limited to {self.max_cand} contours (max_candidates).")
boxes, scores = [], []
kept_boxes_count = 0
for i in range(num_contours_to_process):
contour = contours[i]
pts_mini_box, sside = self._get_mini_boxes(contour)
if sside < self.min_sz:
# print(f" DEBUG OCR: Contour {i} too small (sside {sside:.2f} < min_sz {self.min_sz}). Skipping.") # Can be too verbose
continue
pts_arr = np.array(pts_mini_box)
current_score = self._box_score_fast(pred, pts_arr.reshape(-1,
2)) if self.score_m == "fast" else self._box_score_slow(
pred, contour)
if self.box_thresh > current_score:
# print(f" DEBUG OCR: Contour {i} score {current_score:.4f} < box_thresh {self.box_thresh}. Skipping.") # Can be too verbose
continue
try:
box_unclipped = self._unclip(pts_arr, self.unclip_r).reshape(-1, 1, 2)
except Exception as e_unclip:
# print(f" DEBUG OCR: Contour {i} unclip failed: {e_unclip}. Skipping.") # Can be too verbose
continue
box_final, sside_final = self._get_mini_boxes(box_unclipped)
if sside_final < self.min_sz + 2: # min_sz is 3
# print(f" DEBUG OCR: Contour {i} final size after unclip too small (sside_final {sside_final:.2f} < {self.min_sz + 2}). Skipping.") # Can be too verbose
continue
box_final_arr = np.array(box_final)
box_final_arr[:, 0] = np.clip(np.round(box_final_arr[:, 0] / w * dw), 0, dw)
box_final_arr[:, 1] = np.clip(np.round(box_final_arr[:, 1] / h * dh), 0, dh)
boxes.append(box_final_arr.astype("int32"))
scores.append(current_score)
kept_boxes_count += 1
print(
f" DEBUG OCR: _boxes_from_bitmap: Kept {kept_boxes_count} boxes after all filtering (size, score, unclip). Configured box_thresh: {self.box_thresh}, min_sz: {self.min_sz}.")
return np.array(boxes, dtype="int32"), scores
def _unclip(self, box, ratio):
poly = Polygon(box)
dist = poly.area * ratio / poly.length
offset = pyclipper.PyclipperOffset()
offset.AddPath(box, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
expanded = offset.Execute(dist)
if not expanded:
raise ValueError("Unclip failed")
return np.array(expanded[0])
def _get_mini_boxes(self, contour):
bb = cv2.minAreaRect(contour)
pts = sorted(list(cv2.boxPoints(bb)), key=lambda x: x[0])
i1, i4 = (0, 1) if pts[1][1] > pts[0][1] else (1, 0)
i2, i3 = (2, 3) if pts[3][1] > pts[2][1] else (3, 2)
box = [pts[i1], pts[i2], pts[i3], pts[i4]]
return box, min(bb[1])
def _box_score_fast(self, bmp, box):
h, w = bmp.shape[:2]
xmin = np.clip(np.floor(box[:, 0].min()).astype("int32"), 0, w - 1)
xmax = np.clip(np.ceil(box[:, 0].max()).astype("int32"), 0, w - 1)
ymin = np.clip(np.floor(box[:, 1].min()).astype("int32"), 0, h - 1)
ymax = np.clip(np.ceil(box[:, 1].max()).astype("int32"), 0, h - 1)
mask = np.zeros((ymax - ymin + 1, xmax - xmin + 1), dtype=np.uint8)
box[:, 0] -= xmin
box[:, 1] -= ymin
cv2.fillPoly(mask, box.reshape(1, -1, 2).astype("int32"), 1)
return cv2.mean(bmp[ymin: ymax + 1, xmin: xmax + 1], mask)[0] if np.sum(mask) > 0 else 0.0
def _box_score_slow(self, bmp, contour): # Not used if fast
h, w = bmp.shape[:2]
contour = np.reshape(contour.copy(), (-1, 2))
xmin = np.clip(np.min(contour[:, 0]), 0, w - 1)
xmax = np.clip(np.max(contour[:, 0]), 0, w - 1)
ymin = np.clip(np.min(contour[:, 1]), 0, h - 1)
ymax = np.clip(np.max(contour[:, 1]), 0, h - 1)
mask = np.zeros((ymax - ymin + 1, xmax - xmin + 1), dtype=np.uint8)
contour[:, 0] -= xmin
contour[:, 1] -= ymin
cv2.fillPoly(mask, contour.reshape(1, -1, 2).astype("int32"), 1)
return cv2.mean(bmp[ymin: ymax + 1, xmin: xmax + 1], mask)[0] if np.sum(mask) > 0 else 0.0
def __call__(self, outs_dict, shape_list):
pred = outs_dict['maps'][:, 0, :, :]
seg = pred > self.thresh
# ADDED: More detailed logging
print(
f" DEBUG OCR: _MDR_DBPostProcess: pred map shape: {pred.shape}, seg map shape: {seg.shape}, configured thresh: {self.thresh}")
print(
f" DEBUG OCR: _MDR_DBPostProcess: Number of pixels in seg map above threshold (sum of all batches): {np.sum(seg)}")
boxes_batch = []
for batch_idx in range(pred.shape[0]):
# MODIFIED: Ensure sh, sw are floats for division if they come from shape_list
sh_orig, sw_orig, rh_ratio, rw_ratio = shape_list[batch_idx]
# The dw, dh for _boxes_from_bitmap should be the original image dimensions before DetResizeForTest
# shape_list contains [src_h, src_w, ratio_h, ratio_w]
# So dw = src_w, dh = src_h
dw_orig, dh_orig = sw_orig, sh_orig
current_pred_map = pred[batch_idx]
current_seg_map = seg[batch_idx]
mask = cv2.dilate(np.array(current_seg_map).astype(np.uint8),
self.dila_k) if self.dila_k is not None else current_seg_map
print(
f" DEBUG OCR: _MDR_DBPostProcess (batch {batch_idx}): Input shape to postproc (orig) {dh_orig:.1f}x{dw_orig:.1f}. Sum of mask pixels: {np.sum(mask)}")
if self.box_t == 'poly':
boxes, scores = self._polygons_from_bitmap(current_pred_map, mask, dh_orig, dw_orig)
elif self.box_t == 'quad':
boxes, scores = self._boxes_from_bitmap(current_pred_map, mask, dh_orig, dw_orig)
else:
raise ValueError("box_type must be 'quad' or 'poly'")
print(
f" DEBUG OCR: _MDR_DBPostProcess (batch {batch_idx}): Found {len(boxes)} boxes from bitmap processing.")
boxes_batch.append({'points': boxes})
return boxes_batch
class _MDR_TextDetector(_MDR_PredictBase):
def __init__(self, args):
super().__init__()
self.args = args
pre_ops = [{'DetResizeForTest': {'limit_side_len': args.det_limit_side_len, 'limit_type': args.det_limit_type}},
{'NormalizeImage': {'std': [0.229, 0.224, 0.225], 'mean': [0.485, 0.456, 0.406], 'scale': '1./255.',
'order': 'hwc'}}, {'ToCHWImage': None},
{'KeepKeys': {'keep_keys': ['image', 'shape']}}]
self.pre_op = mdr_ocr_create_operators(pre_ops)
post_params = {'thresh': args.det_db_thresh, 'box_thresh': args.det_db_box_thresh, 'max_candidates': 1000,
'unclip_ratio': args.det_db_unclip_ratio, 'use_dilation': args.use_dilation,
'score_mode': args.det_db_score_mode, 'box_type': args.det_box_type}
self.post_op = _MDR_DBPostProcess(**post_params)
self.sess = self.get_onnx_session(args.det_model_dir, args.use_gpu)
self.input_name = self.get_input_name(self.sess)
self.output_name = self.get_output_name(self.sess)
def _order_pts(self, pts):
r = np.zeros((4, 2), dtype="float32")
s = pts.sum(axis=1)
r[0] = pts[np.argmin(s)]
r[2] = pts[np.argmax(s)]
tmp = np.delete(pts, (np.argmin(s), np.argmax(s)), axis=0)
d = np.diff(np.array(tmp), axis=1)
r[1] = tmp[np.argmin(d)]
r[3] = tmp[np.argmax(d)]
return r
def _clip_pts(self, pts, h, w):
pts[:, 0] = np.clip(pts[:, 0], 0, w - 1)
pts[:, 1] = np.clip(pts[:, 1], 0, h - 1)
return pts
def _filter_quad(self, boxes, shape):
h, w = shape[0:2]
new_boxes = []
for box in boxes:
box = np.array(box) if isinstance(box, list) else box
box = self._order_pts(box)
box = self._clip_pts(box, h, w)
rw = int(np.linalg.norm(box[0] - box[1]))
rh = int(np.linalg.norm(box[0] - box[3]))
if rw <= 3 or rh <= 3:
continue
new_boxes.append(box)
return np.array(new_boxes)
def _filter_poly(self, boxes, shape):
h, w = shape[0:2]
new_boxes = []
for box in boxes:
box = np.array(box) if isinstance(box, list) else box
box = self._clip_pts(box, h, w)
if Polygon(box).area < 10:
continue
new_boxes.append(box)
return np.array(new_boxes)
# In class _MDR_TextDetector:
def __call__(self, img):
ori_im = img.copy()
data = {"image": img}
print(f" DEBUG OCR: _MDR_TextDetector: Original image shape: {ori_im.shape}")
# Preprocessing
try:
data = mdr_ocr_transform(data, self.pre_op)
except Exception as e_preproc:
print(f" DEBUG OCR: _MDR_TextDetector: Error during preprocessing (mdr_ocr_transform): {e_preproc}")
import traceback
traceback.print_exc()
return np.array([])
if data is None:
print(
" DEBUG OCR: _MDR_TextDetector: Preprocessing (mdr_ocr_transform) returned None. No text will be detected.")
return np.array([])
processed_img, shape_list = data
if processed_img is None:
print(" DEBUG OCR: _MDR_TextDetector: Processed image after transform is None. No text will be detected.")
return np.array([])
print(
f" DEBUG OCR: _MDR_TextDetector: Processed image shape for ONNX: {processed_img.shape}, shape_list: {shape_list}")
img_for_onnx = np.expand_dims(processed_img, axis=0)
shape_list_for_onnx = np.expand_dims(shape_list, axis=0)
img_for_onnx = img_for_onnx.copy()
inputs = self.get_input_feed(self.input_name, img_for_onnx)
print(f" DEBUG OCR: _MDR_TextDetector: Running ONNX inference for text detection...")
try:
outputs = self.sess.run(self.output_name, input_feed=inputs)
except Exception as e_infer:
print(f" DEBUG OCR: _MDR_TextDetector: ONNX inference for detection failed: {e_infer}")
import traceback
traceback.print_exc()
return np.array([])
print(f" DEBUG OCR: _MDR_TextDetector: ONNX inference done. Output map shape: {outputs[0].shape}")
preds = {"maps": outputs[0]}
try:
post_res = self.post_op(preds, shape_list_for_onnx)
except Exception as e_postproc:
print(f" DEBUG OCR: _MDR_TextDetector: Error during DBPostProcess: {e_postproc}")
import traceback
traceback.print_exc()
return np.array([])
# --- START: REFINED CHECK ---
# 1. Check if post_res itself is valid and contains the expected structure.
if not post_res or not isinstance(post_res, list) or len(post_res) == 0 or \
not isinstance(post_res[0], dict) or 'points' not in post_res[0]:
print(" DEBUG OCR: _MDR_TextDetector: DBPostProcess returned invalid or empty structure for points.")
return np.array([])
boxes_from_post = post_res[0]['points'] # This is expected to be a np.ndarray or a list of boxes
# 2. Check if boxes_from_post is actually empty.
# For a NumPy array, check its size. For a list, check if it's empty.
no_boxes_found = False
if isinstance(boxes_from_post, np.ndarray):
if boxes_from_post.size == 0:
no_boxes_found = True
elif isinstance(boxes_from_post, list):
if not boxes_from_post: # Empty list
no_boxes_found = True
elif boxes_from_post is None: # Explicitly check for None
no_boxes_found = True
else:
# Should not happen if _MDR_DBPostProcess behaves as expected, but good to log
print(
f" DEBUG OCR: _MDR_TextDetector: 'points' from DBPostProcess is of unexpected type: {type(boxes_from_post)}")
return np.array([])
if no_boxes_found:
print(" DEBUG OCR: _MDR_TextDetector: DBPostProcess returned no actual point data.")
return np.array([])
# --- END: REFINED CHECK ---
print(
f" DEBUG OCR: _MDR_TextDetector: Boxes from DBPostProcess before final filtering: {len(boxes_from_post)}")
# The following check might be redundant now but can be kept for extra safety
# or if boxes_from_post could be other types not handled above.
if not isinstance(boxes_from_post, (list, np.ndarray)) or \
(isinstance(boxes_from_post, np.ndarray) and boxes_from_post.size == 0) or \
(isinstance(boxes_from_post, list) and not boxes_from_post):
print(" DEBUG OCR: _MDR_TextDetector: No boxes from DBPostProcess to filter (secondary check).")
