Spaces:
Sleeping
Sleeping
File size: 44,481 Bytes
443e99e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 |
import os
import json
import signal
import sys
from pathlib import Path
from typing import List, Dict, Tuple, Optional, Sequence, Set, Any
from multiprocessing import Pool, cpu_count
from functools import partial
import fitz # PyMuPDF (Still needed for drawing output PDF)
import pypdfium2 as pdfium
import torch
from doclayout_yolo import YOLOv10
from huggingface_hub import hf_hub_download
from loguru import logger
from PIL import Image
import numpy as np
try:
import pymupdf4llm # type: ignore
except ImportError: # pragma: no cover - optional dependency
pymupdf4llm = None # type: ignore
# ----------------------------------------------------------------------
# CONFIGURATION
# ----------------------------------------------------------------------
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# Model options
MODEL_SIZE = 1024
REPO_ID = "juliozhao/DocLayout-YOLO-DocStructBench"
WEIGHTS_FILE = f"doclayout_yolo_docstructbench_imgsz{MODEL_SIZE}.pt"
# Detection settings
CONF_THRESHOLD = 0.25
# Multiprocessing settings
NUM_WORKERS = None # None = auto (cpu_count - 1), or set to specific number like 4
USE_MULTIPROCESSING = True # Set to False to disable parallel processing entirely
# ----------------------------------------------------------------------
# Color map for the layout classes
# ----------------------------------------------------------------------
CLASS_COLORS = {
"text": (0, 128, 0), # Dark Green
"title": (192, 0, 0), # Dark Red
"figure": (0, 0, 192), # Dark Blue
"table": (218, 165, 32), # Goldenrod (Dark Yellow)
"list": (128, 0, 128), # Purple
"header": (0, 128, 128), # Teal
"footer": (100, 100, 100), # Dark Gray
"figure_caption": (0, 0, 128), # Navy
"table_caption": (139, 69, 19), # Saddle Brown
"table_footnote": (128, 0, 128), # Purple
}
# Global model instance (will be None in worker processes until loaded)
_model = None
_shutdown_requested = False
# ----------------------------------------------------------------------
# Signal handler for graceful shutdown
# ----------------------------------------------------------------------
def signal_handler(signum, frame):
"""Handle interrupt signals gracefully."""
global _shutdown_requested
if not _shutdown_requested:
_shutdown_requested = True
logger.warning("\nβ οΈ Interrupt received! Finishing current page and shutting down gracefully...")
logger.warning("Press Ctrl+C again to force quit (may leave incomplete files)")
else:
logger.error("\nβ Force quit requested. Exiting immediately.")
sys.exit(1)
def setup_signal_handlers():
"""Setup signal handlers for graceful shutdown."""
signal.signal(signal.SIGINT, signal_handler)
signal.signal(signal.SIGTERM, signal_handler)
# ----------------------------------------------------------------------
# Model loader function
# ----------------------------------------------------------------------
def get_model():
"""Lazy load the model (only once per process)."""
global _model
if _model is None:
weights_path = hf_hub_download(repo_id=REPO_ID, filename=WEIGHTS_FILE)
_model = YOLOv10(weights_path)
logger.info(f"β Model loaded in worker process (PID: {os.getpid()})")
return _model
# ----------------------------------------------------------------------
# Worker initialization function
# ----------------------------------------------------------------------
def init_worker():
"""Initialize worker process - loads model once at startup."""
try:
get_model()
logger.success(f"Worker {os.getpid()} ready")
except Exception as e:
logger.error(f"Failed to initialize worker {os.getpid()}: {e}")
raise
# ----------------------------------------------------------------------
# Run layout detection on a single page image (YOLO)
# ----------------------------------------------------------------------
def detect_page(pil_img: Image.Image) -> List[dict]:
"""Detect layout elements using YOLO model."""
model = get_model() # Will return already-loaded model in worker
img_cv = np.array(pil_img)
results = model.predict(
img_cv,
imgsz=MODEL_SIZE,
conf=CONF_THRESHOLD,
device=DEVICE,
verbose=False
)
dets = []
for i, box in enumerate(results[0].boxes):
cls_id = int(box.cls.item())
name = results[0].names[cls_id]
conf = float(box.conf.item())
x0, y0, x1, y1 = box.xyxy[0].cpu().numpy().tolist()
dets.append({
"name": name,
"bbox": [x0, y0, x1, y1],
"conf": conf,
"source": "yolo",
"index": i
})
return dets
# ----------------------------------------------------------------------
# Crop & save figure/table regions (with captions)
# ----------------------------------------------------------------------
def get_union_box(box1: List[float], box2: List[float]) -> List[float]:
"""Get the bounding box enclosing two boxes."""
x0 = min(box1[0], box2[0])
y0 = min(box1[1], box2[1])
x1 = max(box1[2], box2[2])
y1 = max(box1[3], box2[3])
return [x0, y0, x1, y1]
def collect_caption_elements(
element: Dict,
all_dets: List[Dict],
target_name: str,
max_vertical_gap: float = 60.0,
min_overlap: float = 0.25,
) -> List[Dict]:
"""
Collect contiguous caption detections directly below a figure/table.
