Automatic page splitting
This commit is contained in:
sebastien committed
1 parent
d4404e4c27
commit
78f3e4b2f0
13 files changed
+1959
-40
No files matched your search
@@ -0,0 +1,377 @@
|
||||
"""Experimental, non-destructive spread splitting with an HTML review gallery.
|
||||
|
||||
python -m copienator.auto_page_split INPUT OUTPUT [--workers 4]
|
||||
No booklet reordering, secondary cuts, or GUI changes. Unknown formats and
|
||||
uncertain content are retained. OUTPUT must not already exist.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
from concurrent.futures import ProcessPoolExecutor
|
||||
import hashlib
|
||||
import html
|
||||
import json
|
||||
import multiprocessing
|
||||
from pathlib import Path
|
||||
from urllib.parse import quote
|
||||
|
||||
import cv2
|
||||
import numpy as np
|
||||
import pymupdf
|
||||
from PIL import Image, ImageDraw
|
||||
|
||||
from copienator.ink_detection import detect_bounds, detect_page_content
|
||||
|
||||
|
||||
def classify_format(width: float, height: float) -> str:
|
||||
"""Dimensions are displayed PDF points; tolerate 5% scan size error."""
|
||||
w, h = width * 25.4 / 72, height * 25.4 / 72
|
||||
if abs(min(w, h) / 210 - 1) < .05 and abs(max(w, h) / 297 - 1) < .05:
|
||||
return 'a4'
|
||||
if abs(w / 420 - 1) < .05 and abs(h / 297 - 1) < .05:
|
||||
return 'a3'
|
||||
return 'unknown'
|
||||
|
||||
|
||||
def analyse_ink(rgb: np.ndarray, dpi: float, *, spread: bool = False) -> tuple[np.ndarray, dict]:
|
||||
"""Use normal rognage on logical pages; colour cleanup only for overflow.
|
||||
|
||||
Blankness uses the shared crop detector's content-analysis entry point.
|
||||
No independent weak-pixel veto or secondary artifact detector is applied.
|
||||
"""
|
||||
if not spread:
|
||||
return detect_page_content(rgb,dpi)
|
||||
masks = {}
|
||||
result = detect_bounds(rgb, dpi, content_masks=masks,
|
||||
colour_cleanup=True, edge_mark_cleanup=False)
|
||||
ink = masks['foreground']
|
||||
# For ownership, follow connected faint pen strokes beyond rognage's local
|
||||
# 2 mm recovery radius. Do not restore page-wide connected ruling networks.
|
||||
candidate = ink | masks['weak_candidates']
|
||||
n,labels,stats,_ = cv2.connectedComponentsWithStats(candidate.astype(np.uint8),8)
|
||||
seeded = np.zeros(n,bool)
|
||||
seeded[np.unique(labels[ink])] = True
|
||||
seeded[0] = False
|
||||
giant = (stats[:,2] > .65*ink.shape[1]) & (stats[:,3] > .4*ink.shape[0])
|
||||
seeded[giant] = False
|
||||
ink = ink | seeded[labels]
|
||||
return ink, result
|
||||
|
||||
|
||||
def _analyse_page_full(rgb: np.ndarray, dpi: float, kind: str) -> dict:
|
||||
"""Keep spread geometry analysis separate from logical-page blankness."""
|
||||
h, w = rgb.shape[:2]
|
||||
boundaries = [0, w//2, w] if kind == 'a3' else [0, w]
|
||||
content = np.zeros((h,w), bool)
|
||||
evidence = []
|
||||
for x0,x1 in zip(boundaries,boundaries[1:]):
|
||||
foreground, result = analyse_ink(rgb[:,x0:x1],dpi)
|
||||
content[:,x0:x1] = foreground
|
||||
evidence.append(result)
|
||||
ink = analyse_ink(rgb,dpi,spread=True)[0] if kind == 'a3' else content
|
||||
plan = plan_masks(ink,dpi,kind,content_ink=content,rgb=rgb)
|
||||
for output,result in zip(plan['outputs'],evidence):
|
||||
output['crop_analysis'] = result
|
||||
return plan
|
||||
|
||||
|
||||
def analyse_page(rgb: np.ndarray, dpi: float, kind: str) -> dict:
|
||||
"""Analyse a raster with the standard coarse-to-fine splitter."""