return np.array([])
if self.args.det_box_type == 'poly':
final_boxes = self._filter_poly(boxes_from_post, ori_im.shape)
else: # 'quad'
final_boxes = self._filter_quad(boxes_from_post, ori_im.shape)
print(f" DEBUG OCR: _MDR_TextDetector: Boxes after final poly/quad filtering: {len(final_boxes)}")
return final_boxes
class _MDR_ClsPostProcess:
def __init__(self, label_list=None, **kwargs): self.labels = label_list if label_list else {0: '0', 1: '180'}
def __call__(self, preds, label=None, *args, **kwargs):
preds = np.array(preds) if not isinstance(preds, np.ndarray) else preds;
idxs = preds.argmax(axis=1)
return [(self.labels[idx], float(preds[i, idx])) for i, idx in enumerate(idxs)]
class _MDR_TextClassifier(_MDR_PredictBase):
def __init__(self, args):
super().__init__()
self.shape = tuple(map(int, args.cls_image_shape.split(','))) if isinstance(args.cls_image_shape,
str) else args.cls_image_shape
self.batch_num = args.cls_batch_num
self.thresh = args.cls_thresh
self.post_op = _MDR_ClsPostProcess(label_list=args.label_list)
self.sess = self.get_onnx_session(args.cls_model_dir, args.use_gpu)
self.input_name = self.get_input_name(self.sess)
self.output_name = self.get_output_name(self.sess)
def _resize_norm(self, img):
imgC, imgH, imgW = self.shape
h, w = img.shape[:2]
r = w / float(h) if h > 0 else 0
rw = int(ceil(imgH * r))
rw = min(rw, imgW)
resized = cv2.resize(img, (rw, imgH))
resized = resized.astype("float32")
if imgC == 1:
resized = resized / 255.0
resized = resized[np.newaxis, :]
else:
resized = resized.transpose((2, 0, 1)) / 255.0
resized -= 0.5
resized /= 0.5
padding = np.zeros((imgC, imgH, imgW), dtype=np.float32)
padding[:, :, 0:rw] = resized
return padding
def __call__(self, img_list):
if not img_list:
return img_list, []
img_list_cp = copy.deepcopy(img_list)
num = len(img_list_cp)
ratios = [img.shape[1] / float(img.shape[0]) if img.shape[0] > 0 else 0 for img in img_list_cp]
indices = np.argsort(np.array(ratios))
results = [["", 0.0]] * num
batch_n = self.batch_num
for start in range(0, num, batch_n):
end = min(num, start + batch_n)
batch = []
for i in range(start, end):
batch.append(self._resize_norm(img_list_cp[indices[i]])[np.newaxis, :])
if not batch:
continue
batch = np.concatenate(batch, axis=0).copy()
inputs = self.get_input_feed(self.input_name, batch)
outputs = self.sess.run(self.output_name, input_feed=inputs)
cls_out = self.post_op(outputs[0])
for i in range(len(cls_out)):
orig_idx = indices[start + i]
label, score = cls_out[i]
results[orig_idx] = [label, score]
if "180" in label and score > self.thresh:
img_list[orig_idx] = cv2.rotate(img_list[orig_idx], cv2.ROTATE_180)
return img_list, results
class _MDR_BaseRecLabelDecode:
def __init__(self, char_path=None, use_space=False):
self.beg, self.end, self.rev = "sos", "eos", False
self.chars = []
if char_path is None:
self.chars = list("0123456789abcdefghijklmnopqrstuvwxyz")
else:
try:
with open(char_path, "rb") as f:
self.chars = [l.decode("utf-8").strip("\n\r") for l in f]
if use_space:
self.chars.append(" ")
if any("\u0600" <= c <= "\u06FF" for c in self.chars):
self.rev = True
except FileNotFoundError:
print(f"Warn: Dict not found {char_path}")
self.chars = list("0123456789abcdefghijklmnopqrstuvwxyz")
if use_space:
self.chars.append(" ")
d_char = self.add_special_char(list(self.chars))
self.dict = {c: i for i, c in enumerate(d_char)}
self.character = d_char
def add_special_char(self, chars):
return chars
def get_ignored_tokens(self):
return []
def _reverse(self, pred):
res = []
cur = ""
for c in pred:
if not re.search("[a-zA-Z0-9 :*./%+-]", c):
if cur != "":
res.extend([cur, c])
else:
res.extend([c])
cur = ""
else:
cur += c
if cur != "":
res.append(cur)
return "".join(res[::-1])
def decode(self, idxs, probs=None, remove_dup=False):
res = []
ignored = self.get_ignored_tokens()
bs = len(idxs)
for b_idx in range(bs):
sel = np.ones(len(idxs[b_idx]), dtype=bool)
if remove_dup:
sel[1:] = idxs[b_idx][1:] != idxs[b_idx][:-1]
for ig_tok in ignored:
sel &= idxs[b_idx] != ig_tok
char_l = [
self.character[tid]
for tid in idxs[b_idx][sel]
if 0 <= tid < len(self.character)
]
conf_l = probs[b_idx][sel] if probs is not None else [1] * len(char_l)
if len(conf_l) == 0:
conf_l = [0]
txt = "".join(char_l)
if self.rev:
txt = self._reverse(txt)
res.append((txt, float(np.mean(conf_l))))
return res
class _MDR_CTCLabelDecode(_MDR_BaseRecLabelDecode):
def __init__(self, char_path=None, use_space=False, **kwargs): super().__init__(char_path, use_space)
def add_special_char(self, chars): return ["blank"] + chars
def get_ignored_tokens(self): return [0] # blank index
def __call__(self, preds, label=None, *args, **kwargs):
preds = preds[-1] if isinstance(preds, (tuple, list)) else preds;
preds = np.array(preds) if not isinstance(preds, np.ndarray) else preds
idxs = preds.argmax(axis=2);
probs = preds.max(axis=2);
txt = self.decode(idxs, probs, remove_dup=True);
return txt
class _MDR_TextRecognizer(_MDR_PredictBase):
def __init__(self, args):
super().__init__()
shape_str = getattr(args, 'rec_image_shape', "3,48,320")
self.shape = tuple(map(int, shape_str.split(',')))
self.batch_num = getattr(args, 'rec_batch_num', 6)
self.algo = getattr(args, 'rec_algorithm', 'SVTR_LCNet')
self.post_op = _MDR_CTCLabelDecode(char_path=args.rec_char_dict_path,
use_space=getattr(args, 'use_space_char', True))
self.sess = self.get_onnx_session(args.rec_model_dir, args.use_gpu)
self.input_name = self.get_input_name(self.sess)
self.output_name = self.get_output_name(self.sess)
# In class _MDR_TextRecognizer
def _resize_norm(self, img, max_r): # img is a single crop
imgC, imgH, imgW = self.shape # e.g., (3, 48, 320)
h_orig, w_orig = img.shape[:2]
# ADDED: Log input crop shape
print(
f" DEBUG RECOGNIZER: _resize_norm input crop shape: ({h_orig}, {w_orig}), target shape: {self.shape}, max_r_batch: {max_r:.2f}")
# --- START OF FIX ---
MIN_DIM_FOR_RESIZE = 2 # Minimum original height or width to attempt resize
if h_orig < MIN_DIM_FOR_RESIZE or w_orig < MIN_DIM_FOR_RESIZE:
print(
f" DEBUG RECOGNIZER: _resize_norm received degenerate crop ({h_orig}x{w_orig}) with dimension < {MIN_DIM_FOR_RESIZE}. Returning zeros before resize attempt.")
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
# --- END OF FIX ---
# Original check for h_orig == 0 or w_orig == 0 is now covered by the above,
# but can be kept for explicitness or if MIN_DIM_FOR_RESIZE is set to 1.
# If MIN_DIM_FOR_RESIZE is 1, the original check is still useful.
# If MIN_DIM_FOR_RESIZE is > 1, this specific check becomes redundant.
# Let's keep it for safety if MIN_DIM_FOR_RESIZE is changed.
if h_orig == 0 or w_orig == 0: # This check is technically redundant if MIN_DIM_FOR_RESIZE >= 1
print(
f" DEBUG RECOGNIZER: _resize_norm received zero-dimension crop ({h_orig}x{w_orig}) (secondary check). Returning zeros.")
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
r_current = w_orig / float(h_orig) # h_orig is guaranteed > 0 here if MIN_DIM_FOR_RESIZE >=1
tw = min(imgW, int(ceil(imgH * r_current)))
tw = max(1, tw) # Ensure target width is at least 1
# Ensure target height (imgH) is also valid (it comes from self.shape, so should be)
print(f" DEBUG RECOGNIZER: _resize_norm calculated target width (tw): {tw} for target height (imgH): {imgH}")
try:
# Ensure target dimensions for resize are valid
if tw <= 0 or imgH <= 0:
print(
f" DEBUG RECOGNIZER: _resize_norm calculated invalid target resize dimensions (tw: {tw}, imgH: {imgH}). Returning zeros.")
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
resized = cv2.resize(img, (tw, imgH))
except cv2.error as e_resize: # Catch specific cv2 error
print(
f" DEBUG RECOGNIZER: _resize_norm cv2.resize failed: {e_resize}. Original shape ({h_orig},{w_orig}), target ({tw},{imgH}). Returning zeros.")
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
except Exception as e_resize_general: # Catch any other unexpected error
print(
f" DEBUG RECOGNIZER: _resize_norm general error during resize: {e_resize_general}. Original shape ({h_orig},{w_orig}), target ({tw},{imgH}). Returning zeros.")
import traceback
traceback.print_exc()
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
# ... rest of the normalization code ...
resized = resized.astype("float32")
if imgC == 1 and len(resized.shape) == 3: # If target is 1 channel and resized is 3
if resized.shape[2] == 3: # Check if it actually has 3 channels
resized = cv2.cvtColor(resized, cv2.COLOR_BGR2GRAY)
if len(resized.shape) == 2: # If grayscale after potential conversion
resized = resized[:, :, np.newaxis] # Add channel dim
# Ensure resized has 3 channels if imgC is 3, even if input was grayscale or became grayscale
if imgC == 3 and resized.shape[2] == 1:
resized = cv2.cvtColor(resized, cv2.COLOR_GRAY2BGR)
# Final check on channel consistency
if resized.shape[2] != imgC:
print(
f" DEBUG RECOGNIZER: Channel mismatch after processing. Expected {imgC}, got {resized.shape[2]}. Crop shape ({h_orig},{w_orig}). Returning zeros.")
return np.zeros((imgC, imgH, imgW), dtype=np.float32)
resized = resized.transpose((2, 0, 1)) / 255.0
resized -= 0.5
resized /= 0.5
padding = np.zeros((imgC, imgH, imgW), dtype=np.float32)
# Ensure tw is not out of bounds for padding
actual_padded_width = min(tw, imgW)
padding[:, :, 0:actual_padded_width] = resized[:, :, 0:actual_padded_width]
print(f" DEBUG RECOGNIZER: _resize_norm output padded shape: {padding.shape}")
# ... rest of the logging ...
min_px, max_px, mean_px = np.min(padding), np.max(padding), np.mean(padding)
print(f" DEBUG RECOGNIZER: Normalized Crop Properties (before ONNX): "
f"dtype: {padding.dtype}, "
f"MinPx: {min_px:.4f}, "
f"MaxPx: {max_px:.4f}, "
f"MeanPx: {mean_px:.4f}")
if np.all(padding == 0):
print(" DEBUG RECOGNIZER: WARNING - Normalized image is all zeros!")
elif np.abs(max_px - min_px) < 1e-6:
print(f" DEBUG RECOGNIZER: WARNING - Normalized image is a constant value: {mean_px:.4f}")
return padding
def __call__(self, img_list):
if not img_list:
return []
num = len(img_list)
ratios = [img.shape[1] / float(img.shape[0]) if img.shape[0] > 0 else 0 for img in img_list]
indices = np.argsort(np.array(ratios))
results = [["", 0.0]] * num
batch_n = self.batch_num
for start in range(0, num, batch_n):
end = min(num, start + batch_n)
batch = []
max_r_batch = 0
for i in range(start, end):
h, w = img_list[indices[i]].shape[0:2]
if h > 0:
max_r_batch = max(max_r_batch, w / float(h))
for i in range(start, end):
batch.append(self._resize_norm(img_list[indices[i]], max_r_batch)[np.newaxis, :])
if not batch:
continue
batch = np.concatenate(batch, axis=0).copy()
inputs = self.get_input_feed(self.input_name, batch)
outputs = self.sess.run(self.output_name, input_feed=inputs)
rec_out = self.post_op(outputs[0])
for i in range(len(rec_out)):
results[indices[start + i]] = rec_out[i]
return results
# --- MDR ONNX OCR System ---
class _MDR_TextSystem:
def __init__(self, args):
class ArgsObject: # Helper to access dict args with dot notation
def __init__(self, **entries): self.__dict__.update(entries)
if isinstance(args, dict): args = ArgsObject(**args)
self.args = args
self.detector = _MDR_TextDetector(args)
self.recognizer = _MDR_TextRecognizer(args)
self.use_cls = getattr(args, 'use_angle_cls', True)
self.drop_score = getattr(args, 'drop_score', 0.5)
self.classifier = _MDR_TextClassifier(args) if self.use_cls else None
self.crop_idx = 0
self.save_crop = getattr(args, 'save_crop_res', False)
self.crop_dir = getattr(args, 'crop_res_save_dir', "./output/mdr_crop_res")
def __call__(self, img: np.ndarray) -> tuple[list[np.ndarray], list[tuple[str, float]]]:
ori_im = img.copy()
dt_boxes: np.ndarray = self.detector(img)
print(
f" DEBUG TextSystem: Detector found {len(dt_boxes) if dt_boxes is not None and dt_boxes.size > 0 else 0} initial boxes.")
if dt_boxes is None or dt_boxes.size == 0:
return [], []
dt_boxes_sorted: list[np.ndarray] = self._sort_boxes(dt_boxes)
print(f" DEBUG TextSystem: Sorted {len(dt_boxes_sorted)} boxes.")
if not dt_boxes_sorted:
return [], []
# --- Stage 1 Fix: Refined filtering of boxes and creation of crops ---
# Ensure dt_boxes_sorted and img_crop_list are synchronized.
valid_boxes_for_cropping: list[np.ndarray] = []
img_crop_list: list[np.ndarray] = [] # Initialize img_crop_list here
for i, box_pts in enumerate(dt_boxes_sorted):
crop_im = mdr_get_rotated_crop(ori_im, box_pts)
if crop_im is not None and crop_im.shape[0] > 1 and crop_im.shape[1] > 1: # Min height/width for a crop
valid_boxes_for_cropping.append(box_pts)
img_crop_list.append(crop_im) # Directly populate the final img_crop_list
else:
print(
f" DEBUG TextSystem: Crop for box {i} (pts: {box_pts}) was None or too small. Skipping this box.")
dt_boxes_sorted = valid_boxes_for_cropping # Update dt_boxes_sorted to only include those that yielded valid crops
# img_crop_list is now the correctly filtered list of crops, synchronized with dt_boxes_sorted.
# --- End of Stage 1 Fix ---
print(f" DEBUG TextSystem: Created {len(img_crop_list)} valid crops for further processing.")
if not img_crop_list: # If no valid crops were made
print(" DEBUG TextSystem: No valid crops generated. Returning empty.")
return [], []
if self.use_cls and self.classifier is not None:
print(f" DEBUG TextSystem: Applying text classification for {len(img_crop_list)} crops.")
img_crop_list, cls_results = self.classifier(
img_crop_list) # classifier might modify img_crop_list (e.g., rotate)
print(f" DEBUG TextSystem: Classification complete. {len(cls_results if cls_results else [])} results.")
rec_results: list[tuple[str, float]] = []
print(f" DEBUG TextSystem: Recognizing text for {len(img_crop_list)} crops.")
rec_results = self.recognizer(img_crop_list)
print(f" DEBUG TextSystem: Recognizer returned {len(rec_results)} results.")
# --- Start of Stage 2 Fix: Robust handling of rec_results length ---
expected_count = len(dt_boxes_sorted) # This is synchronized with len(img_crop_list) before recognizer
# and should still match len(img_crop_list) after classifier
# if classifier preserves length.
actual_rec_count = len(rec_results)
num_to_process = 0
if actual_rec_count == expected_count:
num_to_process = actual_rec_count
else:
print(f" DEBUG TextSystem: WARNING - Mismatch in lengths after recognition! "
f"Expected (from boxes/crops): {expected_count}, "
f"Recognizer returned: {actual_rec_count} results. ")
num_to_process = min(actual_rec_count, expected_count)
if num_to_process < expected_count:
print(
f" DEBUG TextSystem: Will process {num_to_process} items due to mismatch. Some data might be lost if recognizer dropped results or if there was an issue in earlier stages not caught.")
elif num_to_process < actual_rec_count: # Recognizer returned more than expected
print(
f" DEBUG TextSystem: Will process {num_to_process} items. Recognizer returned more results ({actual_rec_count}) than expected crops ({expected_count}). Extra recognition results will be ignored.")
if num_to_process == 0:
if expected_count > 0: # If there were boxes/crops but no rec results to process
print(
" DEBUG TextSystem: No recognition results to process (num_to_process is 0) despite having input boxes/crops. Returning empty.")
else: # If there were no boxes/crops to begin with
print(
" DEBUG TextSystem: No items to process (no initial boxes or num_to_process is 0). Returning empty.")
return [], []
# --- End of Stage 2 Fix preamble ---
print(
f" DEBUG TextSystem: Filtering {num_to_process} recognition results with drop_score: {self.drop_score}")
final_boxes_to_return: list[np.ndarray] = []
final_recs_to_return: list[tuple[str, float]] = []
final_crops_for_saving: list[np.ndarray] = []
# --- Stage 2 Fix: Modified Loop (No outer strict if/else) ---
for i in range(num_to_process): # Iterate up to the safe number
# It's crucial that dt_boxes_sorted[i], rec_results[i], and img_crop_list[i] correspond
# for the items being processed.
text, confidence = rec_results[i]
print(f" DEBUG TextSystem: Rec item {i} - Text: '{text}', Confidence: {confidence:.4f}")
if confidence >= self.drop_score:
if text and not mdr_is_whitespace(text):
final_boxes_to_return.append(dt_boxes_sorted[i])
final_recs_to_return.append(rec_results[i])
if self.save_crop:
# Ensure img_crop_list[i] is valid if classifier could have changed its length
# However, self.classifier is expected to return img_list of same length as input.
final_crops_for_saving.append(img_crop_list[i])
else:
print(f" DEBUG TextSystem: Item {i} REJECTED (empty/whitespace text).")
else:
print(
f" DEBUG TextSystem: Item {i} REJECTED (confidence {confidence:.4f} < drop_score {self.drop_score}).")