"""
base_box = element["bbox"]
base_bottom = base_box[3]
selected: List[Dict] = []
last_bottom = base_bottom
relevant = [
d for d in all_dets
if d["name"] == target_name and d["bbox"][1] >= base_bottom - 5
]
relevant.sort(key=lambda d: d["bbox"][1])
for cand in relevant:
cand_box = cand["bbox"]
top = cand_box[1]
if selected and top - last_bottom > max_vertical_gap:
break
if selected:
overlap = _horizontal_overlap_ratio(selected[-1]["bbox"], cand_box)
else:
overlap = _horizontal_overlap_ratio(base_box, cand_box)
if overlap < min_overlap:
continue
selected.append(cand)
last_bottom = cand_box[3]
return selected
def collect_title_and_text_segments(
element: Dict,
all_dets: List[Dict],
processed_indices: Set[int],
settings: Optional[Dict[str, float]] = None,
) -> Tuple[List[Dict], List[Dict]]:
"""
Locate a title below the element and any contiguous text blocks directly beneath it.
"""
if settings is None:
settings = TITLE_TEXT_ASSOCIATION
if not element.get("bbox"):
return [], []
figure_box = element["bbox"]
figure_bottom = figure_box[3]
candidates = [
d for d in all_dets
if d.get("bbox") and d["index"] not in processed_indices
]
candidates.sort(key=lambda d: d["bbox"][1])
titles: List[Dict] = []
texts: List[Dict] = []
for idx, det in enumerate(candidates):
if det["name"] != "title":
continue
title_box = det["bbox"]
if title_box[1] < figure_bottom - 5:
continue
vertical_gap = title_box[1] - figure_bottom
if vertical_gap > settings["max_title_gap"]:
break
overlap = _horizontal_overlap_ratio(figure_box, title_box)
if overlap < settings["min_overlap"]:
continue
titles.append(det)
last_bottom = title_box[3]
for follower in candidates[idx + 1 :]:
if follower["name"] == "title":
break
if follower["name"] != "text":
continue
text_box = follower["bbox"]
if text_box[1] < title_box[1]:
continue
gap = text_box[1] - last_bottom
if gap > settings["max_text_gap"]:
break
if _horizontal_overlap_ratio(title_box, text_box) < settings["min_overlap"]:
continue
texts.append(follower)
last_bottom = text_box[3]
break
return titles, texts
def save_layout_elements(pil_img: Image.Image, page_num: int,
dets: List[dict], out_dir: Path) -> List[dict]:
"""Save figure and table crops, merging captions."""
fig_dir = out_dir / "figures"
tab_dir = out_dir / "tables"
os.makedirs(fig_dir, exist_ok=True)
os.makedirs(tab_dir, exist_ok=True)
infos = []
fig_count = 0
tab_count = 0
processed_indices = set()
for i, d in enumerate(dets):
if d["index"] in processed_indices:
continue
name = d["name"].lower()
final_box = d["bbox"]
caption_segments: List[Dict] = []
title_segments: List[Dict] = []
text_segments: List[Dict] = []
if name == "figure":
elem_type = "figure"
path_template = fig_dir / f"page_{page_num + 1}_fig_{fig_count}.png"
fig_count += 1
caption_segments = collect_caption_elements(d, dets, "figure_caption")
for cap in caption_segments:
final_box = get_union_box(final_box, cap["bbox"])
processed_indices.add(cap["index"])
title_segments, text_segments = collect_title_and_text_segments(
d, dets, processed_indices
)
for seg in title_segments + text_segments:
final_box = get_union_box(final_box, seg["bbox"])
processed_indices.add(seg["index"])
elif name == "table":
elem_type = "table"
path_template = tab_dir / f"page_{page_num + 1}_tab_{tab_count}.png"
tab_count += 1
caption_segments = collect_caption_elements(d, dets, "table_caption")
for cap in caption_segments:
final_box = get_union_box(final_box, cap["bbox"])
processed_indices.add(cap["index"])
else:
continue
x0, y0, x1, y1 = map(int, final_box)
crop = pil_img.crop((x0, y0, x1, y1))
if crop.mode == "CMYK":
crop = crop.convert("RGB")
crop.save(path_template)
info_data = {
"type": elem_type,
"page": page_num + 1,
"bbox_pixels": final_box,
"conf": d["conf"],
"source": d.get("source", "yolo"),
"image_path": str(path_template.relative_to(out_dir)),
"width": int(x1 - x0),
"height": int(y1 - y0),
"page_width": pil_img.width,
"page_height": pil_img.height,
}
if caption_segments:
info_data["captions"] = [
{
"bbox": cap["bbox"],
"conf": cap.get("conf"),
"index": cap["index"],
"source": cap.get("source"),
"page": page_num + 1,
}
for cap in caption_segments
]
if title_segments:
info_data["titles"] = [
{
"bbox": seg["bbox"],
"conf": seg.get("conf"),
"index": seg["index"],
"source": seg.get("source"),
"page": page_num + 1,
}
for seg in title_segments
]
if text_segments:
info_data["texts"] = [
{
"bbox": seg["bbox"],
"conf": seg.get("conf"),
"index": seg["index"],
"source": seg.get("source"),
"page": page_num + 1,
}
for seg in text_segments
]
infos.append(info_data)
return infos
TABLE_STITCH_TOLERANCES = {
"x_tol": 60,
"y_tol": 60,
"width_tol": 120,
"height_tol": 120,
}
CROSS_PAGE_CAPTION_THRESHOLDS = {
"max_top_ratio": 0.35,
"max_top_pixels": 220,
"x_tol": 120,
"width_tol": 200,
"min_overlap": 0.05,
}
TITLE_TEXT_ASSOCIATION = {
"max_title_gap": 220,
"max_text_gap": 160,
"min_overlap": 0.2,
}
def _horizontal_overlap_ratio(box1: List[float], box2: List[float]) -> float:
"""Compute horizontal overlap ratio between two bounding boxes."""