|
||||
from copienator.fast_page_split import analyse_page_fast
|
||||
return analyse_page_fast(rgb, dpi, kind)
|
||||
|
||||
|
||||
def _fold_regions(ink: np.ndarray, credible: np.ndarray, dpi: float,
|
||||
rgb: np.ndarray | None = None) -> tuple[list[dict], np.ndarray]:
|
||||
"""Grow local word groups towards page anchors, stopping at supported text.
|
||||
|
||||
Broader gaps are allowed for ownership evidence without merging both pages
|
||||
into one giant line or changing output bounds to include the anchor word.
|
||||
"""
|
||||
import heapq
|
||||
px = dpi/25.4
|
||||
h,w = ink.shape
|
||||
middle = w//2
|
||||
ignored = []
|
||||
if rgb is not None:
|
||||
# Remove weak thin ruling *before* word grouping: otherwise a nearby
|
||||
# dark word can lend its colour/anchor evidence to a pale grid fragment.
|
||||
count,raw_labels,raw_stats,_ = cv2.connectedComponentsWithStats(ink.astype(np.uint8),8)
|
||||
ink = ink.copy()
|
||||
for i,(x,y,bw,bh,area) in enumerate(raw_stats[1:],1):
|
||||
if not (x < middle+2*px and x+bw > middle-2*px and bh<px and bw>4*bh):
|
||||
continue
|
||||
region = raw_labels[y:y+bh,x:x+bw]==i
|
||||
values = rgb[y:y+bh,x:x+bw][region].astype(np.int16)
|
||||
if np.median(np.ptp(values,axis=1))<=80 and np.median(values.min(axis=1))>=110:
|
||||
ink[y:y+bh,x:x+bw][region] = False
|
||||
ignored.append(dict(owner='artifact',box=[int(x),int(y),int(x+bw),int(y+bh)],
|
||||
component_id=0,anchor_boxes=[None,None],anchor_distance_mm=[None,None],reason='weak-thin-ruling'))
|
||||
kernel = np.ones((max(1,round(.5*px)),max(1,round(3*px))),np.uint8)
|
||||
joined = cv2.dilate(ink.astype(np.uint8),kernel)
|
||||
n,labels,stats,_ = cv2.connectedComponentsWithStats(joined,8)
|
||||
boxes = np.zeros((n,4),np.int32)
|
||||
areas = np.bincount(labels[ink],minlength=n)
|
||||
pen = np.zeros(n,bool)
|
||||
for i in range(1,n):
|
||||
x,y,bw,bh,_ = stats[i]
|
||||
yy,xx = np.nonzero((labels[y:y+bh,x:x+bw]==i) & ink[y:y+bh,x:x+bw])
|
||||
boxes[i] = [x+xx.min(),y+yy.min(),x+xx.max()+1,y+yy.max()+1]
|
||||
if rgb is not None and len(xx)>=.4*px*px:
|
||||
values=rgb[y+yy,x+xx].astype(np.int16)
|
||||
# Between the crop detector's weak (60) and strong (100) colour
|
||||
# thresholds: pale blue ruling must not become an ownership anchor.
|
||||
pen[i] = np.median(np.ptp(values,axis=1))>80 or np.median(values.min(axis=1))<110