# --- End of Stage 2 Fix: Modified Loop ---
print(f" DEBUG TextSystem: Kept {len(final_boxes_to_return)} boxes after recognition and filtering.")
if self.save_crop and final_crops_for_saving:
print(f" DEBUG TextSystem: Saving {len(final_crops_for_saving)} filtered crops.")
self._save_crops(final_crops_for_saving, final_recs_to_return)
return final_boxes_to_return, final_recs_to_return
def _sort_boxes(self, boxes):
if boxes is None or len(boxes) == 0: return []
def key(box):
min_y = min(p[1] for p in box); min_x = min(p[0] for p in box); return (min_y, min_x)
try:
return list(sorted(boxes, key=key))
except:
return list(boxes) # Fallback
def _save_crops(self, crops, recs):
mdr_ensure_directory(self.crop_dir)
num = len(crops)
for i in range(num):
txt, score = recs[i]
safe = re.sub(r'\W+', '_', txt)[:20]
fname = f"crop_{self.crop_idx + i}_{safe}_{score:.2f}.jpg"
cv2.imwrite(os.path.join(self.crop_dir, fname), crops[i])
self.crop_idx += num
# --- MDR ONNX OCR Utilities ---
def mdr_get_rotated_crop(img, points):
"""Crops and perspective-transforms a quadrilateral region."""
pts = np.array(points, dtype="float32")
assert len(pts) == 4
w = int(max(np.linalg.norm(pts[0] - pts[1]), np.linalg.norm(pts[2] - pts[3])))
h = int(max(np.linalg.norm(pts[0] - pts[3]), np.linalg.norm(pts[1] - pts[2])))
std = np.float32([[0, 0], [w, 0], [w, h], [0, h]])
M = cv2.getPerspectiveTransform(pts, std)
dst = cv2.warpPerspective(img, M, (w, h), borderMode=cv2.BORDER_CONSTANT, borderValue=(128, 128, 128), flags=cv2.INTER_CUBIC)
dh, dw = dst.shape[0:2]
if dh > 0 and dw > 0 and dh * 1.0 / dw >= 1.5:
dst = cv2.rotate(dst, cv2.ROTATE_90_CLOCKWISE)
return dst
def mdr_get_min_area_crop(img, points):
"""Crops the minimum area rectangle containing the points."""
bb = cv2.minAreaRect(np.array(points).astype(np.int32))
box_pts = cv2.boxPoints(bb)
return mdr_get_rotated_crop(img, box_pts)
# --- MDR Layout Processing ---
_MDR_INCLUDES_MIN_RATE = 0.99
class _MDR_OverlapMatrixContext:
def __init__(self, layouts: list[MDRLayoutElement]):
length = len(layouts);
self.polys: list[Polygon | None] = []
for l in layouts:
try:
p = Polygon(l.rect); self.polys.append(p if p.is_valid else None)
except:
self.polys.append(None)
self.matrix = [[0.0] * length for _ in range(length)];
self.removed = set()
for i in range(length):
p1 = self.polys[i];
if p1 is None: continue; self.matrix[i][i] = 1.0
for j in range(i + 1, length):
p2 = self.polys[j];
if p2 is None: continue
r_ij = self._rate(p1, p2);
r_ji = self._rate(p2, p1);
self.matrix[i][j] = r_ij;
self.matrix[j][i] = r_ji
def _rate(self, p1: Polygon, p2: Polygon) -> float: # Rate p1 covers p2
try:
inter = p1.intersection(p2)
except:
return 0.0
if inter.is_empty or inter.area < 1e-6:
return 0.0
_, _, ix1, iy1 = inter.bounds
iw = ix1 - inter.bounds[0]
ih = iy1 - inter.bounds[1]
_, _, px1, py1 = p2.bounds
pw = px1 - p2.bounds[0]
ph = py1 - p2.bounds[1]
if pw < 1e-6 or ph < 1e-6:
return 0.0
wr = min(iw / pw, 1.0)
hr = min(ih / ph, 1.0)
return (wr + hr) / 2.0
def others(self, idx: int):
for i, r in enumerate(self.matrix[idx]):
if i != idx and i not in self.removed: yield r
def includes(self, idx: int): # Layouts included BY idx
for i, r in enumerate(self.matrix[idx]):
if i != idx and i not in self.removed and r >= _MDR_INCLUDES_MIN_RATE:
if self.matrix[i][idx] < _MDR_INCLUDES_MIN_RATE: yield i
def mdr_remove_overlap_layouts(layouts: list[MDRLayoutElement]) -> list[MDRLayoutElement]:
if not layouts:
return []
ctx = _MDR_OverlapMatrixContext(layouts)
prev_removed = -1
while len(ctx.removed) != prev_removed:
prev_removed = len(ctx.removed)
current_removed = set()
for i in range(len(layouts)):
if i in ctx.removed or i in current_removed:
continue
li = layouts[i]
pi = ctx.polys[i]
if pi is None:
current_removed.add(i)
continue
contained = False
for j in range(len(layouts)):
if i == j or j in ctx.removed or j in current_removed:
continue
if ctx.matrix[j][i] >= _MDR_INCLUDES_MIN_RATE and ctx.matrix[i][j] < _MDR_INCLUDES_MIN_RATE:
contained = True
break
if contained:
current_removed.add(i)
continue
contained_by_i = list(ctx.includes(i))
if contained_by_i:
for j in contained_by_i:
if j not in ctx.removed and j not in current_removed:
li.fragments.extend(layouts[j].fragments)
current_removed.add(j)
li.fragments.sort(key=lambda f: (f.rect.lt[1], f.rect.lt[0]))
ctx.removed.update(current_removed)
return [l for i, l in enumerate(layouts) if i not in ctx.removed]
def _mdr_split_fragments_into_lines(frags: list[MDROcrFragment]) -> Generator[list[MDROcrFragment], None, None]:
if not frags:
return
frags.sort(key=lambda f: (f.rect.lt[1], f.rect.lt[0]))
group, y_sum, h_sum = [], 0.0, 0.0
for f in frags:
_, y1, _, y2 = f.rect.wrapper
h = y2 - y1
med_y = (y1 + y2) / 2.0
if h <= 0:
continue
if not group:
group.append(f)
y_sum, h_sum = med_y, h
else:
g_len = len(group)
avg_med_y = y_sum / g_len
avg_h = h_sum / g_len
max_dev = avg_h * 0.40
if abs(med_y - avg_med_y) > max_dev:
yield group
group, y_sum, h_sum = [f], med_y, h
else:
group.append(f)
y_sum += med_y
h_sum += h
if group:
yield group
def mdr_merge_fragments_into_lines(orig_frags: list[MDROcrFragment]) -> list[MDROcrFragment]:
merged = []
for group in _mdr_split_fragments_into_lines(orig_frags):
if not group:
continue
if len(group) == 1:
merged.append(group[0])
continue
group.sort(key=lambda f: f.rect.lt[0])
min_order = min(f.order for f in group if hasattr(f, 'order')) if group else 0
texts, rank_w, txt_len = [], 0.0, 0
x1, y1, x2, y2 = float("inf"), float("inf"), float("-inf"), float("-inf")
for f in group:
fx1, fy1, fx2, fy2 = f.rect.wrapper
x1, y1, x2, y2 = min(x1, fx1), min(y1, fy1), max(x2, fx2), max(y2, fy2)
t = f.text
l = len(t)
if l > 0:
texts.append(t)
rank_w += f.rank * l
txt_len += l
if txt_len == 0:
continue
m_txt = " ".join(texts)
m_rank = rank_w / txt_len if txt_len > 0 else 0.0
m_rect = MDRRectangle(lt=(x1, y1), rt=(x2, y1), lb=(x1, y2), rb=(x2, y2))
merged.append(MDROcrFragment(order=min_order, text=m_txt, rank=m_rank, rect=m_rect))
merged.sort(key=lambda f: (f.order, f.rect.lt[1], f.rect.lt[0]))
for i, f in enumerate(merged):
f.order = i
return merged
# --- MDR Layout Processing ---
_MDR_CORRECTION_MIN_OVERLAP = 0.5
def mdr_correct_layout_fragments(ocr_engine: 'MDROcrEngine', source_img: Image, layout: MDRLayoutElement):
# --- START OF FIX ---
if not layout.fragments:
# If the layout has no fragments to begin with, there's nothing to correct.
# Attempting to crop and OCR an empty layout region is unnecessary and can lead to errors.
# print(f"Correct: Layout {type(layout.cls).__name__} has no initial fragments. Skipping OCR correction.") # Optional: for debugging
return
# --- END OF FIX ---
try:
x1, y1, x2, y2 = layout.rect.wrapper
margin = 5
# Ensure crop_box dimensions are valid before cropping
crop_x1 = max(0, round(x1) - margin)
crop_y1 = max(0, round(y1) - margin)
crop_x2 = min(source_img.width, round(x2) + margin)
crop_y2 = min(source_img.height, round(y2) + margin)
if crop_x1 >= crop_x2 or crop_y1 >= crop_y2: # If crop dimensions are invalid/empty
print(
f"Correct: Crop box for layout {type(layout.cls).__name__} is invalid/empty ({crop_x1},{crop_y1},{crop_x2},{crop_y2}). Skipping OCR correction.")
return
cropped = source_img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
off_x, off_y = crop_x1, crop_y1
except Exception as e:
print(f"Correct: Crop error for layout {type(layout.cls).__name__}: {e}")
return
# Additional check: if cropped image is too small for OCR
if cropped.width < 5 or cropped.height < 5: # Arbitrary small threshold
print(
f"Correct: Cropped image for layout {type(layout.cls).__name__} is too small ({cropped.width}x{cropped.height}). Skipping OCR correction.")
return
try:
# Ensure conversion to RGB before converting to NumPy array
cropped_np = np.array(cropped.convert("RGB"))[:, :, ::-1] # BGR for OpenCV-based OCR
new_frags_local = list(ocr_engine.find_text_fragments(cropped_np))
except Exception as e:
print(f"Correct: OCR error during correction for layout {type(layout.cls).__name__}: {e}")
# If OCR fails, we should probably keep the original fragments, if any.
# The current logic below will do this if new_frags_local is empty.
return # Exit if OCR itself fails catastrophically
new_frags_global = []
# ... (rest of the function remains the same) ...
for f in new_frags_local:
r = f.rect
lt, rt, lb, rb = r.lt, r.rt, r.lb, r.rb
f.rect = MDRRectangle(lt=(lt[0] + off_x, lt[1] + off_y), rt=(rt[0] + off_x, rt[1] + off_y),
lb=(lb[0] + off_x, lb[1] + off_y), rb=(rb[0] + off_x, rb[1] + off_y))
new_frags_global.append(f)
orig_frags = layout.fragments # These are the fragments that existed before this function call
matched, unmatched_orig = [], []
used_new = set()
# If new_frags_global is empty (e.g. OCR found nothing in the cropped region),
# then all orig_frags will go into unmatched_orig, and layout.fragments will be restored to orig_frags.
# This is generally fine.
for i, orig_f in enumerate(orig_frags):
best_j, best_rate = -1, -1.0
try:
poly_o = Polygon(orig_f.rect)
except:
continue
if not poly_o.is_valid:
continue
for j, new_f in enumerate(new_frags_global):
if j in used_new:
continue
try:
poly_n = Polygon(new_f.rect)
except:
continue
if not poly_n.is_valid:
continue
try:
inter = poly_o.intersection(poly_n)
union = poly_o.union(poly_n)
except:
continue
rate = inter.area / union.area if union.area > 1e-6 else 0.0
if rate > _MDR_CORRECTION_MIN_OVERLAP and rate > best_rate:
best_rate = rate
best_j = j
if best_j != -1:
matched.append((orig_f, new_frags_global[best_j]))
used_new.add(best_j)
else:
unmatched_orig.append(orig_f)
unmatched_new = [f for j, f in enumerate(new_frags_global) if j not in used_new]
final = [n if n.rank >= o.rank else o for o, n in matched]
final.extend(unmatched_orig)
final.extend(unmatched_new)
layout.fragments = final
if layout.fragments: # Only sort if there are fragments
layout.fragments.sort(key=lambda f: (f.rect.lt[1], f.rect.lt[0]))
# --- MDR OCR Engine ---
_MDR_OCR_MODELS = {"det": ("ppocr_onnx", "model", "det_model", "en_PP-OCRv3_det_infer.onnx"),
"cls": ("ppocr_onnx", "model", "cls_model", "ch_ppocr_mobile_v2.0_cls_infer.onnx"),
"rec": ("ppocr_onnx", "model", "rec_model", "en_PP-OCRv3_rec_infer.onnx"),
"keys": ("ppocr_onnx", "ppocr", "utils", "dict", "en_dict.txt")}
_MDR_OCR_URL_BASE = "https://raw.githubusercontent.com/Kazuhito00/PaddleOCR-ONNX-Sample/main/"
@dataclass
class _MDR_ONNXParams:
# Attributes without default values
use_gpu: bool
det_model_dir: str
cls_model_dir: str
rec_model_dir: str
rec_char_dict_path: str
# Attributes with default values (Group 1)
use_angle_cls: bool = True
rec_image_shape: str = "3,48,256"
cls_image_shape: str = "3,48,192"
cls_batch_num: int = 6
cls_thresh: float = 0.9
label_list: List[str] = field(default_factory=lambda: ['0', '180'])
# Attributes with default values (Group 2 - Detection)
det_algorithm: str = "DB"
det_limit_side_len: int = 1280
det_limit_type: str = 'min'
det_db_thresh: float = 0.3
det_db_box_thresh: float = 0.6
det_db_unclip_ratio: float = 1.5
use_dilation: bool = False
det_db_score_mode: str = 'fast'
det_box_type: str = 'quad'
# Attributes with default values (Group 3 - Recognition & General)
rec_batch_num: int = 6
drop_score: float = 0.5
rec_algorithm: str = "SVTR_LCNet"
use_space_char: bool = True
# Attributes with default values (Group 4 - Output & Logging)
save_crop_res: bool = False
crop_res_save_dir: str = "./output/mdr_crop_res"
show_log: bool = False
use_onnx: bool = True
class MDROcrEngine:
"""Handles OCR detection and recognition using ONNX models."""
def __init__(self, device: Literal["cpu", "cuda"], model_dir_path: str):
self._device = device;
self._model_dir = mdr_ensure_directory(model_dir_path)
self._text_system: _MDR_TextSystem | None = None;
self._onnx_params: _MDR_ONNXParams | None = None
self._ensure_models();
self._get_system() # Init on creation
def _ensure_models(self):
for key, parts in _MDR_OCR_MODELS.items():
fp = Path(self._model_dir) / Path(*parts)
if not fp.exists(): print(f"Downloading MDR OCR model: {fp.name}..."); url = _MDR_OCR_URL_BASE + "/".join(
parts); mdr_download_model(url, fp)
def _get_system(self) -> _MDR_TextSystem | None:
if self._text_system is None:
paths = {k: str(Path(self._model_dir) / Path(*p)) for k, p in _MDR_OCR_MODELS.items()}
# In MDROcrEngine._get_system()
self._onnx_params = _MDR_ONNXParams(
use_gpu=(self._device == "cuda"),
det_model_dir=paths["det"],
cls_model_dir=paths["cls"],
rec_model_dir=paths["rec"],
rec_char_dict_path=paths["keys"],
# much lower thresholds so we actually get some candidate masks:
det_db_thresh=0.3,
det_db_box_thresh=0.5,
drop_score=0.1,
use_angle_cls=False,
)
try:
self._text_system = _MDR_TextSystem(self._onnx_params); print(f"MDR OCR System initialized.")
except Exception as e:
print(f"ERROR initializing MDR OCR System: {e}"); self._text_system = None
return self._text_system
# In class MDROcrEngine:
def find_text_fragments(self, image_np: np.ndarray) -> Generator[MDROcrFragment, None, None]:
"""Finds and recognizes text fragments in a NumPy image (BGR)."""
system = self._get_system()
if system is None:
print(" DEBUG OCR Engine: MDR OCR System unavailable. No fragments will be found.")
return
img_for_system = self._preprocess(image_np)
print(f" DEBUG OCR Engine: Image preprocessed for TextSystem. Shape: {img_for_system.shape}")
try:
boxes, recs = system(img_for_system)
except Exception as e:
print(f" DEBUG OCR Engine: Error during TextSystem prediction: {e}")
import traceback
traceback.print_exc()
return
if not boxes or not recs:
print(
f" DEBUG OCR Engine: TextSystem returned no boxes ({len(boxes) if boxes is not None else 'None'}) or no recs ({len(recs) if recs is not None else 'None'}). No fragments generated.")
return
if len(boxes) != len(recs):
print(
f" DEBUG OCR Engine: Mismatch between boxes ({len(boxes)}) and recs ({len(recs)}) from TextSystem. This is problematic. No fragments generated.")
return
print(
f" DEBUG OCR Engine: TextSystem returned {len(boxes)} boxes and {len(recs)} recognition results. Converting to MDROcrFragment.")
fragments_generated_count = 0
for i, (box_pts, rec_tuple) in enumerate(zip(boxes, recs)):
if not isinstance(rec_tuple, (list, tuple)) or len(rec_tuple) != 2:
print(f" DEBUG OCR Engine: Rec item {i} is not a valid (text, score) tuple: {rec_tuple}. Skipping.")
continue
txt, conf = rec_tuple
if not txt or mdr_is_whitespace(txt):
# print(f" DEBUG OCR Engine: Fragment {i} has empty/whitespace text after system call. Text: '{txt}'. Skipping.") # Already logged in TextSystem
continue
try:
pts = [(float(p[0]), float(p[1])) for p in box_pts]
if len(pts) == 4:
r = MDRRectangle(lt=pts[0], rt=pts[1], lb=pts[2], rb=pts[3])
if r.is_valid and r.area > 1:
yield MDROcrFragment(order=-1, text=txt, rank=float(conf), rect=r)
fragments_generated_count += 1
# else:
# print(f" DEBUG OCR Engine: Fragment {i} has invalid/small rectangle. Area: {r.area:.2f}. Valid: {r.is_valid}. Skipping.")