x_left = max(box1[0], box2[0])
x_right = min(box1[2], box2[2])
overlap = max(0.0, x_right - x_left)
if overlap <= 0:
return 0.0
width_union = max(box1[2], box2[2]) - min(box1[0], box2[0])
if width_union <= 0:
return 0.0
return overlap / width_union
def _bbox_to_rect(bbox: List[float]) -> Tuple[int, int, int, int]:
"""Convert [x0, y0, x1, y1] into (x, y, w, h)."""
x0, y0, x1, y1 = bbox
return int(x0), int(y0), int(x1 - x0), int(y1 - y0)
def _open_table_image(elem: Dict, out_dir: Path) -> Optional[Image.Image]:
"""Open a table image relative to the output directory."""
image_path = out_dir / elem["image_path"]
if not image_path.exists():
logger.warning(f"Missing table crop for stitching: {image_path}")
return None
img = Image.open(image_path)
if img.mode != "RGB":
img = img.convert("RGB")
return img
def _pad_width(img: Image.Image, target_width: int) -> Image.Image:
if img.width >= target_width:
return img
canvas = Image.new("RGB", (target_width, img.height), color=(255, 255, 255))
canvas.paste(img, (0, 0))
return canvas
def _pad_height(img: Image.Image, target_height: int) -> Image.Image:
if img.height >= target_height:
return img
canvas = Image.new("RGB", (img.width, target_height), color=(255, 255, 255))
canvas.paste(img, (0, 0))
return canvas
def _append_segment_image(
base_img: Image.Image,
segment_img: Image.Image,
resize_to_base: bool = False,
) -> Image.Image:
"""Append segment image below base image with optional width alignment."""
if base_img.mode != "RGB":
base_img = base_img.convert("RGB")
if segment_img.mode != "RGB":
segment_img = segment_img.convert("RGB")
if resize_to_base and segment_img.width > 0 and base_img.width > 0:
segment_img = segment_img.resize(
(
base_img.width,
max(1, int(segment_img.height * (base_img.width / segment_img.width))),
),
Image.Resampling.LANCZOS,
)
target_width = max(base_img.width, segment_img.width)
base_img = _pad_width(base_img, target_width)
segment_img = _pad_width(segment_img, target_width)
stitched = Image.new(
"RGB",
(target_width, base_img.height + segment_img.height),
color=(255, 255, 255),
)
stitched.paste(base_img, (0, 0))
stitched.paste(segment_img, (0, base_img.height))
return stitched
def _render_pdf_page(
pdf_doc: pdfium.PdfDocument,
page_index: int,
scale: float,
cache: Dict[int, Image.Image],
) -> Optional[Image.Image]:
"""Render a PDF page to a PIL image with caching."""
if page_index in cache:
return cache[page_index]
try:
page = pdf_doc[page_index]
bitmap = page.render(scale=scale)
pil_img = bitmap.to_pil()
page.close()
except Exception as exc:
logger.error(f"Failed to render page {page_index + 1} for caption stitching: {exc}")
return None
cache[page_index] = pil_img
return pil_img
def _crop_pdf_region(
page_img: Optional[Image.Image], bbox: List[float]
) -> Optional[Image.Image]:
"""Crop a region from a rendered PDF page."""
if page_img is None:
return None
x0, y0, x1, y1 = map(int, bbox)
x0 = max(0, x0)
y0 = max(0, y0)
x1 = min(page_img.width, max(x0 + 1, x1))
y1 = min(page_img.height, max(y0 + 1, y1))
if x0 >= x1 or y0 >= y1:
return None
crop = page_img.crop((x0, y0, x1, y1))
if crop.mode == "CMYK":
crop = crop.convert("RGB")
return crop
def write_markdown_document(pdf_path: Path, out_dir: Path) -> Optional[Path]:
"""
Extract markdown text from a PDF using PyMuPDF4LLM and write it to disk.
