Automatic page splitting

copienator/auto_page_split.py

@@ -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())

copienator/commands/automatic_page_splitter.py

@@ -0,0 +1,183 @@
+"""Publish automatic splits into an evaluation, preserving sources and review data."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import shutil
+import tempfile
+from pathlib import Path
+
+from copienator import CliError, ExitCode
+from copienator.auto_page_split import run as analyse_batch
+from copienator.copy_errors import clear_copy_error, copy_errors, mark_copy_error
+from copienator.filesystem import publish_file_sets
+from copienator.orientation import report_path
+
+
+def decision_path(workspace, stem):
+    return workspace.metadata_dir / "page_splitting" / f"{stem}.json"
+
+
+def load_decisions(workspace, source):
+    """Only reuse decisions for the exact preserved input and published output."""
+    try:
+        data = json.loads(decision_path(workspace, source.stem).read_text())
+        if data["source_sha256"] != hashlib.sha256(source.read_bytes()).hexdigest():
+            return None
+        output = workspace.copies_dir / source.name
+        digest = (
+            hashlib.sha256(output.read_bytes()).hexdigest() if output.exists() else None
+        )
+        if digest != data["output_sha256"]:
+            return None
+        pages = data["pages"]
+        if not isinstance(pages, list) or not all(
+            isinstance(row, dict)
+            and type(row.get("source_page")) is int
+            and isinstance(row.get("outputs"), list)
+            and all(
+                isinstance(o, dict)
+                and o.get("side") in {"left", "right", "whole"}
+                and type(o.get("keep")) is bool
+                and isinstance(o.get("crop_points"), list)
+                and len(o["crop_points"]) == 4
+                for o in row["outputs"]
+            )
+            for row in pages
+        ):
+            return None
+        return pages
+    except (OSError, ValueError, KeyError, TypeError, AttributeError):
+        return None
+
+
+def run(workspace, inputs: list[Path], *, workers=4):
+    if workers < 1:
+        raise CliError(
+            "Le nombre de traitements parallèles doit être positif.",
+            ExitCode.INVALID_ARGUMENTS,
+        )
+    if not inputs:
+        print("Aucune copie PDF à traiter.", flush=True)
+        return ExitCode.SUCCESS
+    if len({p.name for p in inputs}) != len(inputs):
+        raise CliError(
+            "Plusieurs copies portent le même nom.", ExitCode.INVALID_ARGUMENTS
+        )
+    copy_errors(workspace)  # Validate existing review metadata before publishing.
+    workspace.runs_dir.mkdir(parents=True, exist_ok=True)
+    run_dir = Path(tempfile.mkdtemp(prefix="page-split-", dir=workspace.runs_dir))
+    snapshots = run_dir / "inputs"
+    snapshots.mkdir()
+    for source in inputs:
+        shutil.copy2(source, snapshots / source.name)
+    result = run_dir / "result"
+    print(f"Découpage automatique de {len(inputs)} copie(s).", flush=True)
+    summary = analyse_batch(snapshots, result, workers=workers)
+    records = json.loads((result / "report.json").read_text())["pages"]
+    stages = {
+        name: run_dir / name for name in ("copies", "originals", "origins", "decisions")
+    }
+    for stage in stages.values():
+        stage.mkdir()
+    for source in inputs:
+        rows = [r for r in records if r["file"] == source.name]
+        expected = rows[0]["source_sha256"]
+        if hashlib.sha256(source.read_bytes()).hexdigest() != expected:
+            raise CliError(
+                f"{source.name} a changé pendant l’analyse. Aucun PDF remplacé."
+            )
+        shutil.copy2(snapshots / source.name, stages["originals"] / source.name)
+        for directory, label in (
+            (workspace.copies_dir, "previous_copies"),
+            (workspace.original_copies_dir, "previous_originals"),
+        ):
+            prior = directory / source.name
+            if prior.exists():
+                backup = run_dir / label
+                backup.mkdir(exist_ok=True)
+                shutil.copy2(prior, backup / source.name)
+        output = result / "pdfs" / source.name
+        digest = (
+            hashlib.sha256(output.read_bytes()).hexdigest() if output.exists() else None
+        )
+        if digest:
+            shutil.copy2(output, stages["copies"] / source.name)
+        origins = [
+            {"origin": r["source_page"], "crop_points": o["crop_points"]}
+            for r in rows
+            for o in r["outputs"]
+            if o["keep"]
+        ]
+        if digest:
+            (stages["origins"] / (source.stem + ".json")).write_text(
+                json.dumps({"version": 1, "sha256": digest, "pages": origins}),
+                encoding="utf-8",
+            )
+        (stages["decisions"] / (source.stem + ".json")).write_text(
+            json.dumps(
+                {
+                    "version": 1,
+                    "source_sha256": expected,
+                    "output_sha256": digest,
+                    "pages": rows,
+                }
+            ),
+            encoding="utf-8",
+        )
+    names = [p.name for p in inputs]
+    for source in inputs:
+        if source.read_bytes() != (snapshots / source.name).read_bytes():
+            raise CliError(
+                f"{source.name} a changé pendant l’analyse. Aucun PDF remplacé."
+            )
+    publish_file_sets(
+        {
+            workspace.copies_dir: (stages["copies"], names),
+            workspace.original_copies_dir: (stages["originals"], []),
+            workspace.metadata_dir / "page_origins": (
+                stages["origins"],
+                [p.stem + ".json" for p in inputs],
+            ),
+            workspace.metadata_dir / "page_splitting": (stages["decisions"], []),
+        }
+    )
+    for source in inputs:
+        if (
+            source.parent.resolve()
+            not in {
+                workspace.copies_dir.resolve(),
+                workspace.original_copies_dir.resolve(),
+            }
+            and source.read_bytes() == (snapshots / source.name).read_bytes()
+        ):
+            # Root scans become preserved originals, as in the manual workflow.
+            source.unlink()
+        rows = [r for r in records if r["file"] == source.name]
+        output = workspace.copies_dir / source.name
+        clear_copy_error(workspace, output, source="orientation")
+        report_path(workspace, source.stem).unlink(missing_ok=True)
+        reasons = {reason for r in rows for reason in r["review_reasons"]}
+        if reasons or not output.exists():
+            mark_copy_error(
+                workspace,
+                output,
+                "Découpage automatique à vérifier manuellement."
+                if reasons
+                else "Toutes les pages semblent vides : vérifier l’original.",
+                source="page_splitter",
+            )
+        else:
+            clear_copy_error(workspace, output, source="page_splitter")
+    print(
+        f"Terminé : {summary['output_pages']} pages conservées, {summary['blank_removed']} pages vides retirées. "
+        "Ordre des pages inchangé.",
+        flush=True,
+    )
+    print(
+        f"Rapport et sauvegardes : {run_dir}\n"
+        "Choisissez Manuellement pour vérifier ou modifier les propositions.",
+        flush=True,
+    )
+    return ExitCode.SUCCESS