# else:
# print(f" DEBUG OCR Engine: Fragment {i} box_pts not length 4: {len(pts)}. Skipping.")
except Exception as e_frag:
print(f" DEBUG OCR Engine: Error creating MDROcrFragment for item {i}: {e_frag}")
continue
print(f" DEBUG OCR Engine: Generated {fragments_generated_count} MDROcrFragment objects.")
def _preprocess(self, img: np.ndarray) -> np.ndarray:
if len(img.shape) == 3 and img.shape[2] == 4:
a = img[:, :, 3] / 255.0
bg = (255, 255, 255)
new = np.zeros_like(img[:, :, :3])
[setattr(new[:, :, i], 'flags.writeable', True) for i in range(3)]
[np.copyto(new[:, :, i], (bg[i] * (1 - a) + img[:, :, i] * a)) for i in range(3)]
img = new.astype(np.uint8)
elif len(img.shape) == 2:
img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
elif not (len(img.shape) == 3 and img.shape[2] == 3):
raise ValueError("Unsupported image format")
return img
# --- MDR Layout Reading Internals ---
_MDR_MAX_LEN = 510;
_MDR_CLS_ID = 0;
_MDR_SEP_ID = 2;
_MDR_PAD_ID = 1
def mdr_boxes_to_reader_inputs(boxes: List[List[int]], max_len=_MDR_MAX_LEN) -> Dict[str, torch.Tensor]:
t_boxes = boxes[:max_len];
i_boxes = [[0, 0, 0, 0]] + t_boxes + [[0, 0, 0, 0]]
i_ids = [_MDR_CLS_ID] + [_MDR_PAD_ID] * len(t_boxes) + [_MDR_SEP_ID]
a_mask = [1] * len(i_ids);
pad_len = (max_len + 2) - len(i_ids)
if pad_len > 0: i_boxes.extend([[0, 0, 0, 0]] * pad_len); i_ids.extend([_MDR_PAD_ID] * pad_len); a_mask.extend(
[0] * pad_len)
return {"bbox": torch.tensor([i_boxes]), "input_ids": torch.tensor([i_ids]),
"attention_mask": torch.tensor([a_mask])}
def mdr_prepare_reader_inputs(inputs: Dict[str, torch.Tensor], model: LayoutLMv3ForTokenClassification) -> Dict[
str, torch.Tensor]:
return {k: v.to(model.device) if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
def mdr_parse_reader_logits(logits: torch.Tensor, length: int) -> List[int]:
print(f"mdr_parse_reader_logits: Called with logits shape: {logits.shape}, length: {length}")
if length == 0:
print("mdr_parse_reader_logits: length is 0, returning empty list.")
return []
print(f"mdr_parse_reader_logits: Attempting to slice logits with [1 : {length + 1}, :{length}]")
try:
rel_logits = logits[1: length + 1, :length]
print(f"mdr_parse_reader_logits: rel_logits shape: {rel_logits.shape}")
except IndexError as e:
print(f"mdr_parse_reader_logits: IndexError during rel_logits slicing! Error: {e}")
import traceback
traceback.print_exc()
# Depending on desired behavior, either raise or return empty/fallback
return list(range(length)) # Fallback to sequential order if slicing fails
orders = rel_logits.argmax(dim=1).tolist()
print(f"mdr_parse_reader_logits: Initial orders calculated. Count: {len(orders)}")
# ADDED: Loop safeguard
loop_count = 0
# Max loops: if N items, N^2 is a generous limit for pairwise comparisons/adjustments.
# For N=33, N^2 = 1089. For N=21, N^2 = 441. This matches the logs.
# A tighter bound might be N * (N-1) / 2 or N * some_factor.
# Let's use N * N as seen in logs, or a fixed large number if N is small.
max_loops = max(50, length * length) # Ensure at least 50 loops for small N
while True:
loop_count += 1
if loop_count > max_loops:
print(
f"mdr_parse_reader_logits: Exceeded max_loops ({max_loops}), breaking while loop to prevent infinite loop.")
break
# print(f"mdr_parse_reader_logits: While loop iteration: {loop_count}") # Can be too verbose
conflicts = defaultdict(list)
[conflicts[order].append(idx) for idx, order in enumerate(orders)]
# Filter to find actual conflicting orders (where multiple original indices map to the same target order)
conflicting_orders_map = {o: idxs for o, idxs in conflicts.items() if len(idxs) > 1}
if not conflicting_orders_map:
# print("mdr_parse_reader_logits: No conflicting orders, breaking while loop.") # Verbose
break
# Log only if there are actual conflicts to resolve
if loop_count == 1 or loop_count % 10 == 0: # Log first and every 10th iteration with conflicts
print(
f"mdr_parse_reader_logits: While loop iteration: {loop_count}. Found {len(conflicting_orders_map)} conflicting orders.")
for order_val, c_idxs in conflicting_orders_map.items():
# This logic seems to pick the one with the highest score for that conflicting order.
# It might need more sophisticated tie-breaking if scores are identical or very close.
# The original logic was:
# best_idx = -1; max_score = -float('inf')
# for c_idx in c_idxs:
# score = rel_logits[c_idx, order_val].item()
# if score > max_score: max_score = score; best_idx = c_idx
# for c_idx in c_idxs:
# if c_idx != best_idx: orders[c_idx] = -1 # Mark for re-evaluation or different assignment
# Simpler approach: keep the first one, mark others to be reassigned.
# This might not be optimal but could break cycles.
# A more robust solution might involve graph-based cycle detection or a different assignment strategy.
# For now, let's stick to a slight modification of the implied original logic:
# The one with the highest confidence for *that specific order_val* keeps it.
# Others get their order reset to their own index (diagonal) or -1 to be re-evaluated.
if not c_idxs: continue
best_c_idx_for_this_order = -1
max_confidence_for_this_order = -float('inf')
for current_c_idx in c_idxs:
confidence = rel_logits[current_c_idx, order_val].item()
if confidence > max_confidence_for_this_order:
max_confidence_for_this_order = confidence
best_c_idx_for_this_order = current_c_idx
# Now, for all conflicting indices for this 'order_val',
# if they are not the 'best_c_idx_for_this_order',
# they need a new order. A simple strategy is to make them point to themselves initially.
# Or, find their next best alternative.
for current_c_idx in c_idxs:
if current_c_idx != best_c_idx_for_this_order:
# Option 1: Reset to self (might not resolve complex cycles)
# orders[current_c_idx] = current_c_idx
# Option 2: Find next best order for this current_c_idx, excluding the conflicting 'order_val'
# Create a temporary copy of its logits row, set the conflicting order's logit to -inf
temp_logits_row = rel_logits[current_c_idx, :].clone()
temp_logits_row[order_val] = -float('inf')
orders[current_c_idx] = temp_logits_row.argmax().item()
print(
f"mdr_parse_reader_logits: While loop finished after {loop_count} iterations. Returning {len(orders)} orders.")
return orders
# --- MDR Layout Reading Engine ---
@dataclass
class _MDR_ReaderBBox: layout_index: int; fragment_index: int; virtual: bool; order: int; value: tuple[
float, float, float, float]
class MDRLayoutReader:
"""Determines reading order of layout elements using LayoutLMv3."""
def __init__(self, model_path: str):
self._model_path = model_path
self._model: LayoutLMv3ForTokenClassification | None = None
# Determine device more robustly, self._device will be 'cuda' or 'cpu'
if torch.cuda.is_available(): # Check if CUDA is actually available at runtime
self._device = "cuda"
print("MDRLayoutReader: CUDA is available. Setting device to cuda.")
else:
self._device = "cpu"
print("MDRLayoutReader: CUDA not available. Setting device to cpu.")
# In class MDRLayoutReader:
def _get_model(self) -> LayoutLMv3ForTokenClassification | None:
if self._model is None:
# MODIFIED: Use self._model_path for the layoutreader's specific cache,
# and ensure it's a directory. self._model_path is passed during MDRLayoutReader init.
layoutreader_cache_dir = Path(self._model_path) # self._model_path is like "./mdr_models/layoutreader"
mdr_ensure_directory(str(layoutreader_cache_dir)) # Ensure this specific directory exists
name = "lakshya-rawat/document-qa-model"
print(f"MDRLayoutReader: Attempting to load LayoutLMv3 model '{name}'. Cache dir: {layoutreader_cache_dir}")
try:
self._model = LayoutLMv3ForTokenClassification.from_pretrained(
name,
cache_dir=str(layoutreader_cache_dir),
local_files_only=False,
num_labels=_MDR_MAX_LEN + 1
)
self._model.to(torch.device(self._device))
self._model.eval()
print(f"MDR LayoutReader model '{name}' loaded successfully on device: {self._model.device}.")
except Exception as e:
print(f"ERROR loading MDR LayoutReader model '{name}': {e}")
import traceback
traceback.print_exc()
self._model = None
return self._model
# In class MDRLayoutReader:
def determine_reading_order(self, layouts: list[MDRLayoutElement], size: tuple[int, int]) -> list[MDRLayoutElement]:
w, h = size
if w <= 0 or h <= 0:
print("MDRLayoutReader: Invalid image size (w or h <= 0), returning layouts as is.")
return layouts
if not layouts:
print("MDRLayoutReader: No layouts to process, returning empty list.")
return []
model = self._get_model()
if model is None:
print("MDRLayoutReader: Model not available, returning layouts sorted geometrically.")
layouts.sort(key=lambda l: (l.rect.lt[1], l.rect.lt[0])) # Sort by top-left y, then x
return layouts
print("MDRLayoutReader: Preparing bboxes...")
# bbox_list contains _MDR_ReaderBBox objects, each with .value = (x0,y0,x1,y1) in original pixels
bbox_list = self._prepare_bboxes(layouts, w, h)
if bbox_list is None or len(bbox_list) == 0:
print("MDRLayoutReader: No bboxes prepared from layouts, returning layouts as is (sorted geometrically).")
layouts.sort(key=lambda l: (l.rect.lt[1], l.rect.lt[0]))
return layouts
print(f"MDRLayoutReader: Prepared {len(bbox_list)} bboxes.")
# --- START: SCALING LOGIC as in the prompt ---
scaled_bboxes: list[list[int]] = []
if w > 0 and h > 0:
for bbox_item in bbox_list:
x0, y0, x1, y1 = bbox_item.value
x0_c = max(0.0, min(x0, float(w)))
y0_c = max(0.0, min(y0, float(h)))
x1_c = max(0.0, min(x1, float(w)))
y1_c = max(0.0, min(y1, float(h)))
scaled_x0 = max(0, min(1000, int(1000 * x0_c / w)))
scaled_y0 = max(0, min(1000, int(1000 * y0_c / h)))
scaled_x1 = max(scaled_x0, min(1000, int(1000 * x1_c / w)))
scaled_y1 = max(scaled_y0, min(1000, int(1000 * y1_c / h)))
scaled_bboxes.append([scaled_x0, scaled_y0, scaled_x1, scaled_y1])
else:
# This branch should ideally not be reached due to the initial w,h check
print(
"MDRLayoutReader: Warning - Invalid image dimensions (w or h is zero) for scaling bboxes. Cannot determine reading order.")
layouts.sort(key=lambda l: (l.rect.lt[1], l.rect.lt[0]))
return layouts
# --- END: SCALING LOGIC ---
if not scaled_bboxes: # Handles if bbox_list was empty
print(
"MDRLayoutReader: No scaled bboxes available after scaling step. Returning geometrically sorted layouts.")
layouts.sort(key=lambda l: (l.rect.lt[1], l.rect.lt[0]))
return layouts
# --- START OF FIX ---
# Check if scaled_bboxes are problematic (e.g., all identical and degenerate)
bypass_model_inference = False
if len(scaled_bboxes) > 0:
num_s_bboxes = len(scaled_bboxes)
# Check if all scaled_bboxes are identical to the first one
first_s_bbox_str = str(scaled_bboxes[0])
all_identical = all(str(s_b) == first_s_bbox_str for s_b in scaled_bboxes)
if all_identical:
# Check if this identical box is degenerate (zero width or height)
s_x0, s_y0, s_x1, s_y1 = scaled_bboxes[0]
if (s_x1 - s_x0 == 0) or (s_y1 - s_y0 == 0):
bypass_model_inference = True
print("MDRLayoutReader: All scaled bboxes are identical and degenerate. Bypassing LayoutLMv3.")
if not bypass_model_inference and num_s_bboxes > 1: # Check for high proportion of degenerate if not all identical
degenerate_count = 0
for s_b in scaled_bboxes:
if (s_b[2] - s_b[0] == 0) or (s_b[3] - s_b[1] == 0): # x1-x0 or y1-y0
degenerate_count += 1
# If, for example, more than 90% of bboxes are degenerate
if degenerate_count / num_s_bboxes > 0.9:
bypass_model_inference = True
print(
f"MDRLayoutReader: High percentage ({degenerate_count / num_s_bboxes * 100:.1f}%) of scaled bboxes are degenerate. Bypassing LayoutLMv3.")
if bypass_model_inference:
print("MDRLayoutReader: Applying fallback sequential order due to problematic scaled_bboxes.")