"""
if pymupdf4llm is None:
logger.warning(
"Skipping markdown extraction for %s because pymupdf4llm is not installed.",
pdf_path.name,
)
return None
try:
markdown_content = pymupdf4llm.to_markdown(str(pdf_path))
except Exception as exc:
logger.error(f" Failed to create markdown for {pdf_path.name}: {exc}")
return None
if isinstance(markdown_content, list):
markdown_content = "\n\n".join(
part for part in markdown_content if isinstance(part, str)
)
if not isinstance(markdown_content, str):
logger.error(
f" Unexpected markdown output type {type(markdown_content)} for {pdf_path.name}"
)
return None
markdown_content = markdown_content.strip()
if not markdown_content:
logger.warning(f" No textual content extracted from {pdf_path.name}")
return None
if not markdown_content.endswith("\n"):
markdown_content += "\n"
md_path = out_dir / f"{pdf_path.stem}.md"
md_path.write_text(markdown_content, encoding="utf-8")
logger.info(f" Saved markdown to {md_path.name}")
return md_path
def _collect_text_under_title_cross_page(
title_det: Dict,
sorted_dets: List[Dict],
start_idx: int,
page_idx: int,
used_indices: Set[Tuple[int, int]],
settings: Optional[Dict[str, float]] = None,
) -> List[Dict]:
"""Collect text elements directly below a title on the next page."""
if settings is None:
settings = TITLE_TEXT_ASSOCIATION
texts: List[Dict] = []
title_box = title_det["bbox"]
last_bottom = title_box[3]
for follower in sorted_dets[start_idx + 1 :]:
det_index = follower.get("index")
if det_index is None or (page_idx, det_index) in used_indices:
continue
if follower["name"] == "title":
break
if follower["name"] != "text":
continue
text_box = follower["bbox"]
if text_box[1] < title_box[1]:
continue
gap = text_box[1] - last_bottom
if gap > settings["max_text_gap"]:
break
if _horizontal_overlap_ratio(title_box, text_box) < settings["min_overlap"]:
continue
texts.append(follower)
last_bottom = text_box[3]
return texts
def attach_cross_page_figure_captions(
elements: List[Dict],
all_dets: Sequence[Optional[List[Dict[str, Any]]]],
pdf_bytes: bytes,
out_dir: Path,
scale: float,
) -> List[Dict]:
"""
If a figure caption appears on the next page, stitch it to the prior figure.
"""
figures = [elem for elem in elements if elem.get("type") == "figure"]
if not figures or not all_dets:
return elements
try:
pdf_doc = pdfium.PdfDocument(pdf_bytes)
except Exception as exc:
logger.error(f"Unable to reopen PDF for figure caption stitching: {exc}")
return elements
page_cache: Dict[int, Image.Image] = {}
used_following_ids: Set[Tuple[int, int]] = set()
# Mark existing caption/title/text detections as used
for elem in figures:
for key in ("captions", "titles", "texts"):
for seg in elem.get(key, []) or []:
idx = seg.get("index")
page_no = seg.get("page")
if idx is None or page_no is None:
continue
used_following_ids.add((page_no - 1, idx))
for elem in figures:
page_no = elem.get("page")
bbox = elem.get("bbox_pixels")
if page_no is None or bbox is None:
continue
current_idx = page_no - 1
next_idx = current_idx + 1
if next_idx >= len(all_dets):
continue
next_dets = all_dets[next_idx]
if not next_dets:
continue
fig_width = bbox[2] - bbox[0]
page_img = _render_pdf_page(pdf_doc, next_idx, scale, page_cache)
if page_img is None:
continue
next_page_height = page_img.height
max_top_allowed = min(
CROSS_PAGE_CAPTION_THRESHOLDS["max_top_pixels"],
int(next_page_height * CROSS_PAGE_CAPTION_THRESHOLDS["max_top_ratio"]),
)
sorted_next = sorted(
[det for det in next_dets if det.get("bbox")],
key=lambda det: det["bbox"][1],
)
caption_candidate: Optional[Tuple[Dict, int]] = None
caption_candidates = []
for det in sorted_next:
if det.get("name") != "figure_caption":
continue
det_index = det.get("index")
if det_index is None or (next_idx, det_index) in used_following_ids:
continue
det_bbox = det.get("bbox")
if not det_bbox or det_bbox[1] > max_top_allowed:
continue
overlap = _horizontal_overlap_ratio(bbox, det_bbox)
x_diff = abs(bbox[0] - det_bbox[0])
width_diff = abs((bbox[2] - bbox[0]) - (det_bbox[2] - det_bbox[0]))
if overlap < CROSS_PAGE_CAPTION_THRESHOLDS["min_overlap"]:
if (
x_diff > CROSS_PAGE_CAPTION_THRESHOLDS["x_tol"]
or width_diff > CROSS_PAGE_CAPTION_THRESHOLDS["width_tol"]
):
continue
score = width_diff + 0.5 * x_diff
caption_candidates.append((score, det, det_index))
if caption_candidates:
caption_candidates.sort(key=lambda item: item[0])
_, best_det, best_index = caption_candidates[0]
caption_candidate = (best_det, best_index)
title_candidate: Optional[Tuple[Dict, int]] = None
title_texts: List[Dict] = []
for idx_sorted, det in enumerate(sorted_next):
if det.get("name") != "title":
continue
det_index = det.get("index")
if det_index is None or (next_idx, det_index) in used_following_ids:
continue
det_bbox = det.get("bbox")
if not det_bbox or det_bbox[1] > max_top_allowed:
continue
overlap = _horizontal_overlap_ratio(bbox, det_bbox)
x_diff = abs(bbox[0] - det_bbox[0])
if (
overlap < TITLE_TEXT_ASSOCIATION["min_overlap"]
and x_diff > CROSS_PAGE_CAPTION_THRESHOLDS["x_tol"]
):
continue
title_candidate = (det, det_index)
title_texts = _collect_text_under_title_cross_page(
det, sorted_next, idx_sorted, next_idx, used_following_ids
)
break
if not caption_candidate and not title_candidate and not title_texts:
continue
figure_path = out_dir / elem["image_path"]
if not figure_path.exists():
continue
figure_img = Image.open(figure_path)
if figure_img.mode == "CMYK":
figure_img = figure_img.convert("RGB")
segments_added = False
if caption_candidate:
cap_det, cap_index = caption_candidate
caption_crop = _crop_pdf_region(page_img, cap_det["bbox"])
if caption_crop is not None:
figure_img = _append_segment_image(
figure_img, caption_crop, resize_to_base=True
)
elem.setdefault("captions", [])
elem["captions"].append(
{
"bbox": cap_det["bbox"],
"conf": cap_det.get("conf"),
"index": cap_index,
"source": cap_det.get("source"),
"page": next_idx + 1,
}
)
used_following_ids.add((next_idx, cap_index))
segments_added = True
if title_candidate:
title_det, title_index = title_candidate
title_crop = _crop_pdf_region(page_img, title_det["bbox"])
if title_crop is not None:
figure_img = _append_segment_image(figure_img, title_crop)
elem.setdefault("titles", [])
elem["titles"].append(
{
"bbox": title_det["bbox"],
"conf": title_det.get("conf"),
"index": title_index,
"source": title_det.get("source"),
"page": next_idx + 1,
}
)
used_following_ids.add((next_idx, title_index))
segments_added = True
for text_det in title_texts:
text_index = text_det.get("index")
text_crop = _crop_pdf_region(page_img, text_det["bbox"])
if text_crop is None:
continue
figure_img = _append_segment_image(figure_img, text_crop)
elem.setdefault("texts", [])
elem["texts"].append(
{
"bbox": text_det["bbox"],
"conf": text_det.get("conf"),
"index": text_index,
"source": text_det.get("source"),
"page": next_idx + 1,
}
)
if text_index is not None:
used_following_ids.add((next_idx, text_index))
segments_added = True
if not segments_added:
continue
figure_img.save(figure_path)
elem["width"] = figure_img.width
elem["height"] = figure_img.height
span = elem.get("page_span")
if span:
if next_idx + 1 not in span:
span.append(next_idx + 1)
else:
base_page = elem.get("page")
new_span = [page for page in (base_page, next_idx + 1) if page is not None]
elem["page_span"] = new_span
pdf_doc.close()
return elements
def _stitch_table_pair(
base_elem: Dict,
candidate_elem: Dict,
out_dir: Path,
merge_index: int,
stitch_type: str,
) -> Optional[Dict]:
"""Stitch two table crops either vertically or horizontally."""