copienator/commands/page_splitter.py

@@ -27,7 +27,7 @@ from copienator import (
 )
 from copienator.platform import launch_pdf_arranger
 from copienator.copy_errors import marked_copy_paths, clear_copy_error
-from copienator.orientation import save_provenance, report_path
+from copienator.orientation import save_provenance, report_path, provenance_path
 
 # --- Constants ---
 # Conversion factor: 1 cm to points (1 inch = 2.54 cm, 72 points = 1 inch)
@@ -49,7 +49,7 @@ def _temporary_sibling(path: Path, purpose: str) -> Path:
 def commit_processed_pdf(
     workspace: EvaluationWorkspace,
     original_path: Path,
-    generated_path: Path,
+    generated_path: Path | None,
 ) -> Path:
     """Commit a processed copy and its original backup with rollback."""
     backup_path = workspace.original_copies_dir / original_path.name
@@ -73,8 +73,9 @@ def commit_processed_pdf(
             backup_replaced = True
         if output_path.exists():
             output_path.replace(saved_output)
-        generated_path.replace(output_path)
-        output_replaced = True
+        if generated_path is not None:
+            generated_path.replace(output_path)
+            output_replaced = True
         if original_path.resolve() not in {
             backup_path.resolve(),
             output_path.resolve(),
@@ -120,6 +121,10 @@ class PDFPreviewer:
         try:
             self.doc = pymupdf.open(self.pdf_path)
             self.original_page_ids = {page.xref: index + 1 for index, page in enumerate(self.doc)}
+            from copienator.commands.automatic_page_splitter import load_decisions
+            saved = load_decisions(self.workspace, self.pdf_path)
+            self.automatic_plans = {r['source_page']: r for r in saved} if saved else {}
+            self.automatic_review = bool(self.automatic_plans)
         except (OSError, RuntimeError, ValueError) as e:
             self.failed = True
             self._temporary_directory.cleanup()
@@ -176,6 +181,12 @@ class PDFPreviewer:
             f"{fmt('add_secondary_splitter')}: add secondary line, {fmt('delete_secondary_splitter')}: delete newest (arrows move newest; main line red, active secondary orange)\n"
             f"'{fmt('reverse_pages')}': reverse page order and restart from the new first page\n"
         )
+        if self.automatic_review:
+            instructions += (
+                "\nPropositions automatiques : la touche suivante valide le choix affiché. "
+                "Croix rouge : page retirée.\n"
+                "Déplacer une ligne reprend le découpage manuel pour cette page."
+            )
 