# Assign sequential order based on _prepare_bboxes's sort (y, then x)
for i in range(len(bbox_list)):
bbox_list[i].order = i
# Use _apply_order to apply this simple sequential ordering
result_layouts = self._apply_order(layouts, bbox_list)
return result_layouts
# --- END OF FIX ---
orders: list[int] = []
try:
with torch.no_grad():
print("MDRLayoutReader: Creating reader inputs...")
inputs = mdr_boxes_to_reader_inputs(scaled_bboxes)
print("MDRLayoutReader: Preparing inputs for model device...")
inputs = mdr_prepare_reader_inputs(inputs, model)
print("MDRLayoutReader: Running model inference...")
logits = model(**inputs).logits.cpu().squeeze(0)
print("MDRLayoutReader: Model inference complete. Parsing logits...")
orders = mdr_parse_reader_logits(logits, len(bbox_list)) # len(bbox_list) is correct here
print(f"MDRLayoutReader: Logits parsed. Orders count: {len(orders)}")
if len(orders) == len(bbox_list):
for i, order_val in enumerate(orders):
bbox_list[i].order = order_val
else:
print(
f"MDRLayoutReader: Warning - Mismatch between orders ({len(orders)}) and bbox_list ({len(bbox_list)}). Using sequential order.")
for i in range(len(bbox_list)):
bbox_list[i].order = i
except Exception as e:
print(f"MDR LayoutReader prediction error: {e}")
import traceback
traceback.print_exc()
for i in range(len(bbox_list)):
bbox_list[i].order = i
print("MDRLayoutReader: Applying fallback sequential order due to error...")
result_layouts = self._apply_order(layouts, bbox_list)
return result_layouts
print("MDRLayoutReader: Applying order...")
result_layouts = self._apply_order(layouts, bbox_list)
print("MDRLayoutReader: Order applied. Returning layouts.")
return result_layouts
def _prepare_bboxes(self, layouts: list[MDRLayoutElement], w: int, h: int) -> list[_MDR_ReaderBBox] | None:
line_h = self._estimate_line_h(layouts)
bbox_list = []
for i, l in enumerate(layouts):
if l.cls == MDRLayoutClass.PLAIN_TEXT and l.fragments:
[bbox_list.append(_MDR_ReaderBBox(i, j, False, -1, f.rect.wrapper)) for j, f in enumerate(l.fragments)]
else:
bbox_list.extend(self._gen_virtual(l, i, line_h, w, h))
if len(bbox_list) > _MDR_MAX_LEN:
print(f"Too many boxes ({len(bbox_list)}>{_MDR_MAX_LEN})")
return None
bbox_list.sort(key=lambda b: (b.value[1], b.value[0]))
return bbox_list
# In class MDRLayoutReader
def _apply_order(self, original_layouts_list: list[MDRLayoutElement],
ordered_bbox_list_with_final_orders: list[_MDR_ReaderBBox]) -> list[MDRLayoutElement]:
# layout_map: maps original layout index to a list of its _MDR_ReaderBBox objects (which now have final .order)
layout_map = defaultdict(list)
for bbox_item in ordered_bbox_list_with_final_orders:
layout_map[bbox_item.layout_index].append(bbox_item)
# Determine the new order of layouts themselves
# The .order in bbox_item here is the *within-layout* order for fragments/virtual boxes.
# We need the median of these *final reading orders* to sort the layouts.
# The .order attribute of _MDR_ReaderBBox should have been updated by mdr_parse_reader_logits.
layout_median_orders = []
for original_layout_idx, bboxes_for_this_layout in layout_map.items():
if bboxes_for_this_layout: # Ensure there are bboxes
# Each bbox_item.order here is its final reading order determined by LayoutLM
median_order_for_layout = self._median([b.order for b in bboxes_for_this_layout])
layout_median_orders.append((original_layout_idx, median_order_for_layout))
layout_median_orders.sort(key=lambda x: x[1]) # Sort layouts by their median reading order
# Create the new list of sorted layouts
# Important: We are reordering the original_layouts_list.
# The fragment objects within these layouts are the ones we need to sort.
final_sorted_layouts = [original_layouts_list[idx] for idx, _ in layout_median_orders]
# Now, sort fragments within each layout
# nfo (next fragment order) is a global counter for the absolute order of fragments across all layouts
nfo = 0
for layout_obj in final_sorted_layouts:
if not layout_obj.fragments: # Skip layouts with no fragments
continue
# Get the _MDR_ReaderBBox items that correspond to this specific layout_obj
# We need the original index of layout_obj from the input `original_layouts_list`
# This assumes original_layouts_list has not been reordered yet by this function.
try:
# Find the original index of the current layout_obj
# This is safe if original_layouts_list is the list passed into this function
original_idx_of_current_layout = original_layouts_list.index(layout_obj)
except ValueError:
# This should not happen if layout_obj came from original_layouts_list via layout_median_orders
print(
f" ERROR: Could not find layout in original list during fragment sort. Skipping fragment sort for this layout.")
# Assign sequential order as a fallback for fragments in this layout
for i_frag, frag_in_layout in enumerate(layout_obj.fragments):
frag_in_layout.order = nfo + i_frag
nfo += len(layout_obj.fragments)
continue
# Get the _MDR_ReaderBBox items for this layout, which contain the final .order for each fragment_index
reader_bboxes_for_this_layout = [
b for b in layout_map[original_idx_of_current_layout] if not b.virtual
]
if reader_bboxes_for_this_layout:
# Create a map from original_fragment_index to its new_reading_order
frag_idx_to_new_order_map = {
b.fragment_index: b.order for b in reader_bboxes_for_this_layout
}
# Sort the actual MDROcrFragment objects in layout_obj.fragments
# The key for sorting should use the original index of the fragment
# to look up its new_reading_order from the map.
# We assume layout_obj.fragments has not been reordered yet by this function for this layout.
# We need to sort a list of (fragment_object, original_index) tuples first.
fragments_with_original_indices = list(enumerate(layout_obj.fragments))
fragments_with_original_indices.sort(
key=lambda item: frag_idx_to_new_order_map.get(item[0], float('inf')) # item[0] is original index
)
# Reconstruct the sorted list of fragment objects
layout_obj.fragments = [item[1] for item in
fragments_with_original_indices] # item[1] is fragment object
else: # No corresponding reader_bboxes (e.g. layout was all virtual or had no frags initially)
# or if the layout was created as a fallback and has no reader_bboxes.
print(
f" LayoutReader ApplyOrder: No reader_bboxes for layout (orig_idx {original_idx_of_current_layout}). Sorting frags geometrically.")
layout_obj.fragments.sort(key=lambda f: (f.rect.lt[1], f.rect.lt[0])) # Fallback geometric sort
# Assign the final absolute order (nfo)
for frag in layout_obj.fragments:
frag.order = nfo
nfo += 1
return final_sorted_layouts
def _estimate_line_h(self, layouts: list[MDRLayoutElement]) -> float:
heights = [f.rect.size[1] for l in layouts for f in l.fragments if f.rect.size[1] > 0]
return self._median(heights) if heights else 15.0
def _gen_virtual(self, l: MDRLayoutElement, l_idx: int, line_h: float, pw: int, ph: int) -> Generator[
_MDR_ReaderBBox, None, None]:
x0, y0, x1, y1 = l.rect.wrapper
lh = y1 - y0
lw = x1 - x0
if lh <= 0 or lw <= 0 or line_h <= 0:
yield _MDR_ReaderBBox(l_idx, -1, True, -1, (x0, y0, x1, y1))
return
lines = 1
if lh > line_h * 1.5:
if lh <= ph * 0.25 or lw >= pw * 0.5:
lines = 3
elif lw > pw * 0.25:
lines = 3 if lw > pw * 0.4 else 2
elif lw <= pw * 0.25:
lines = max(1, int(lh / (line_h * 1.5))) if lh / lw > 1.5 else 2
else:
lines = max(1, int(round(lh / line_h)))
lines = max(1, lines)
act_line_h = lh / lines
cur_y = y0
for i in range(lines):
ly0 = max(0, min(ph, cur_y))
ly1 = max(0, min(ph, cur_y + act_line_h))
lx0 = max(0, min(pw, x0))
lx1 = max(0, min(pw, x1))
if ly1 > ly0 and lx1 > lx0:
yield _MDR_ReaderBBox(l_idx, -1, True, -1, (lx0, ly0, lx1, ly1))
cur_y += act_line_h
def _median(self, nums: list[float | int]) -> float:
if not nums:
return 0.0
s_nums = sorted(nums)
n = len(s_nums)
return float(s_nums[n // 2]) if n % 2 == 1 else float((s_nums[n // 2 - 1] + s_nums[n // 2]) / 2.0)
# --- MDR LaTeX Extractor ---
class MDRLatexExtractor:
"""Extracts LaTeX from formula images using pix2tex."""
def __init__(self, model_path: str):
self._model_path = model_path;
self._model: LatexOCR | None = None
self._device = "cuda" if torch.cuda.is_available() else "cpu"
def extract(self, image: Image) -> str | None:
if LatexOCR is None: return None;
image = mdr_expand_image(image, 0.1)
model = self._get_model()
if model is None: return None;
try:
with torch.no_grad():
img_rgb = image.convert('RGB') if image.mode != 'RGB' else image; latex = model(
img_rgb); return latex if latex else None
except Exception as e:
print(f"MDR LaTeX error: {e}"); return None
def _get_model(self) -> LatexOCR | None:
if self._model is None and LatexOCR is not None:
mdr_ensure_directory(self._model_path)
wp = Path(self._model_path) / "weights.pth"
rp = Path(self._model_path) / "image_resizer.pth"
cp = Path(self._model_path) / "config.yaml"
if not wp.exists() or not rp.exists():
print("Downloading MDR LaTeX models...")
self._download()
if not cp.exists():
print(f"Warn: MDR LaTeX config not found {self._model_path}")
try:
args = Munch({"config": str(cp), "checkpoint": str(wp), "device": self._device,
"no_cuda": self._device == "cuda", "no_resize": False, "temperature": 0.0})
self._model = LatexOCR(args)
print(f"MDR LaTeX loaded on {self._device}.")
except Exception as e:
print(f"ERROR initializing MDR LatexOCR: {e}")
self._model = None
return self._model
def _download(self):
tag = "v0.0.1"
base = f"https://github.com/lukas-blecher/LaTeX-OCR/releases/download/{tag}/"
files = {"weights.pth": base + "weights.pth", "image_resizer.pth": base + "image_resizer.pth"}
mdr_ensure_directory(self._model_path)
[mdr_download_model(url, Path(self._model_path) / name) for name, url in files.items() if
not (Path(self._model_path) / name).exists()]
# --- MDR Table Parser ---
MDRTableOutputFormat = Literal["latex", "markdown", "html"]
class MDRTableParser:
"""Parses table structure/content from images using StructTable model."""
def __init__(self, device: Literal["cpu", "cuda"], model_path: str):
self._model: Any | None = None;
self._model_path = mdr_ensure_directory(model_path)
self._device = device if torch.cuda.is_available() and device == "cuda" else "cpu"
self._disabled = self._device == "cuda"
if self._disabled: print("Warning: MDR Table parsing requires CUDA. Disabled.")
def parse_table_image(self, image: Image, format: MDRTableLayoutParsedFormat) -> str | None:
if self._disabled: return None;
fmt: MDRTableOutputFormat | None = None
if format == MDRTableLayoutParsedFormat.LATEX:
fmt = "latex"
elif format == MDRTableLayoutParsedFormat.MARKDOWN:
fmt = "markdown"
elif format == MDRTableLayoutParsedFormat.HTML:
fmt = "html"
else:
return None
image = mdr_expand_image(image, 0.05)
model = self._get_model()
if model is None: return None;
try:
img_rgb = image.convert('RGB') if image.mode != 'RGB' else image
with torch.no_grad():
results = model([img_rgb], output_format=fmt)
return results[0] if results else None
except Exception as e:
print(f"MDR Table parsing error: {e}"); return None
def _get_model(self):
if self._model is None and not self._disabled:
try:
from struct_eqtable import build_model # Dynamic import
name = "U4R/StructTable-InternVL2-1B";
local = any(Path(self._model_path).iterdir())
print(f"Loading MDR StructTable model '{name}'...")
model = build_model(model_ckpt=name, max_new_tokens=1024, max_time=30, lmdeploy=False, flash_attn=True,
batch_size=1, cache_dir=self._model_path, local_files_only=local)
self._model = model.to(self._device);
print(f"MDR StructTable loaded on {self._device}.")
except ImportError:
print("ERROR: struct_eqtable not found."); self._disabled = True; self._model = None
except Exception as e:
print(f"ERROR loading MDR StructTable: {e}"); self._model = None
return self._model
# --- MDR Image Optimizer ---
_MDR_TINY_ROTATION = 0.005
@dataclass
class _MDR_RotationContext: to_origin: MDRRotationAdjuster; to_new: MDRRotationAdjuster; fragment_origin_rectangles: \
list[MDRRectangle]
class MDRImageOptimizer:
"""Handles image rotation detection and coordinate adjustments."""
def __init__(self, raw_image: Image, adjust_points: bool):
self._raw = raw_image;
self._image = raw_image;
self._adjust_points = adjust_points
self._fragments: list[MDROcrFragment] = [];
self._rotation: float = 0.0;
self._rot_ctx: _MDR_RotationContext | None = None
@property
def image(self) -> Image:
return self._image
@property
def adjusted_image(self) -> Image | None:
return self._image if self._rot_ctx is not None else None
@property
def rotation(self) -> float:
return self._rotation
@property
def image_np(self) -> np.ndarray:
img_rgb = np.array(self._raw.convert("RGB")); return cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
def receive_fragments(self, fragments: list[MDROcrFragment]):
self._fragments = fragments
if not fragments:
return
self._rotation = mdr_calculate_image_rotation(fragments)
if abs(self._rotation) < _MDR_TINY_ROTATION:
self._rotation = 0.0
return
orig_sz = self._raw.size
try:
self._image = self._raw.rotate(-np.degrees(self._rotation), resample=PILResampling.BICUBIC,
fillcolor=(255, 255, 255), expand=True)
except Exception as e:
print(f"Optimizer rotation error: {e}")
self._rotation = 0.0
self._image = self._raw
return
new_sz = self._image.size
self._rot_ctx = _MDR_RotationContext(
fragment_origin_rectangles=[f.rect for f in fragments],
to_new=MDRRotationAdjuster(orig_sz, new_sz, self._rotation, False),
to_origin=MDRRotationAdjuster(orig_sz, new_sz, self._rotation, True))
adj = self._rot_ctx.to_new
[setattr(f, 'rect',
MDRRectangle(lt=adj.adjust(r.lt), rt=adj.adjust(r.rt), lb=adj.adjust(r.lb), rb=adj.adjust(r.rb))) for f
in fragments if (r := f.rect)]
def finalize_layout_coords(self, layouts: list[MDRLayoutElement]):
if self._rot_ctx is None: # If no rotation context, nothing to do
return
if not self._adjust_points: # If we are NOT adjusting points back to original,
# then restore original fragment rectangles
if len(self._fragments) == len(self._rot_ctx.fragment_origin_rectangles):
for f, orig_r in zip(self._fragments, self._rot_ctx.fragment_origin_rectangles):
f.rect = orig_r
# And adjust layout rectangles to origin coordinates
adj = self._rot_ctx.to_origin
for l in layouts:
if (r := l.rect): # Check if rect exists
l.rect = MDRRectangle(lt=adj.adjust(r.lt), rt=adj.adjust(r.rt), lb=adj.adjust(r.lb),
rb=adj.adjust(r.rb))
# If self._adjust_points is True, the coordinates (already adjusted to the rotated image) are kept as is.
# No further action is needed for the True case here, as the adjustments happened in receive_fragments.
# --- MDR Image Clipping ---
def mdr_clip_from_image(image: Image, rect: MDRRectangle, wrap_w: float = 0.0, wrap_h: float = 0.0) -> Image:
"""Clips a potentially rotated rectangle from an image."""
try:
h_rot, _ = mdr_calculate_rectangle_rotation(rect)
avg_w, avg_h = rect.size
if avg_w <= 0 or avg_h <= 0:
return new_image("RGB", (1, 1), (255, 255, 255))
tx, ty = rect.lt
trans_orig = np.array([[1, 0, -tx], [0, 1, -ty], [0, 0, 1]])
cos_r = cos(-h_rot)
sin_r = sin(-h_rot)
rot = np.array([[cos_r, -sin_r, 0], [sin_r, cos_r, 0], [0, 0, 1]])
pad_dx = wrap_w / 2.0
pad_dy = wrap_h / 2.0
trans_pad = np.array([[1, 0, pad_dx], [0, 1, pad_dy], [0, 0, 1]])
matrix = trans_pad @ rot @ trans_orig
try:
inv_matrix = np.linalg.inv(matrix)
except np.linalg.LinAlgError:
x0, y0, x1, y1 = rect.wrapper
return image.crop((round(x0), round(y0), round(x1), round(y1)))
p_mat = (
inv_matrix[0, 0], inv_matrix[0, 1], inv_matrix[0, 2], inv_matrix[1, 0], inv_matrix[1, 1], inv_matrix[1, 2])
out_w = ceil(avg_w + wrap_w)
out_h = ceil(avg_h + wrap_h)
return image.transform((out_w, out_h), PILTransform.AFFINE, p_mat, PILResampling.BICUBIC,
fillcolor=(255, 255, 255))
except Exception as e:
print(f"MDR Clipping error: {e}")
return new_image("RGB", (10, 10), (255, 255, 255))
def mdr_clip_layout(res: MDRExtractionResult, layout: MDRLayoutElement, wrap_w: float = 0.0,
wrap_h: float = 0.0) -> Image:
"""Clips a layout region from the MDRExtractionResult image."""