base_img = _open_table_image(base_elem, out_dir)
candidate_img = _open_table_image(candidate_elem, out_dir)
if base_img is None or candidate_img is None:
return None
tables_dir = out_dir / "tables"
tables_dir.mkdir(parents=True, exist_ok=True)
if stitch_type == "vertical":
target_width = max(base_img.width, candidate_img.width)
base_img = _pad_width(base_img, target_width)
candidate_img = _pad_width(candidate_img, target_width)
merged_height = base_img.height + candidate_img.height
stitched = Image.new("RGB", (target_width, merged_height), color=(255, 255, 255))
stitched.paste(base_img, (0, 0))
stitched.paste(candidate_img, (0, base_img.height))
else:
target_height = max(base_img.height, candidate_img.height)
base_img = _pad_height(base_img, target_height)
candidate_img = _pad_height(candidate_img, target_height)
merged_width = base_img.width + candidate_img.width
stitched = Image.new("RGB", (merged_width, target_height), color=(255, 255, 255))
stitched.paste(base_img, (0, 0))
stitched.paste(candidate_img, (base_img.width, 0))
merged_name = (
f"page_{base_elem['page']}_to_{candidate_elem['page']}_"
f"table_merged_{merge_index}.png"
)
merged_path = tables_dir / merged_name
stitched.save(merged_path)
# Remove original partial crops to avoid duplicates
(out_dir / base_elem["image_path"]).unlink(missing_ok=True)
(out_dir / candidate_elem["image_path"]).unlink(missing_ok=True)
new_bbox = [
min(base_elem["bbox_pixels"][0], candidate_elem["bbox_pixels"][0]),
min(base_elem["bbox_pixels"][1], candidate_elem["bbox_pixels"][1]),
max(base_elem["bbox_pixels"][2], candidate_elem["bbox_pixels"][2]),
max(base_elem["bbox_pixels"][3], candidate_elem["bbox_pixels"][3]),
]
merged_elem = base_elem.copy()
merged_elem["page_span"] = [base_elem["page"], candidate_elem["page"]]
merged_elem["box_refs"] = [
{"page": base_elem["page"], "image_path": base_elem["image_path"]},
{"page": candidate_elem["page"], "image_path": candidate_elem["image_path"]},
]
merged_elem["bbox_pixels"] = new_bbox
merged_elem["image_path"] = str(merged_path.relative_to(out_dir))
merged_elem["width"] = stitched.width
merged_elem["height"] = stitched.height
merged_elem["page_height"] = stitched.height
merged_elem["conf"] = min(
base_elem.get("conf", 1.0), candidate_elem.get("conf", 1.0)
)
return merged_elem
def merge_spanning_tables(elements: List[Dict], out_dir: Path) -> List[Dict]:
"""
Stitch table crops that continue across adjacent pages using the heuristic
from the legacy OpenCV-based extractor.
"""
if not elements:
return elements
tables_by_page: Dict[int, List[Dict]] = {}
non_tables: List[Dict] = []
for elem in elements:
if elem.get("type") != "table":
non_tables.append(elem)
continue
page = elem.get("page")
if not isinstance(page, int):
non_tables.append(elem)
continue
tables_by_page.setdefault(page, []).append(elem)
merged_results: List[Dict] = []
used_next: Dict[int, set[int]] = {}
merge_counter = 0
for page in sorted(tables_by_page.keys()):
current_tables = tables_by_page.get(page, [])
next_page_tables = tables_by_page.get(page + 1, [])
next_used_indices = used_next.get(page + 1, set())
current_used_indices = used_next.get(page, set())
for idx_current, table_elem in enumerate(current_tables):
if idx_current in current_used_indices:
continue
if not next_page_tables:
merged_results.append(table_elem)
continue
x, y, w, h = _bbox_to_rect(table_elem["bbox_pixels"])
matched = False
for idx, candidate in enumerate(next_page_tables):
if idx in next_used_indices:
continue
if candidate.get("type") != "table":
continue
cx, cy, cw, ch = _bbox_to_rect(candidate["bbox_pixels"])
vertical_match = (
abs(x - cx) <= TABLE_STITCH_TOLERANCES["x_tol"]
and abs((x + w) - (cx + cw)) <= TABLE_STITCH_TOLERANCES["width_tol"]
)
horizontal_match = (
abs(y - cy) <= TABLE_STITCH_TOLERANCES["y_tol"]
and abs((y + h) - (cy + ch))
<= TABLE_STITCH_TOLERANCES["height_tol"]
)
stitch_type = "vertical" if vertical_match else None
if not stitch_type and horizontal_match:
stitch_type = "horizontal"
if not stitch_type:
continue
merge_counter += 1
merged_elem = _stitch_table_pair(
table_elem, candidate, out_dir, merge_counter, stitch_type
)
if merged_elem is None:
continue
merged_results.append(merged_elem)
next_used_indices.add(idx)
matched = True
break
if not matched:
merged_results.append(table_elem)
used_next[page + 1] = next_used_indices
merged_results.extend(non_tables)
return merged_results
# ----------------------------------------------------------------------
# Draw layout boxes on the original PDF
# ----------------------------------------------------------------------
def draw_layout_pdf(pdf_bytes: bytes, all_dets: List[List[dict]],
scale: float, out_path: Path):
"""Annotate PDF with semi-transparent bounding boxes and labels."""