|
||||
support = credible & ink
|
||||
# The blank classifier deliberately ignores short isolated pen marks.
|
||||
# A two-dimensional letter/terminal stroke can still link to its page here.
|
||||
support |= pen[labels] & ink
|
||||
left_limit,right_limit = max(0,round(middle-12*px)),min(w,round(middle+12*px))
|
||||
left_counts = np.bincount(labels[:,:left_limit][support[:,:left_limit]],minlength=n)
|
||||
right_counts = np.bincount(labels[:,right_limit:][support[:,right_limit:]],minlength=n)
|
||||
anchors = np.c_[left_counts>=max(2,.2*px*px),right_counts>=max(2,.2*px*px)]
|
||||
backed = np.bincount(labels[support],minlength=n)>=max(2,.12*px*px)
|
||||
backed[0] = False
|
||||
near = max(1,round(2*px))
|
||||
roots = np.unique(labels[:,max(0,middle-near):middle+near])
|
||||
regions = ignored
|
||||
for root in roots:
|
||||
if root == 0:continue
|
||||
x0,y0,x1,y1 = boxes[root]
|
||||
if x1 < middle-near or x0 > middle+near:continue
|
||||
thin_unanchored = (not anchors[root].any() and
|
||||
((not pen[root] and y1-y0<px and x1-x0>4*(y1-y0)) or areas[root]<.4*px*px))
|
||||
if thin_unanchored:
|
||||
regions.append(dict(owner='artifact',box=boxes[root].tolist(),component_id=int(root),anchor_boxes=[None,None],anchor_distance_mm=[None,None]))
|
||||
continue
|
||||
distances = np.full(n,np.inf)
|
||||
distances[root]=0
|
||||
queue=[(0.,int(root))]
|
||||
reached=[np.inf,np.inf]
|
||||
evidence=[None,None]
|
||||
# Only supported writing may carry a connection onwards. Otherwise
|
||||
# rows of grid fragments could form a bridge all the way into an answer.
|
||||
allowed = backed.copy();allowed[root]=True
|
||||
while queue:
|
||||
cost,node=heapq.heappop(queue)
|
||||
if cost != distances[node] or cost>45*px:continue
|
||||
if anchors[node].any():
|
||||
for side in (0,1):
|
||||
if anchors[node,side] and cost<reached[side]:
|
||||
reached[side]=cost;evidence[side]=boxes[node].tolist()
|
||||
continue
|
||||
a,b,c,d = boxes[node]
|
||||
dx=np.maximum(0,np.maximum(boxes[:,0]-c,a-boxes[:,2]))
|
||||
dy=np.maximum(0,np.maximum(boxes[:,1]-d,b-boxes[:,3]))
|
||||
# Wide mathematical spaces are plausible between strong pen
|
||||
# regions; unsupported shadows get only the shorter word gap.
|
||||
gap = 32*px if pen[root] else 8*px
|
||||
neighbours=np.flatnonzero(allowed & (dx<=gap) & (dy<=2*px))
|
||||
for other in neighbours:
|
||||
step=float(np.hypot(dx[other],3*dy[other]))+1
|
||||
new=cost+step
|
||||
if new<distances[other]:
|
||||
distances[other]=new;heapq.heappush(queue,(new,int(other)))
|
||||
if not np.isfinite(reached).any():owner='artifact'
|
||||
elif abs(reached[0]-reached[1])<=px:owner='uncertain'
|
||||
else:owner='left' if reached[0]<reached[1] else 'right'
|
||||
regions.append(dict(owner=owner,box=[int(x0),int(y0),int(x1),int(y1)],
|
||||
component_id=int(root),anchor_boxes=evidence,anchor_distance_mm=[round(c/px,2) if np.isfinite(c) else None for c in reached]))
|
||||
return regions,labels
|
||||
|
||||
|
||||
def plan_masks(ink: np.ndarray, dpi: float, kind: str, *,
|
||||
content_ink: np.ndarray | None = None,
|
||||
rgb: np.ndarray | None = None,
|
||||
_remaining_content: dict | None = None) -> dict:
|
||||
"""Produce rectangles in image coordinates and explicit review reasons."""