img = res.adjusted_image if res.adjusted_image else res.extracted_image
return mdr_clip_from_image(img, layout.rect, wrap_w, wrap_h)
# --- MDR Debug Plotting ---
_MDR_FRAG_COLOR = (0x49, 0xCF, 0xCB, 200);
_MDR_LAYOUT_COLORS = {MDRLayoutClass.TITLE: (0x0A, 0x12, 0x2C, 255), MDRLayoutClass.PLAIN_TEXT: (0x3C, 0x67, 0x90, 255),
MDRLayoutClass.ABANDON: (0xC0, 0xBB, 0xA9, 180), MDRLayoutClass.FIGURE: (0x5B, 0x91, 0x3C, 255),
MDRLayoutClass.FIGURE_CAPTION: (0x77, 0xB3, 0x54, 255),
MDRLayoutClass.TABLE: (0x44, 0x17, 0x52, 255),
MDRLayoutClass.TABLE_CAPTION: (0x81, 0x75, 0xA0, 255),
MDRLayoutClass.TABLE_FOOTNOTE: (0xEF, 0xB6, 0xC9, 255),
MDRLayoutClass.ISOLATE_FORMULA: (0xFA, 0x38, 0x27, 255),
MDRLayoutClass.FORMULA_CAPTION: (0xFF, 0x9D, 0x24, 255)};
_MDR_DEFAULT_COLOR = (0x80, 0x80, 0x80, 255);
_MDR_RGBA = tuple[int, int, int, int]
def mdr_plot_layout(image: Image, layouts: Iterable[MDRLayoutElement]) -> None:
"""Draws layout and fragment boxes onto an image for debugging."""
if not layouts: return;
try:
l_font = load_default(size=25)
f_font = load_default(size=15) # Not used currently, but kept for potential future use
draw = ImageDraw.Draw(image, mode="RGBA")
except Exception as e:
print(f"MDR Plot init error: {e}")
return
def _draw_num(pos: MDRPoint, num: int, font: FreeTypeFont, color: _MDR_RGBA):
try:
x, y = pos
txt = str(num)
txt_pos = (round(x) + 3, round(y) + 1)
bbox = draw.textbbox(txt_pos, txt, font=font)
bg_rect = (bbox[0] - 2, bbox[1] - 1, bbox[2] + 2, bbox[3] + 1)
bg_color = (color[0], color[1], color[2], 180)
draw.rectangle(bg_rect, fill=bg_color)
draw.text(txt_pos, txt, font=font, fill=(255, 255, 255, 255))
except Exception as e:
print(f"MDR Draw num error: {e}")
for i, l in enumerate(layouts):
try:
l_color = _MDR_LAYOUT_COLORS.get(l.cls, _MDR_DEFAULT_COLOR)
draw.polygon([p for p in l.rect], outline=l_color, width=3)
_draw_num(l.rect.lt, i + 1, l_font, l_color)
except Exception as e:
print(f"MDR Layout draw error: {e}")
for l in layouts:
for f in l.fragments:
try:
draw.polygon([p for p in f.rect], outline=_MDR_FRAG_COLOR, width=1)
except Exception as e:
print(f"MDR Fragment draw error: {e}")
# --- MDR Extraction Engine ---
class MDRExtractionEngine:
"""Core engine for extracting structured information from a document image."""
def __init__(self, model_dir_path: str, device: Literal["cpu", "cuda"] = "cpu", ocr_for_each_layouts: bool = True,
extract_formula: bool = True, extract_table_format: MDRTableLayoutParsedFormat | None = None):
self._model_dir = model_dir_path # Base directory for all models
self._device = device if torch.cuda.is_available() else "cpu"
self._ocr_each = ocr_for_each_layouts;
self._ext_formula = extract_formula;
self._ext_table = extract_table_format
self._yolo: YOLOv10 | None = None
# Initialize sub-modules, passing the main model_dir_path
self._ocr_engine = MDROcrEngine(device=self._device, model_dir_path=os.path.join(self._model_dir, "onnx_ocr"))
self._table_parser = MDRTableParser(device=self._device,
model_path=os.path.join(self._model_dir, "struct_eqtable"))
self._latex_extractor = MDRLatexExtractor(model_path=os.path.join(self._model_dir, "latex"))
self._layout_reader = MDRLayoutReader(model_path=os.path.join(self._model_dir, "layoutreader"))
print(f"MDR Extraction Engine initialized on device: {self._device}")
# In class MDRExtractionEngine:
def _get_yolo_model(self) -> Any | None: # Return type will be ultralytics.YOLO
"""Loads the YOLOv10b-DocLayNet layout detection model using ultralytics.YOLO."""
if self._yolo is None:
# Using hantian/yolo-doclaynet (or ppaanngggg if that's the one you have the .pt for)
# Ensure these match the model you intend to use
repo_id = "hantian/yolo-doclaynet"
filename = "yolov10b-doclaynet.pt" # Or the exact .pt filename from the repo
yolo_cache_dir = Path(self._model_dir) / "yolo_hf_cache_doclaynet"
mdr_ensure_directory(str(yolo_cache_dir))
print(f"Attempting to load YOLO model '{filename}' from repo '{repo_id}' using ultralytics.YOLO...")
print(f"Hugging Face Hub cache directory for YOLO: {yolo_cache_dir}")
try:
yolo_model_filepath = hf_hub_download(
repo_id=repo_id,
filename=filename,
cache_dir=yolo_cache_dir,
local_files_only=False,
force_download=False, # Set to True if you suspect a corrupted download
)
print(f"YOLO model file path: {yolo_model_filepath}")
from ultralytics import YOLO as UltralyticsYOLO # Import here
self._yolo = UltralyticsYOLO(yolo_model_filepath) # This is the line that fails with SCDown
print("MDR YOLOv10b-DocLayNet model loaded successfully using ultralytics.YOLO.")
except ImportError:
print("ERROR: ultralytics library not found. Cannot load YOLOv10b-DocLayNet.")
print("Please ensure it's installed: pip install ultralytics (matching version if possible)")
self._yolo = None
except HfHubHTTPError as e:
print(f"ERROR: Failed to download YOLO model '{filename}' via Hugging Face Hub: {e}")
self._yolo = None
except Exception as e: # Catch other model loading errors (like the SCDown error)
print(f"ERROR: Failed to load YOLO model '{yolo_model_filepath}' with ultralytics.YOLO: {e}")
import traceback
traceback.print_exc()
self._yolo = None # Ensure self._yolo is None on failure
return self._yolo
def analyze_image(self, image: Image, adjust_points: bool = False) -> MDRExtractionResult:
"""Analyzes a single page image to extract layout and content."""
print(" Engine: Analyzing image...")
# --- START: ADDED CLAHE PREPROCESSING ---
# Convert PIL Image to OpenCV BGR format
#ori_im_cv = cv2.cvtColor(np.array(image.convert("RGB")), cv2.COLOR_RGB2BGR)
#gray_cv = cv2.cvtColor(ori_im_cv, cv2.COLOR_BGR2GRAY)
#clahe_obj = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
#enhanced_gray_cv = clahe_obj.apply(gray_cv)
# Convert back to BGR for downstream components that might expect 3 channels
# (even if they only use one, like the detector)
# And then back to PIL Image for the optimizer
#processed_cv_bgr = cv2.cvtColor(enhanced_gray_cv, cv2.COLOR_GRAY2BGR)
# Convert the processed OpenCV image back to PIL Image for the optimizer
# The optimizer expects a PIL Image.
# The image passed to optimizer will now be the CLAHE'd version.
processed_pil_image = image #pil_fromarray(cv2.cvtColor(processed_cv_bgr, cv2.COLOR_BGR2RGB))
print(" Engine: CLAHE preprocessing applied to input image.")
optimizer = MDRImageOptimizer(processed_pil_image, adjust_points)
print(" Engine: Initial OCR...")
frags = list(self._ocr_engine.find_text_fragments(optimizer.image_np))
print(f" Engine: {len(frags)} fragments found.")
optimizer.receive_fragments(frags)
frags = optimizer._fragments # Use adjusted fragments
print(" Engine: Layout detection...")
yolo = self._get_yolo_model()
raw_layouts = []
if yolo:
try:
raw_layouts = list(self._run_yolo_detection(optimizer.image, yolo))
print(f" Engine: {len(raw_layouts)} raw layouts found.")
except Exception:
import traceback, sys
traceback.print_exc(file=sys.stderr)
print(" Engine: Matching fragments...")
layouts = self._match_fragments_to_layouts(frags, raw_layouts)
if not layouts and frags:
# treat the whole page as one plain-text layout
page_rect = MDRRectangle(
lt=(0, 0), rt=(optimizer.image.width, 0),
lb=(0, optimizer.image.height), rb=(optimizer.image.width, optimizer.image.height)
)
dummy = MDRPlainLayoutElement(
cls=MDRLayoutClass.PLAIN_TEXT, rect=page_rect, fragments=frags.copy()
)
layouts.append(dummy)
print(" Engine: Removing overlaps...")
layouts = mdr_remove_overlap_layouts(layouts)
print(f" Engine: {len(layouts)} layouts after overlap removal.")
if self._ocr_each and layouts:
print(" Engine: OCR correction...")
self._run_ocr_correction(optimizer.image, layouts)
print(" Engine: Determining reading order...")
layouts = self._layout_reader.determine_reading_order(layouts, optimizer.image.size)
layouts = [l for l in layouts if self._should_keep_layout(l)]
print(f" Engine: {len(layouts)} layouts after filtering.")
if self._ext_table or self._ext_formula:
print(" Engine: Parsing tables/formulas...")
self._parse_special_layouts(layouts, optimizer)
print(" Engine: Merging fragments...")
[setattr(l, 'fragments', mdr_merge_fragments_into_lines(l.fragments)) for l in layouts]
print(" Engine: Finalizing coords...")
optimizer.finalize_layout_coords(layouts)
print(" Engine: Analysis complete.")
return MDRExtractionResult(rotation=optimizer.rotation, layouts=layouts, extracted_image=image,
adjusted_image=optimizer.adjusted_image)
# In class MDRExtractionEngine
def _run_yolo_detection(self, img: Image, yolo: Any): # yolo is an ultralytics.YOLO instance
img_rgb = img.convert("RGB")
res_list = yolo.predict(source=img_rgb, imgsz=1024, conf=0.25, # Adjust conf as needed
device=self._device, verbose=False)
if not res_list or not hasattr(res_list[0], 'boxes') or res_list[0].boxes is None:
print(" Engine: YOLO detection (ultralytics) returned no results or no boxes.")
return
results = res_list[0]
model_class_names = {}
if hasattr(results, 'names') and isinstance(results.names, dict):
model_class_names = results.names
print(f" Engine: YOLO model class names from ultralytics: {model_class_names}")
else:
# This fallback is a major source of potential error if results.names isn't populated.
# It's better to fail or have a very explicit warning if names aren't found.
print(
" Engine: CRITICAL WARNING - Could not get class names from YOLO model. Layout mapping will likely be incorrect.")
# Forcing a known DocLayNet order as a last resort (HIGHLY UNRELIABLE without verification)
_doclaynet_names_fallback = ['Caption', 'Footnote', 'Formula', 'List-item', 'Page-footer',
'Page-header', 'Picture', 'Section-header', 'Table', 'Text', 'Title']
model_class_names = {i: name for i, name in enumerate(_doclaynet_names_fallback)}
print(f" Engine: Using FALLBACK class names (VERIFY!): {model_class_names}")
plain_mdr_classes: set[MDRLayoutClass] = {
MDRLayoutClass.TITLE, MDRLayoutClass.PLAIN_TEXT,
MDRLayoutClass.FIGURE_CAPTION, MDRLayoutClass.TABLE_CAPTION,
MDRLayoutClass.TABLE_FOOTNOTE, MDRLayoutClass.FORMULA_CAPTION,
}
if results.boxes.cls is None or results.boxes.xyxy is None:
print(" Engine: YOLO results.boxes.cls or .xyxy is None.")
return
print(f" Engine: Processing {len(results.boxes.cls)} detected YOLO boxes...")
for i in range(len(results.boxes.cls)):
yolo_cls_id = int(results.boxes.cls[i].item())
xyxy_tensor = results.boxes.xyxy[i]
yolo_cls_name = model_class_names.get(yolo_cls_id, f"UnknownID-{yolo_cls_id}")
mdr_cls = None
# --- THIS MAPPING IS BASED ON STANDARD DOCLAYNET ---
# --- VERIFY IT AGAINST `model_class_names` PRINTED ABOVE ---
if yolo_cls_name == 'Text':
mdr_cls = MDRLayoutClass.PLAIN_TEXT
elif yolo_cls_name == 'Title':
mdr_cls = MDRLayoutClass.TITLE
elif yolo_cls_name == 'Section-header':
mdr_cls = MDRLayoutClass.TITLE
elif yolo_cls_name == 'List-item':
mdr_cls = MDRLayoutClass.PLAIN_TEXT
elif yolo_cls_name == 'Table':
mdr_cls = MDRLayoutClass.TABLE
elif yolo_cls_name == 'Picture':
mdr_cls = MDRLayoutClass.FIGURE
elif yolo_cls_name == 'Formula':
mdr_cls = MDRLayoutClass.ISOLATE_FORMULA
elif yolo_cls_name == 'Caption':
mdr_cls = MDRLayoutClass.FIGURE_CAPTION # Needs context to be TABLE_CAPTION
elif yolo_cls_name == 'Footnote':
mdr_cls = MDRLayoutClass.TABLE_FOOTNOTE # Needs context
elif yolo_cls_name in ['Page-header', 'Page-footer']:
mdr_cls = MDRLayoutClass.ABANDON
if mdr_cls is None:
# print(f" Skipping YOLO box: class '{yolo_cls_name}' (ID {yolo_cls_id}) - not mapped.")
continue
# print(f" Detected: {yolo_cls_name} (ID {yolo_cls_id}) -> {mdr_cls.name}")
x1, y1, x2, y2 = map(float, xyxy_tensor)
rect = MDRRectangle(lt=(x1, y1), rt=(x2, y1), lb=(x1, y2), rb=(x2, y2))
if rect.area < 10: continue
if mdr_cls == MDRLayoutClass.TABLE:
yield MDRTableLayoutElement(rect=rect, fragments=[], parsed=None, cls=mdr_cls)
elif mdr_cls == MDRLayoutClass.ISOLATE_FORMULA:
yield MDRFormulaLayoutElement(rect=rect, fragments=[], latex=None, cls=mdr_cls)
elif mdr_cls == MDRLayoutClass.FIGURE:
yield MDRPlainLayoutElement(cls=mdr_cls, rect=rect, fragments=[])
elif mdr_cls in plain_mdr_classes:
yield MDRPlainLayoutElement(cls=mdr_cls, rect=rect, fragments=[])
elif mdr_cls == MDRLayoutClass.ABANDON:
yield MDRPlainLayoutElement(cls=mdr_cls, rect=rect, fragments=[])
def _match_fragments_to_layouts(self, frags: list[MDROcrFragment], layouts: list[MDRLayoutElement]) -> list[
MDRLayoutElement]:
if not frags or not layouts:
return layouts
layout_polys = [(Polygon(l.rect) if l.rect.is_valid else None) for l in layouts]
for frag in frags:
try:
frag_poly = Polygon(frag.rect)
frag_area = frag_poly.area
except:
continue
if not frag_poly.is_valid or frag_area < 1e-6:
continue
candidates = [] # (layout_idx, layout_area, overlap_ratio)
for idx, l_poly in enumerate(layout_polys):
if l_poly is None:
continue
try:
inter_area = frag_poly.intersection(l_poly).area
except:
continue
overlap = inter_area / frag_area if frag_area > 0 else 0
if overlap > 0.85:
candidates.append((idx, l_poly.area, overlap))
if candidates:
candidates.sort(key=lambda x: (x[1], -x[2]))
best_idx = candidates[0][0]
layouts[best_idx].fragments.append(frag)
for l in layouts:
l.fragments.sort(key=lambda f: (f.rect.lt[1], f.rect.lt[0]))
return layouts
def _run_ocr_correction(self, img: Image, layouts: list[MDRLayoutElement]):
for i, l in enumerate(layouts):
if l.cls == MDRLayoutClass.FIGURE: continue
try:
mdr_correct_layout_fragments(self._ocr_engine, img, l)
except Exception as e:
print(f" Engine: OCR correction error layout {i}: {e}")
def _parse_special_layouts(self, layouts: list[MDRLayoutElement], optimizer: MDRImageOptimizer):
img_to_clip = optimizer.image
for l in layouts:
if isinstance(l, MDRFormulaLayoutElement) and self._ext_formula:
try:
f_img = mdr_clip_from_image(img_to_clip, l.rect)
l.latex = self._latex_extractor.extract(f_img) if f_img.width > 1 and f_img.height > 1 else None
except Exception as e:
print(f" Engine: LaTeX extract error: {e}")
elif isinstance(l, MDRTableLayoutElement) and self._ext_table is not None:
try:
t_img = mdr_clip_from_image(img_to_clip, l.rect)
parsed = self._table_parser.parse_table_image(t_img,
self._ext_table) if t_img.width > 1 and t_img.height > 1 else None
except Exception as e:
print(f" Engine: Table parse error: {e}")
parsed = None
if parsed:
l.parsed = (parsed, self._ext_table)
def _should_keep_layout(self, l: MDRLayoutElement) -> bool:
if l.fragments and not all(mdr_is_whitespace(f.text) for f in l.fragments): return True
return l.cls in [MDRLayoutClass.FIGURE, MDRLayoutClass.TABLE, MDRLayoutClass.ISOLATE_FORMULA]
# --- MDR Page Section Linking ---
class _MDR_LinkedShape:
"""Internal helper for managing layout linking across pages."""
def __init__(self, layout: MDRLayoutElement): self.layout = layout; self.pre: list[MDRLayoutElement | None] = [None,
None]; self.nex: \
list[MDRLayoutElement | None] = [None, None]
@property
def distance2(self) -> float: x, y = self.layout.rect.lt; return x * x + y * y
class MDRPageSection:
"""Represents a page's layouts for framework detection via linking."""
def __init__(self, page_index: int, layouts: Iterable[MDRLayoutElement]):
self._page_index = page_index;
self._shapes = [_MDR_LinkedShape(l) for l in layouts];
self._shapes.sort(key=lambda s: (s.layout.rect.lt[1], s.layout.rect.lt[0]))
@property
def page_index(self) -> int:
return self._page_index
def find_framework_elements(self) -> list[MDRLayoutElement]:
"""Identifies framework layouts based on links to other pages."""
return [s.layout for s in self._shapes if any(s.pre) or any(s.nex)]
def link_to_next(self, next_section: 'MDRPageSection', offset: int) -> None:
"""Links matching shapes between this section and the next."""
if offset not in (1, 2):
return
matches_matrix = [[sn for sn in next_section._shapes if self._shapes_match(ss, sn)] for ss in self._shapes]
origin_pair = self._find_origin_pair(matches_matrix, next_section._shapes)
if origin_pair is None:
return
orig_s, orig_n = origin_pair
orig_s_pt = orig_s.layout.rect.lt
orig_n_pt = orig_n.layout.rect.lt
for i, s1 in enumerate(self._shapes):
potentials = matches_matrix[i]
if not potentials:
continue
r1_rel = self._relative_rect(orig_s_pt, s1.layout.rect)
best_s2 = None
max_ovr = -1.0
for s2 in potentials:
r2_rel = self._relative_rect(orig_n_pt, s2.layout.rect)
ovr = self._symmetric_iou(r1_rel, r2_rel)
if ovr > max_ovr:
max_ovr = ovr
best_s2 = s2
if max_ovr >= 0.80 and best_s2 is not None:
s1.nex[offset - 1] = best_s2.layout
best_s2.pre[offset - 1] = s1.layout # Link both ways
def _shapes_match(self, s1: _MDR_LinkedShape, s2: _MDR_LinkedShape) -> bool:
l1 = s1.layout
l2 = s2.layout
sz1 = l1.rect.size
sz2 = l2.rect.size
thresh = 0.90
if mdr_similarity_ratio(sz1[0], sz2[0]) < thresh or mdr_similarity_ratio(sz1[1], sz2[1]) < thresh:
return False
f1 = l1.fragments
f2 = l2.fragments
c1 = len(f1)
c2 = len(f2)
if c1 == 0 and c2 == 0:
return True
if c1 == 0 or c2 == 0:
return False
matches = 0
used_f2 = [False] * c2
for frag1 in f1:
best_j = -1
max_sim = -1.0
for j, frag2 in enumerate(f2):
if not used_f2[j]:
sim = self._fragment_sim(l1, l2, frag1, frag2)
if sim > max_sim:
max_sim = sim
best_j = j
if max_sim > 0.75:
matches += 1
if best_j != -1:
used_f2[best_j] = True
max_c = max(c1, c2)
rate_frags = matches / max_c
return self._check_match_threshold(rate_frags, max_c, (0.0, 0.45, 0.45, 0.6, 0.8, 0.95))
def _fragment_sim(self, l1: MDRLayoutElement, l2: MDRLayoutElement, f1: MDROcrFragment,
f2: MDROcrFragment) -> float:
r1_rel = self._relative_rect(l1.rect.lt, f1.rect)
r2_rel = self._relative_rect(l2.rect.lt, f2.rect)
geom_sim = self._symmetric_iou(r1_rel, r2_rel)
text_sim, _ = mdr_check_text_similarity(f1.text, f2.text)
return (geom_sim + text_sim) / 2.0
def _find_origin_pair(self, matches_matrix: list[list[_MDR_LinkedShape]], next_shapes: list[_MDR_LinkedShape]) -> \
tuple[_MDR_LinkedShape, _MDR_LinkedShape] | None:
best_pair = None
min_dist2 = float('inf')
for i, s1 in enumerate(self._shapes):
match_list = matches_matrix[i]
if not match_list:
continue
for s2 in match_list:
dist2 = s1.distance2 + s2.distance2
if dist2 < min_dist2:
min_dist2 = dist2
best_pair = (s1, s2)
return best_pair
def _check_match_threshold(self, rate: float, count: int, thresholds: Sequence[float]) -> bool:
if not thresholds: return False; idx = min(count, len(thresholds) - 1); return rate >= thresholds[idx]
def _relative_rect(self, origin: MDRPoint, rect: MDRRectangle) -> MDRRectangle:
ox, oy = origin
r = rect
return MDRRectangle(lt=(r.lt[0] - ox, r.lt[1] - oy), rt=(r.rt[0] - ox, r.rt[1] - oy),
lb=(r.lb[0] - ox, r.lb[1] - oy), rb=(r.rb[0] - ox, r.rb[1] - oy))
def _symmetric_iou(self, r1: MDRRectangle, r2: MDRRectangle) -> float:
try:
p1 = Polygon(r1)
p2 = Polygon(r2)
except:
return 0.0
if not p1.is_valid or not p2.is_valid:
return 0.0
try:
inter = p1.intersection(p2)
union = p1.union(p2)
except:
return 0.0
if inter.is_empty or inter.area < 1e-6:
return 0.0
union_area = union.area
return inter.area / union_area if union_area > 1e-6 else 1.0
# --- MDR Document Iterator ---
_MDR_CONTEXT_PAGES = 2 # Look behind/ahead pages for context
@dataclass
class MDRProcessingParams:
"""Parameters for processing a document."""
pdf: str | FitzDocument;
page_indexes: Iterable[int] | None;
report_progress: MDRProgressReportCallback | None
class MDRDocumentIterator:
"""Iterates through document pages, handles context, and calls the extraction engine."""
def __init__(self, device: Literal["cpu", "cuda"], model_dir_path: str, ocr_level: MDROcrLevel,
extract_formula: bool, extract_table_format: MDRTableLayoutParsedFormat | None,
debug_dir_path: str | None):
self._debug_dir = debug_dir_path
self._engine = MDRExtractionEngine(device=device, model_dir_path=model_dir_path,
ocr_for_each_layouts=(ocr_level == MDROcrLevel.OncePerLayout),
extract_formula=extract_formula, extract_table_format=extract_table_format)
def iterate_sections(self, params: MDRProcessingParams) -> Generator[
tuple[int, MDRExtractionResult, list[MDRLayoutElement]], None, None]:
"""Yields page index, extraction result, and content layouts for each requested page."""
# In MDRDocumentIterator.iterate_sections
for res, sec in self._process_and_link_sections(params):
# Get the IDs of the framework elements
framework_element_ids = {id(fw_el) for fw_el in sec.find_framework_elements()}
# Filter content layouts by checking if their ID is not in the set of framework element IDs
content = [l for l in res.layouts if id(l) not in framework_element_ids]
yield sec.page_index, res, content
def _process_and_link_sections(self, params: MDRProcessingParams) -> Generator[
tuple[MDRExtractionResult, MDRPageSection], None, None]:
queue: list[tuple[MDRExtractionResult, MDRPageSection]] = []
for page_idx, res in self._run_extraction_on_pages(params):
cur_sec = MDRPageSection(page_idx, res.layouts)
for i, (_, prev_sec) in enumerate(queue):
offset = len(queue) - i
if offset <= _MDR_CONTEXT_PAGES:
prev_sec.link_to_next(cur_sec, offset)
queue.append((res, cur_sec))
if len(queue) > _MDR_CONTEXT_PAGES:
yield queue.pop(0)
for res, sec in queue:
yield res, sec
def _run_extraction_on_pages(self, params: MDRProcessingParams) -> Generator[
tuple[int, MDRExtractionResult], None, None]:
if self._debug_dir: mdr_ensure_directory(self._debug_dir)
doc, should_close = None, False
if isinstance(params.pdf, str):
try:
doc = fitz.open(params.pdf); should_close = True
except Exception as e:
print(f"ERROR: PDF open failed: {e}"); return
elif isinstance(params.pdf, FitzDocument):
doc = params.pdf
else:
print(f"ERROR: Invalid PDF type: {type(params.pdf)}"); return
scan_idxs, enable_idxs = self._get_page_ranges(doc, params.page_indexes)
enable_set = set(enable_idxs);
total_scan = len(scan_idxs)
try:
for i, page_idx in enumerate(scan_idxs):
print(f" Iterator: Processing page {page_idx + 1}/{doc.page_count} (Scan {i + 1}/{total_scan})...")
try:
page = doc.load_page(page_idx)
img = self._render_page_image(page, 300)
res = self._engine.analyze_image(image=img, adjust_points=False) # Engine analyzes image
if self._debug_dir:
self._save_debug_plot(img, page_idx, res, self._debug_dir)
if page_idx in enable_set:
yield page_idx, res # Yield result for requested pages
if params.report_progress:
params.report_progress(i + 1, total_scan)
except Exception as e:
print(f" Iterator: Page {page_idx + 1} processing error: {e}")
finally:
if should_close and doc: doc.close()
def _get_page_ranges(self, doc: FitzDocument, idxs: Iterable[int] | None) -> tuple[Sequence[int], Sequence[int]]:
count = doc.page_count
if idxs is None:
all_p = list(range(count))
return all_p, all_p
enable = set()
scan = set()
for i in idxs:
if 0 <= i < count:
enable.add(i)
[scan.add(j) for j in range(max(0, i - _MDR_CONTEXT_PAGES), min(count, i + _MDR_CONTEXT_PAGES + 1))]
return sorted(list(scan)), sorted(list(enable))
def _render_page_image(self, page: FitzPage, dpi: int) -> Image:
mat = FitzMatrix(dpi / 72.0, dpi / 72.0)
pix = page.get_pixmap(matrix=mat, alpha=False)
return frombytes("RGB", (pix.width, pix.height), pix.samples)
def _save_debug_plot(self, img: Image, idx: int, res: MDRExtractionResult, path: str):
try:
plot_img = res.adjusted_image.copy() if res.adjusted_image else img.copy()
mdr_plot_layout(plot_img, res.layouts)
plot_img.save(os.path.join(path, f"mdr_plot_page_{idx + 1}.png"))
except Exception as e:
print(f" Iterator: Plot generation error page {idx + 1}: {e}")
# --- MagicDataReadiness Main Processor ---
class MagicPDFProcessor:
"""
Main class for processing PDF documents to extract structured data blocks
using the MagicDataReadiness pipeline.
"""
def __init__(self, device: Literal["cpu", "cuda"] = "cuda", model_dir_path: str = "./mdr_models",
ocr_level: MDROcrLevel = MDROcrLevel.Once, extract_formula: bool = True,
extract_table_format: MDRExtractedTableFormat | None = None, debug_dir_path: str | None = None):
"""
Initializes the MagicPDFProcessor.
Args:
device: Computation device ('cpu' or 'cuda'). Defaults to 'cuda'. Fallbacks to 'cpu' if CUDA not available.
model_dir_path: Path to directory for storing/caching downloaded models. Defaults to './mdr_models'.
ocr_level: Level of OCR application (Once per page or Once per layout). Defaults to Once per page.
extract_formula: Whether to attempt LaTeX extraction from formula images. Defaults to True.
extract_table_format: Desired format for extracted table content (LATEX, MARKDOWN, HTML, DISABLE, or None).
Defaults to LATEX if CUDA is available, otherwise DISABLE.
debug_dir_path: Optional path to save debug plots and intermediate files. Defaults to None (disabled).
"""
actual_dev = device if torch.cuda.is_available() else "cpu";
print(f"MagicPDFProcessor using device: {actual_dev}.")
if extract_table_format is None: extract_table_format = MDRExtractedTableFormat.LATEX if actual_dev == "cuda" else MDRExtractedTableFormat.DISABLE
table_fmt_internal: MDRTableLayoutParsedFormat | None = None
if extract_table_format == MDRExtractedTableFormat.LATEX:
table_fmt_internal = MDRTableLayoutParsedFormat.LATEX
elif extract_table_format == MDRExtractedTableFormat.MARKDOWN:
table_fmt_internal = MDRTableLayoutParsedFormat.MARKDOWN
elif extract_table_format == MDRExtractedTableFormat.HTML:
table_fmt_internal = MDRTableLayoutParsedFormat.HTML
self._iterator = MDRDocumentIterator(device=actual_dev, model_dir_path=model_dir_path, ocr_level=ocr_level,
extract_formula=extract_formula, extract_table_format=table_fmt_internal,
debug_dir_path=debug_dir_path)
print("MagicPDFProcessor initialized.")
def process_document(self, pdf_input: str | FitzDocument,
report_progress: MDRProgressReportCallback | None = None) -> Generator[
MDRStructuredBlock, None, None]:
"""
Processes the entire PDF document and yields all extracted structured blocks.
Args:
pdf_input: Path to the PDF file or a loaded fitz.Document object.
report_progress: Optional callback function for progress updates (receives completed_scan_pages, total_scan_pages).
Yields:
MDRStructuredBlock: An extracted block (MDRTextBlock, MDRTableBlock, etc.).