doc = fitz.open(stream=pdf_bytes, filetype="pdf")
for page_no, dets in enumerate(all_dets):
page = doc[page_no]
for d in dets:
rgb = CLASS_COLORS.get(d["name"], (0, 0, 0))
rect = fitz.Rect([c / scale for c in d["bbox"]])
border_color = [c / 255 for c in rgb]
fill_color = [c / 255 for c in rgb]
fill_opacity = 0.15
border_width = 1.5
page.draw_rect(
rect,
color=border_color,
fill=fill_color,
width=border_width,
overlay=True,
fill_opacity=fill_opacity
)
label = f"{d['name']} {d['conf']:.2f}"
if d.get("source"):
label += f" [{d['source'][0].upper()}]"
text_bg = fitz.Rect(rect.x0, rect.y0 - 10, rect.x0 + 60, rect.y0)
page.draw_rect(text_bg, color=None, fill=(1, 1, 1, 0.6), overlay=True)
page.insert_text(
(rect.x0 + 2, rect.y0 - 8),
label,
fontsize=6.5,
color=border_color,
overlay=True
)
doc.save(str(out_path))
doc.close()
# ----------------------------------------------------------------------
# Process a single PDF Page (for parallel execution)
# ----------------------------------------------------------------------
def process_page(task_data: Tuple[int, bytes, float, Path, str]) -> Optional[Tuple[int, List[dict], List[dict]]]:
"""
Process a single page of a PDF in a worker process.
Returns: (page_number, detections, elements) or None on failure
"""
pno, pdf_bytes, scale, out_dir, pdf_name = task_data
if _shutdown_requested:
return None
pdf_pdfium = None
try:
pdf_pdfium = pdfium.PdfDocument(pdf_bytes)
page = pdf_pdfium[pno]
bitmap = page.render(scale=scale)
pil = bitmap.to_pil()
dets = detect_page(pil)
elements = save_layout_elements(pil, pno, dets, out_dir)
page_figures = len([d for d in dets if d['name'] == 'figure'])
page_tables = len([d for d in dets if d['name'] == 'table'])
logger.info(f" [{pdf_name}] Page {pno + 1}: {page_figures} figs, {page_tables} tables")
page.close()
pdf_pdfium.close()
return (pno, dets, elements)
except Exception as e:
logger.error(f"Failed to process page {pno + 1} of {pdf_name}: {e}")
if pdf_pdfium:
pdf_pdfium.close()
return None
# ----------------------------------------------------------------------
# Process a full PDF using the persistent worker pool
# ----------------------------------------------------------------------
def process_pdf_with_pool(
pdf_path: Path,
out_dir: Path,
pool: Optional[Pool] = None,
*,
extract_images: bool = True,
extract_markdown: bool = True,
):
"""
Main processing pipeline for a PDF file.
If pool is provided, uses it. Otherwise processes serially.
"""
if _shutdown_requested:
logger.warning(f"Skipping {pdf_path.name} due to shutdown request")
return
stem = pdf_path.stem
logger.info(f"Processing {pdf_path.name}")
pdf_bytes = pdf_path.read_bytes()
doc = None
try:
doc = pdfium.PdfDocument(pdf_bytes)
page_count = len(doc)
except Exception as e:
logger.error(f"Failed to open PDF {pdf_path.name}: {e}. Skipping.")
return
finally:
if doc is not None:
doc.close()
scale = 2.0
all_elements: List[Dict] = []
filtered_dets: List[List[dict]] = []
if extract_images:
all_dets: List[Optional[List[dict]]] = [None] * page_count
if pool is not None and USE_MULTIPROCESSING:
logger.info(f" Using worker pool for {page_count} pages...")
tasks = [
(pno, pdf_bytes, scale, out_dir, pdf_path.name)
for pno in range(page_count)
]
try:
results = pool.map(process_page, tasks)
for res in results:
if res:
pno, dets, elements = res
all_dets[pno] = dets
all_elements.extend(elements)
except KeyboardInterrupt:
logger.warning("Processing interrupted during parallel execution")
raise
else:
logger.info("Using serial processing...")
try:
pdf_pdfium = pdfium.PdfDocument(pdf_bytes)
for pno in range(page_count):
if _shutdown_requested:
logger.warning(
f"Stopping at page {pno + 1}/{page_count} due to shutdown request"
)
break
try:
logger.info(f" Processing page {pno + 1}/{page_count}")
page = pdf_pdfium[pno]
bitmap = page.render(scale=scale)
pil = bitmap.to_pil()
dets = detect_page(pil)
all_dets[pno] = dets
elements = save_layout_elements(pil, pno, dets, out_dir)
all_elements.extend(elements)
page_figures = len([d for d in dets if d["name"] == "figure"])
page_tables = len([d for d in dets if d["name"] == "table"])
logger.info(
f" Found {page_figures} figures and {page_tables} tables"
)
page.close()
except Exception as e:
logger.error(f"Failed to process page {pno + 1}: {e}. Skipping page.")