|
||||
h, w = ink.shape
|
||||
px = dpi / 25.4
|
||||
middle = w // 2
|
||||
reasons = []
|
||||
crossings = []
|
||||
left_end, right_start = middle, middle
|
||||
owned = (ink if content_ink is None else content_ink).copy()
|
||||
left_owned, right_owned = owned[:, :middle].copy(), owned[:, middle:].copy()
|
||||
if kind == 'a3':
|
||||
pad = round(2*px)
|
||||
crossings,group_labels = _fold_regions(ink,owned,dpi,rgb)
|
||||
for crossing in crossings:
|
||||
owner = crossing['owner']
|
||||
x0,y0,x1,y1 = crossing['box']
|
||||
if owner == 'artifact':
|
||||
continue
|
||||
if owner == 'left':
|
||||
left_end = max(left_end,min(w,x1+pad))
|
||||
if x1>middle:
|
||||
right_owned[y0:y1,max(x0,middle)-middle:x1-middle] &= ~(ink[y0:y1,max(x0,middle):x1] & (group_labels[y0:y1,max(x0,middle):x1]==crossing['component_id']))
|
||||
elif owner == 'right':
|
||||
right_start = min(right_start,max(0,x0-pad))
|
||||
if x0<middle:
|
||||
left_owned[y0:y1,x0:min(x1,middle)] &= ~(ink[y0:y1,x0:min(x1,middle)] & (group_labels[y0:y1,x0:min(x1,middle)]==crossing['component_id']))
|
||||
else:
|
||||
left_end = max(left_end,min(w,x1+pad))
|
||||
right_start = min(right_start,max(0,x0-pad))
|
||||
reasons.append('ambiguous-centre-writing')
|
||||
regions = [('left', [0,0,left_end,h], left_owned),
|
||||
('right', [right_start,0,w,h], right_owned)]
|
||||
else:
|
||||
regions = [('whole', [0,0,w,h], owned)]
|
||||
if kind == 'unknown':
|
||||
reasons.append('unknown-format-or-orientation')
|
||||
|
||||
outputs = []
|
||||
if _remaining_content is not None:
|
||||
_remaining_content.update({side: content for side,_,content in regions})
|
||||
for side, rect, content in regions:
|
||||
state = 'content' if content.any() else 'blank'
|
||||
if kind == 'unknown':
|
||||
state = 'uncertain'
|
||||
outputs.append(dict(side=side, rect_px=rect, state=state,
|
||||
keep=state != 'blank', ink_pixels=int(content.sum())))
|
||||
return dict(format=kind, size_px=[w,h], outputs=outputs, crossings=crossings,
|
||||
review_reasons=sorted(set(reasons)),
|
||||
extensions_mm=[round((left_end-middle)/px,2),
|
||||
round((middle-right_start)/px,2)] if kind == 'a3' else [0,0])
|
||||
|
||||
|
||||
def save_preview(rgb: np.ndarray, plan: dict, path: Path) -> None:
|
||||
im = Image.fromarray(rgb)
|
||||
im.thumbnail((1200, 1000))
|
||||
sx, sy = im.width/plan['size_px'][0], im.height/plan['size_px'][1]
|
||||
overlay = Image.new('RGBA', im.size)
|
||||
draw = ImageDraw.Draw(overlay)
|
||||
for output in plan['outputs']:
|
||||
x0,y0,x1,y1 = output['rect_px']
|
||||
box = (x0*sx,y0*sy,x1*sx-1,y1*sy-1)
|
||||
colour = '#187941' if output['side'] == 'left' else '#a832a6'
|
||||
draw.rectangle(box, outline=colour, width=3,
|
||||
fill=(255,30,30,65) if not output['keep'] else None)
|
||||
for crossing in plan['crossings']:
|
||||
x0,y0,x1,y1 = crossing['box']
|
||||
draw.rectangle((x0*sx,y0*sy,x1*sx,y1*sy), outline='#999999' if crossing['owner']=='artifact' else '#e88b00', width=2)
|
||||
Image.alpha_composite(im.convert('RGBA'), overlay).convert('RGB').save(path, quality=90)