"""
print(f"Processing document: {pdf_input if isinstance(pdf_input, str) else 'FitzDocument object'}")
for _, blocks, _ in self.process_document_pages(pdf_input=pdf_input, report_progress=report_progress,
page_indexes=None):
yield from blocks
print("Document processing complete.")
def process_document_pages(self, pdf_input: str | FitzDocument, page_indexes: Iterable[int] | None = None,
report_progress: MDRProgressReportCallback | None = None) -> Generator[
tuple[int, list[MDRStructuredBlock], Image], None, None]:
"""
Processes specific pages (or all if page_indexes is None) of the PDF document.
Yields results page by page, including the page index, extracted blocks, and the original page image.
Args:
pdf_input: Path to the PDF file or a loaded fitz.Document object.
page_indexes: An iterable of 0-based page indices to process. If None, processes all pages.
report_progress: Optional callback function for progress updates.
Yields:
tuple[int, list[MDRStructuredBlock], Image]:
- page_index (0-based)
- list of extracted MDRStructuredBlock objects for that page
- PIL Image object of the original rendered page
"""
params = MDRProcessingParams(pdf=pdf_input, page_indexes=page_indexes, report_progress=report_progress)
page_count = 0
for page_idx, extraction_result, content_layouts in self._iterator.iterate_sections(params):
page_count += 1
print(f"Processor: Converting layouts to blocks for page {page_idx + 1}...")
blocks = self._create_structured_blocks(extraction_result, content_layouts)
print(f"Processor: Analyzing paragraph structure for page {page_idx + 1}...")
self._analyze_paragraph_structure(blocks)
print(f"Processor: Yielding results for page {page_idx + 1}.")
yield page_idx, blocks, extraction_result.extracted_image # Yield original image
print(f"Processor: Finished processing {page_count} pages.")
def _create_structured_blocks(self, result: MDRExtractionResult, layouts: list[MDRLayoutElement]) -> list[
MDRStructuredBlock]:
"""Converts MDRLayoutElement objects into MDRStructuredBlock objects."""
temp_store: list[tuple[MDRLayoutElement, MDRStructuredBlock]] = []
for layout in layouts:
if isinstance(layout, MDRPlainLayoutElement):
self._add_plain_block(temp_store, layout, result)
elif isinstance(layout, MDRTableLayoutElement):
temp_store.append((layout, self._create_table_block(layout, result)))
elif isinstance(layout, MDRFormulaLayoutElement):
temp_store.append((layout, self._create_formula_block(layout, result)))
self._assign_relative_font_sizes(temp_store)
return [block for _, block in temp_store]
def _analyze_paragraph_structure(self, blocks: list[MDRStructuredBlock]):
"""
Calculates indentation and line-end heuristics for MDRTextBlocks
based on page-level text boundaries and average line height.
"""
# Define constants for clarity and maintainability
MIN_VALID_HEIGHT = 1e-6
# Heuristic: Indent if first line starts more than 1.0 * avg line height from page text start
INDENTATION_THRESHOLD_FACTOR = 1.0
# Heuristic: Last line touches end if it ends less than 1.0 * avg line height from page text end
LINE_END_THRESHOLD_FACTOR = 1.0
# Calculate average line height and text boundaries for the relevant text blocks on the page
page_avg_line_height, page_min_x, page_max_x = self._calculate_text_range(
(b for b in blocks if isinstance(b, MDRTextBlock) and b.kind != MDRTextKind.ABANDON)
)
# Avoid calculations if page metrics are invalid (e.g., no text, zero height)
if page_avg_line_height <= MIN_VALID_HEIGHT:
return
# Iterate through each block to determine its paragraph properties
for block in blocks:
# Process only valid text blocks with actual text content
if not isinstance(block, MDRTextBlock) or block.kind == MDRTextKind.ABANDON or not block.texts:
continue
# Use calculated page-level metrics for consistency in thresholds
avg_line_height = page_avg_line_height
page_text_start_x = page_min_x
page_text_end_x = page_max_x
# Get the first and last text spans (assumed to be lines after merging) within the block
first_text_span = block.texts[0]
last_text_span = block.texts[-1]
try:
# --- Calculate Indentation ---
# Estimate the starting x-coordinate of the first line (average of left top/bottom)
first_line_start_x = (first_text_span.rect.lt[0] + first_text_span.rect.lb[0]) / 2.0
# Calculate the difference between the first line's start and the page's text start boundary
indentation_delta = first_line_start_x - page_text_start_x
# Determine indentation based on the heuristic threshold relative to average line height
block.has_paragraph_indentation = indentation_delta > (avg_line_height * INDENTATION_THRESHOLD_FACTOR)
# --- Calculate Last Line End ---
# Estimate the ending x-coordinate of the last line (average of right top/bottom)
last_line_end_x = (last_text_span.rect.rt[0] + last_text_span.rect.rb[0]) / 2.0
# Calculate the difference between the page's text end boundary and the last line's end
line_end_delta = page_text_end_x - last_line_end_x
# Determine if the last line reaches near the end based on the heuristic threshold
block.last_line_touch_end = line_end_delta < (avg_line_height * LINE_END_THRESHOLD_FACTOR)
except Exception as e:
# Handle potential errors during calculation (e.g., invalid rect data)
print(f"Warn: Error calculating paragraph structure for block: {e}")
# Default to False if calculation fails to ensure attributes are set
block.has_paragraph_indentation = False
block.last_line_touch_end = False
def _calculate_text_range(self, blocks_iter: Iterable[MDRStructuredBlock]) -> tuple[float, float, float]:
"""Calculates average line height and min/max x-coordinates for text."""
h_sum = 0.0
count = 0
x1 = float('inf')
x2 = float('-inf')
for b in blocks_iter:
if not isinstance(b, MDRTextBlock) or b.kind == MDRTextKind.ABANDON:
continue
for t in b.texts:
_, h = t.rect.size
if h > 1e-6: # Use small threshold for valid height
h_sum += h
count += 1
tx1, _, tx2, _ = t.rect.wrapper
x1 = min(x1, tx1)
x2 = max(x2, tx2)
if count == 0:
return 0.0, 0.0, 0.0
mean_h = h_sum / count
x1 = 0.0 if x1 == float('inf') else x1
x2 = 0.0 if x2 == float('-inf') else x2
return mean_h, x1, x2
def _add_plain_block(self, store: list[tuple[MDRLayoutElement, MDRStructuredBlock]], layout: MDRPlainLayoutElement,
result: MDRExtractionResult):
"""Creates MDRStructuredBlocks for plain layout types."""
cls = layout.cls
texts = self._convert_fragments_to_spans(layout.fragments)
if cls == MDRLayoutClass.TITLE:
store.append((layout, MDRTextBlock(layout.rect, texts, 0.0, MDRTextKind.TITLE)))
elif cls == MDRLayoutClass.PLAIN_TEXT:
store.append((layout, MDRTextBlock(layout.rect, texts, 0.0, MDRTextKind.PLAIN_TEXT)))
elif cls == MDRLayoutClass.ABANDON:
store.append((layout, MDRTextBlock(layout.rect, texts, 0.0, MDRTextKind.ABANDON)))
elif cls == MDRLayoutClass.FIGURE:
store.append((layout, MDRFigureBlock(layout.rect, [], 0.0, mdr_clip_layout(result, layout))))
elif cls == MDRLayoutClass.FIGURE_CAPTION:
block = self._find_previous_block(store, MDRFigureBlock)
if block: block.texts.extend(texts)
elif cls == MDRLayoutClass.TABLE_CAPTION or cls == MDRLayoutClass.TABLE_FOOTNOTE:
block = self._find_previous_block(store, MDRTableBlock)
if block: block.texts.extend(texts)
elif cls == MDRLayoutClass.FORMULA_CAPTION:
block = self._find_previous_block(store, MDRFormulaBlock)
if block: block.texts.extend(texts)
def _find_previous_block(self, store: list[tuple[MDRLayoutElement, MDRStructuredBlock]],
block_type: type) -> MDRStructuredBlock | None:
"""Finds the most recent block of a specific type."""
for i in range(len(store) - 1, -1, -1):
_, block = store[i]
if isinstance(block, block_type):
return block
return None
def _create_table_block(self, layout: MDRTableLayoutElement, result: MDRExtractionResult) -> MDRTableBlock:
"""Converts MDRTableLayoutElement to MDRTableBlock."""
fmt = MDRTableFormat.UNRECOGNIZABLE
content = ""
if layout.parsed:
p_content, p_fmt = layout.parsed
can_use = not (p_fmt == MDRTableLayoutParsedFormat.LATEX and mdr_contains_cjka(
"".join(f.text for f in layout.fragments)))
if can_use:
content = p_content
if p_fmt == MDRTableLayoutParsedFormat.LATEX:
fmt = MDRTableFormat.LATEX
elif p_fmt == MDRTableLayoutParsedFormat.MARKDOWN:
fmt = MDRTableFormat.MARKDOWN
elif p_fmt == MDRTableLayoutParsedFormat.HTML:
fmt = MDRTableFormat.HTML
return MDRTableBlock(layout.rect, [], 0.0, fmt, content, mdr_clip_layout(result, layout))
def _create_formula_block(self, layout: MDRFormulaLayoutElement, result: MDRExtractionResult) -> MDRFormulaBlock:
"""Converts MDRFormulaLayoutElement to MDRFormulaBlock."""
content = layout.latex if layout.latex and not mdr_contains_cjka(
"".join(f.text for f in layout.fragments)) else None
return MDRFormulaBlock(layout.rect, [], 0.0, content, mdr_clip_layout(result, layout))
def _assign_relative_font_sizes(self, store: list[tuple[MDRLayoutElement, MDRStructuredBlock]]):
"""Calculates and assigns relative font size (0-1) to blocks."""
sizes = []
for l, _ in store:
heights = [f.rect.size[1] for f in l.fragments if f.rect.size[1] > 1e-6] # Use small threshold
avg_h = sum(heights) / len(heights) if heights else 0.0
sizes.append(avg_h)
valid = [s for s in sizes if s > 1e-6]
min_s, max_s = (min(valid), max(valid)) if valid else (0.0, 0.0)
rng = max_s - min_s
if rng < 1e-6:
[setattr(b, 'font_size', 0.0) for _, b in store]
else:
[setattr(b, 'font_size', (s - min_s) / rng if s > 1e-6 else 0.0) for s, (_, b) in zip(sizes, store)]
def _convert_fragments_to_spans(self, frags: list[MDROcrFragment]) -> list[MDRTextSpan]:
"""Converts MDROcrFragment list to MDRTextSpan list."""
return [MDRTextSpan(f.text, f.rank, f.rect) for f in frags]
# --- MagicDataReadiness Example Usage ---
if __name__ == '__main__':
print("=" * 60)
print(" MagicDataReadiness PDF Processor - Example Usage")
print("=" * 60)
# --- 1. Configuration (!!! MODIFY THESE PATHS WHEN OUTSIDE HF !!!) ---
# Directory where models are stored or will be downloaded
# IMPORTANT: Create this directory or ensure it's writable!
MDR_MODEL_DIRECTORY = "./mdr_pipeline_models"
# Path to the PDF file you want to process
# IMPORTANT: Place a PDF file here for testing!
# Create a dummy PDF if it doesn't exist for the example to run
MDR_INPUT_PDF = "example_input.pdf" # <--- CHANGE THIS
if not Path(MDR_INPUT_PDF).exists():
try:
print(f"Creating dummy PDF: {MDR_INPUT_PDF}")
doc = fitz.new_document()
page = doc.new_page()
page.insert_text((72, 72), "This is a dummy PDF for testing.")
doc.save(MDR_INPUT_PDF)
doc.close()
except Exception as e:
print(f"Warning: Could not create dummy PDF: {e}")
# Optional: Directory to save debug plots (set to None to disable)
MDR_DEBUG_DIRECTORY = "./mdr_debug_output"
# Specify device ('cuda' or 'cpu').
MDR_DEVICE = "cuda"
# Specify desired table format
MDR_TABLE_FORMAT = MDRExtractedTableFormat.MARKDOWN
# Specify pages (list of 0-based indices, or None for all)
MDR_PAGES = None # Example: [0, 1] for first two pages
# --- 2. Setup & Pre-checks ---
print(f"Model Directory: {os.path.abspath(MDR_MODEL_DIRECTORY)}")
print(f"Input PDF: {os.path.abspath(MDR_INPUT_PDF)}")
print(f"Debug Output: {os.path.abspath(MDR_DEBUG_DIRECTORY) if MDR_DEBUG_DIRECTORY else 'Disabled'}")
print(f"Target Device: {MDR_DEVICE}")
print(f"Table Format: {MDR_TABLE_FORMAT.name}")
print(f"Pages: {'All' if MDR_PAGES is None else MDR_PAGES}")
print("-" * 60)
mdr_ensure_directory(MDR_MODEL_DIRECTORY)
if MDR_DEBUG_DIRECTORY:
mdr_ensure_directory(MDR_DEBUG_DIRECTORY)
if not Path(MDR_INPUT_PDF).is_file():
print(f"ERROR: Input PDF not found at '{MDR_INPUT_PDF}'. Please place a PDF file there or update the path.")
exit(1)
# --- 3. Progress Callback ---
def mdr_progress_update(completed, total):
perc = (completed / total) * 100 if total > 0 else 0
print(f" [Progress] Scanned {completed}/{total} pages ({perc:.1f}%)")
# --- 4. Initialize Processor ---
print("Initializing MagicPDFProcessor...")
init_start = time.time()
try:
mdr_processor = MagicPDFProcessor(
device=MDR_DEVICE,
model_dir_path=MDR_MODEL_DIRECTORY,
debug_dir_path=MDR_DEBUG_DIRECTORY,
extract_table_format=MDR_TABLE_FORMAT
)
print(f"Initialization took {time.time() - init_start:.2f}s")
except Exception as e:
print(f"FATAL ERROR during initialization: {e}")
import traceback
traceback.print_exc()
exit(1)
# --- 5. Process Document ---
print("\nStarting document processing...")
proc_start = time.time()
all_blocks_count = 0
processed_pages_count = 0
try:
# Use the main processing method
block_generator = mdr_processor.process_document_pages(
pdf_input=MDR_INPUT_PDF,
page_indexes=MDR_PAGES,
report_progress=mdr_progress_update
)
# Iterate through pages and blocks
for page_idx, page_blocks, page_img in block_generator:
processed_pages_count += 1
print(f"\n--- Page {page_idx + 1} Results ---")
if not page_blocks:
print(" No blocks extracted.")
continue
print(f" Extracted {len(page_blocks)} blocks:")
for block_idx, block in enumerate(page_blocks):
all_blocks_count += 1
info = f" - Block {block_idx + 1}: {type(block).__name__}"
if isinstance(block, MDRTextBlock):
preview = block.texts[0].content[:70].replace('\n', ' ') + "..." if block.texts else "[EMPTY]"
info += f" (Kind: {block.kind.name}, FontSz: {block.font_size:.2f}, Indent: {block.has_paragraph_indentation}, EndTouch: {block.last_line_touch_end}) | Text: '{preview}'" # Added indent/endtouch
elif isinstance(block, MDRTableBlock):
info += f" (Format: {block.format.name}, HasContent: {bool(block.content)}, FontSz: {block.font_size:.2f})"
# if block.content: print(f" Content:\n{block.content}") # Uncomment to see content
elif isinstance(block, MDRFormulaBlock):
info += f" (HasLatex: {bool(block.content)}, FontSz: {block.font_size:.2f})"
# if block.content: print(f" LaTeX: {block.content}") # Uncomment to see content
elif isinstance(block, MDRFigureBlock):
info += f" (FontSz: {block.font_size:.2f})"
print(info)
proc_time = time.time() - proc_start
print("\n" + "=" * 60)
print(" Processing Summary")
print(f" Total time: {proc_time:.2f} seconds")
print(f" Pages processed: {processed_pages_count}")
print(f" Total blocks extracted: {all_blocks_count}")
print("=" * 60)
except Exception as e:
print(f"\nFATAL ERROR during processing: {e}")
import traceback
traceback.print_exc()
exit(1)