pdf_pdfium.close()
except Exception as e:
logger.error(f"Fatal error processing {pdf_path.name}: {e}")
if "pdf_pdfium" in locals() and pdf_pdfium:
pdf_pdfium.close()
return
dets_per_page: List[Optional[List[Dict[str, Any]]]] = [
det if det is not None else None for det in all_dets
]
filtered_dets = [d for d in all_dets if d is not None]
if all_elements:
all_elements = merge_spanning_tables(all_elements, out_dir)
all_elements = attach_cross_page_figure_captions(
all_elements, dets_per_page, pdf_bytes, out_dir, scale
)
if all_elements:
content_list_path = out_dir / f"{stem}_content_list.json"
with open(content_list_path, "w", encoding="utf-8") as f:
json.dump(all_elements, f, ensure_ascii=False, indent=4)
logger.info(f" Saved {len(all_elements)} elements to JSON")
if filtered_dets:
draw_layout_pdf(
pdf_bytes, filtered_dets, scale, out_dir / f"{stem}_layout.pdf"
)
logger.info(" Generated annotated PDF")
else:
logger.warning(f"No detections found for {stem}. Skipping layout PDF.")
else:
logger.info(" Image extraction skipped per configuration.")
markdown_path = None
if extract_markdown:
markdown_path = write_markdown_document(pdf_path, out_dir)
if markdown_path is None:
logger.warning(f" Markdown extraction yielded no content for {stem}.")
if _shutdown_requested:
logger.warning(f"β οΈ Partial results saved for {stem} β {out_dir}")
else:
if extract_images:
logger.success(
f"β {stem} β {out_dir} ({len(all_elements)} elements extracted)"
)
else:
logger.success(f"β {stem} β {out_dir} (image extraction skipped)")
# ----------------------------------------------------------------------
# Main
# ----------------------------------------------------------------------
if __name__ == "__main__":
# Important for multiprocessing on Windows/macOS
torch.multiprocessing.set_start_method('spawn', force=True)
# Setup signal handlers for graceful shutdown
setup_signal_handlers()
INPUT_DIR = Path("./pdfs")
OUTPUT_DIR = Path("./output")
os.makedirs(INPUT_DIR, exist_ok=True)
os.makedirs(OUTPUT_DIR, exist_ok=True)
pdf_files = list(INPUT_DIR.glob("*.pdf"))
if not pdf_files:
logger.warning("No PDF files found in ./pdfs")
logger.info("Please add PDF files to the ./pdfs directory")
logger.info("The script will exit gracefully. No errors occurred.")
sys.exit(0)
logger.info(f"Found {len(pdf_files)} PDF file(s) to process")
logger.info(f"Settings: MODEL_SIZE={MODEL_SIZE}, CONF={CONF_THRESHOLD}")
# Determine worker count
total_cpus = cpu_count()
if NUM_WORKERS is None:
num_workers = max(1, total_cpus - 1)
else:
num_workers = max(1, min(NUM_WORKERS, total_cpus))
# Decide whether to use multiprocessing
use_pool = USE_MULTIPROCESSING and DEVICE == "cpu" and total_cpus >= 4
if use_pool:
logger.info(f"π Creating persistent worker pool with {num_workers} workers...")
else:
if not USE_MULTIPROCESSING:
logger.info("Multiprocessing disabled by configuration")
elif DEVICE != "cpu":
logger.info(f"Using serial GPU processing (device: {DEVICE})")
else:
logger.info(f"Using serial CPU processing (CPU count {total_cpus} too low)")
pool = None
try:
# Create persistent pool ONCE for all PDFs
if use_pool:
pool = Pool(processes=num_workers, initializer=init_worker)
logger.success(f"β Worker pool ready with {num_workers} workers\n")
else:
# Load model in main process for serial execution
logger.info("Initializing model in main process...")
get_model()
logger.success(f"β Model loaded (device: {DEVICE})\n")
# Process all PDFs using the same pool
for i, pdf_path in enumerate(pdf_files, 1):
if _shutdown_requested:
logger.warning(f"\nShutdown requested. Processed {i-1}/{len(pdf_files)} files.")
break
logger.info(f"\n{'='*60}")
logger.info(f"π File {i}/{len(pdf_files)}: {pdf_path.name}")
logger.info(f"{'='*60}")
sub_out = OUTPUT_DIR / pdf_path.stem
os.makedirs(sub_out, exist_ok=True)
try:
process_pdf_with_pool(pdf_path, sub_out, pool)
except KeyboardInterrupt:
logger.warning(f"\nInterrupted while processing {pdf_path.name}")
break
except Exception as e:
logger.error(f"Error processing {pdf_path.name}: {e}")
if _shutdown_requested:
break
logger.info("Continuing with next file...")
continue
if _shutdown_requested:
logger.warning(f"\nβ οΈ Processing interrupted. Partial results saved in {OUTPUT_DIR}")
else:
logger.success(f"\n⨠All done! Results are in {OUTPUT_DIR}")
except KeyboardInterrupt:
logger.error("\nβ Processing interrupted by user")
sys.exit(1)
except Exception as e:
logger.error(f"\nβ Fatal error: {e}")
sys.exit(1)
finally:
# Clean up pool if it exists
if pool is not None:
logger.info("\nπ§Ή Shutting down worker pool...")
pool.close()
pool.join()
logger.success("β Worker pool closed cleanly") |