|
||||
|
||||
|
||||
def process_pdf(job: tuple[str, str, int]) -> list[dict]:
|
||||
source_s, destination_s, dpi = job
|
||||
source, destination = Path(source_s), Path(destination_s)
|
||||
cv2.setNumThreads(1)
|
||||
digest = hashlib.sha256(source.read_bytes()).hexdigest()
|
||||
records = []
|
||||
with pymupdf.open(source) as doc, pymupdf.open() as result:
|
||||
for i, page in enumerate(doc):
|
||||
rotation = page.rotation
|
||||
kind = classify_format(page.rect.width, page.rect.height)
|
||||
from copienator.fast_page_split import analyse_pdf_page_fast
|
||||
plan,rgb = analyse_pdf_page_fast(page,dpi,kind)
|
||||
pixel_width,pixel_height = plan['size_px']
|
||||
plan.update(file=source.name, source_page=i+1, source_sha256=digest,
|
||||
source_rotation=rotation, source_cropbox=list(page.cropbox))
|
||||
preview = f'{source.stem}-{i+1:03}.jpg'
|
||||
plan['preview'] = f'review/{preview}'
|
||||
save_preview(rgb, plan, destination/'review'/preview)
|
||||
# Normalize through a full-page PDF form: remove_rotation() loses
|
||||
# non-default CropBoxes on quarter-turn pages in some MuPDF builds.
|
||||
w,h = page.rect.width,page.rect.height
|
||||
normalized = pymupdf.open()
|
||||
flat = normalized.new_page(width=w,height=h)
|
||||
page.set_rotation(0)
|
||||
if page.get_contents():
|
||||
flat.show_pdf_page(flat.rect,doc,i,rotate=-rotation)
|
||||
for output in plan['outputs']:
|
||||
x0,y0,x1,y1 = output['rect_px']
|
||||
clip = pymupdf.Rect(x0/pixel_width*w,y0/pixel_height*h,
|
||||
x1/pixel_width*w,y1/pixel_height*h)
|
||||
output['crop_points'] = list(clip)
|
||||
if output['keep']:
|
||||
target = result.new_page(width=clip.width,height=clip.height)
|
||||
# Truly empty PDF pages have no content stream to show.
|
||||
if flat.get_contents():
|
||||
target.show_pdf_page(target.rect,normalized,0,clip=clip)
|
||||
output['output_page'] = len(result)
|
||||
records.append(plan)
|
||||
normalized.close()
|
||||
# An entirely blank copy has no output PDF; its report is still kept.
|
||||
if len(result):
|
||||
result.save(destination/'pdfs'/source.name, garbage=3, deflate=True)
|
||||
if hashlib.sha256(source.read_bytes()).hexdigest() != digest:
|
||||
raise RuntimeError(f'Source changed during analysis: {source}')
|
||||
output_path = destination/'pdfs'/source.name
|
||||
expected = sum(o['keep'] for r in records for o in r['outputs'])
|
||||
if expected:
|
||||
with pymupdf.open(output_path) as check:
|
||||
if len(check) != expected or any(p.rect.is_empty for p in check):
|
||||
raise RuntimeError(f'Invalid output: {output_path}')
|
||||
return records
|
||||
|
||||
|
||||
def run(source: Path, destination: Path, *, dpi: int = 200, workers: int = 1) -> dict:
|
||||
if dpi < 100 or workers < 1:
|
||||
raise ValueError('Use dpi >= 100 and workers >= 1')
|
||||
sources = sorted(source.glob('*.pdf')) if source.is_dir() else [source]
|
||||
if not sources or any(not p.is_file() or p.suffix.lower() != '.pdf' for p in sources):
|
||||
raise ValueError('No input PDFs found')
|
||||
# Refuse reuse: this makes accidental source replacement impossible.
|
||||
destination.mkdir(parents=True, exist_ok=False)
|
||||
(destination/'review').mkdir()
|
||||
(destination/'pdfs').mkdir()
|
||||
jobs = [(str(p),str(destination),dpi) for p in sources]
|
||||
records = []
|
||||
if workers == 1:
|
||||
batches = map(process_pdf, jobs)
|
||||
for batch in batches:
|
||||
records.extend(batch)
|
||||
print(f'{batch[0]["file"]}: {len(batch)} scanned pages', flush=True)
|
||||
else:
|
||||
with ProcessPoolExecutor(max_workers=workers,
|
||||
mp_context=multiprocessing.get_context('spawn')) as pool:
|
||||
for batch in pool.map(process_pdf,jobs):
|
||||
records.extend(batch)
|
||||
print(f'{batch[0]["file"]}: {len(batch)} scanned pages', flush=True)
|
||||
return write_report(records,destination,len(sources))
|
||||
|
||||
|
||||
def write_report(records: list[dict], destination: Path, file_count: int) -> dict:
|
||||
"""Rebuild the gallery from saved decisions without reanalysing scans."""
|
||||
summary = dict(files=file_count, source_pages=len(records),
|
||||
a3=sum(r['format']=='a3' for r in records), a4=sum(r['format']=='a4' for r in records),
|
||||
output_pages=sum(o['keep'] for r in records for o in r['outputs']),
|
||||
blank_removed=sum(not o['keep'] for r in records for o in r['outputs']),
|
||||
review_pages=sum(bool(r['review_reasons']) for r in records),
|
||||
extended_spreads=sum(any(r['extensions_mm']) for r in records))
|
||||
(destination/'report.json').write_text(json.dumps(dict(summary=summary,pages=records),indent=2)+'\n')
|
||||
cards = []
|
||||
for row in records:
|
||||
name = html.escape(row['file'])
|
||||
states = ', '.join(f'{o["side"]}: {o["state"]}' for o in row['outputs'])
|
||||
reasons = ', '.join(row['review_reasons']) or 'No review flag'
|
||||
links = ' · '.join(f'<a href="pdfs/{quote(row["file"])}#page={o["output_page"]}">{o["side"]} PDF</a>'
|
||||
for o in row['outputs'] if o['keep'])
|
||||
cards.append(f'<article class="{"review" if row["review_reasons"] else "ok"}">'
|
||||
f'<h2>{name} / {row["source_page"]}</h2><p>{row["format"]} | {states}</p>'
|
||||
f'<p>Extensions L/R: {row["extensions_mm"]} mm | {reasons}</p><p>{links}</p>'
|
||||
f'<a href="{html.escape(row["preview"])}"><img loading="lazy" src="{html.escape(row["preview"])}"></a></article>')
|
||||
(destination/'index.html').write_text('<!doctype html><meta charset="utf-8">'
|
||||
'<title>Automatic page splitting - review</title><style>'
|
||||
'body{font:16px system-ui;margin:24px;background:#eee}main{display:grid;grid-template-columns:repeat(auto-fit,minmax(550px,1fr));gap:16px}'
|
||||
'article{background:white;padding:16px}article.review{border-top:5px solid #e88b00}img{width:100%}h2{font-size:18px}p{font-size:13px}</style>'
|
||||
'<h1>Automatic page splitting - experimental</h1><p>Source order; no booklet reordering. '
|
||||
'Green: left output. Purple: right/whole output. Orange: centre writing. Red shading: removed blank.</p>'
|
||||
f'<pre>{html.escape(json.dumps(summary,indent=2))}</pre><main>'+''.join(cards)+'</main>')
|
||||
return summary
|
||||
|
||||
|
||||
def main(argv=None) -> int:
|
||||
parser = argparse.ArgumentParser(description=__doc__)
|
||||
parser.add_argument('input',type=Path)
|
||||
parser.add_argument('output',type=Path)
|
||||
parser.add_argument('--dpi',type=int,default=200)
|
||||
parser.add_argument('--workers',type=int,default=1)
|
||||
args = parser.parse_args(argv)
|
||||
print(json.dumps(run(args.input,args.output,dpi=args.dpi,workers=args.workers),indent=2))
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
raise SystemExit(main())
|
||||
Reference in new issue
Block a user