         self.info_label = tk.Label(master, text=instructions, justify=tk.LEFT)
         self.info_label.pack(pady=5, side=tk.TOP)
@@ -255,6 +266,25 @@ class PDFPreviewer:
             self.current_line_x = page.rect.width / 2
             self.current_rotation = 0
             self.secondary_splitters = []
+            origin = getattr(self, 'original_page_ids', {}).get(page.xref, self.current_page_index+1)
+            plan = getattr(self, 'automatic_plans', {}).get(origin)
+            self.current_automatic_outputs = plan['outputs'] if plan else None
+
+    def _automatic_outputs(self):
+        """Express saved asymmetric crops in the current preview orientation."""
+        outputs = getattr(self, 'current_automatic_outputs', None)
+        if outputs is None:
+            return None
+        flipped = (self.current_rotation+self.file_rotation+self.global_rotation) % 360
+        if not flipped:
+            return outputs
+        width, height = self.doc[self.current_page_index].rect.width, self.doc[self.current_page_index].rect.height
+        result = []
+        for output in outputs:
+            x0,y0,x1,y1 = output['crop_points']
+            result.append(dict(output, side={'left':'right','right':'left'}.get(output['side'], 'whole'),
+                               crop_points=[width-x1,height-y1,width-x0,height-y0]))
+        return result
 
     def load_page(self):
         """Loads and displays the current page on the canvas, scaled to fit."""
@@ -296,9 +326,20 @@ class PDFPreviewer:
         # --- Draw Line and Rotate ---
         draw = ImageDraw.Draw(img)
         # The line position is scaled by the same zoom factor
-        line_x_scaled = self.current_line_x * self.current_zoom
-        draw.line([(line_x_scaled, 0), (line_x_scaled, pix.height)],
-                  fill="red", width=3)
+        proposed = self._automatic_outputs()
+        if proposed is not None:
+            for output in proposed:
+                x0,y0,x1,y1 = output['crop_points']
+                box = tuple(v*self.current_zoom for v in (x0,y0,x1,y1))
+                draw.rectangle(box, outline='green' if output['side']=='left' else 'purple', width=3)
+                if not output['keep']:
+                    draw.line((box[0],box[1],box[2],box[3]), fill='red', width=3)
+                    draw.line((box[0],box[3],box[2],box[1]), fill='red', width=3)
+            self.page_label.config(text=self.page_label.cget('text')+' — Proposition automatique (croix rouge : page retirée)')
+        else:
+            line_x_scaled = self.current_line_x * self.current_zoom
+            draw.line([(line_x_scaled, 0), (line_x_scaled, pix.height)],
+                      fill="red", width=3)
         for index, position in enumerate(self.secondary_splitters):
             x = position * self.current_zoom
             active = index == len(self.secondary_splitters) - 1
@@ -353,6 +394,7 @@ class PDFPreviewer:
     def add_secondary_splitter(self, event=None):
         """Start a new line in the widest remaining interval of this page."""
         width = self.doc.load_page(self.current_page_index).rect.width
+        self.current_automatic_outputs = None
         boundaries = sorted({0, width, self.current_line_x, *self.secondary_splitters})
         left, right = max(zip(boundaries, boundaries[1:]), key=lambda pair: pair[1] - pair[0])
         self.secondary_splitters.append((left + right) / 2)
@@ -364,6 +406,7 @@ class PDFPreviewer:
             self.load_page()
 
     def _move_active_line(self, delta):
+        self.current_automatic_outputs = None
         width = self.doc.load_page(self.current_page_index).rect.width
         if self.secondary_splitters:
             self.secondary_splitters[-1] = min(width, max(0, self.secondary_splitters[-1] + delta))
@@ -415,13 +458,19 @@ class PDFPreviewer:
         # Reload the canvas to show the new page
         self.load_page()
 
-    def confirm_and_next_page(self, event=None, keep="both"):
+    def confirm_and_next_page(self, event=None, keep=None):
         """Saves the settings for the current page and moves to the next."""
+        proposed = self._automatic_outputs()
+        if keep is None:
+            sides = {o['side'] for o in proposed if o['keep']} if proposed is not None else {'left','right'}
+            keep = ('as_is' if 'whole' in sides else 'both' if len(sides)==2
+                    else next(iter(sides)) if sides else 'none')
         self.page_settings.append({
             "line_x": self.current_line_x,
             "secondary_splitters": list(self.secondary_splitters),
             "rotation": self.current_rotation,
-            "keep": keep
+            "keep": keep,
+            "automatic_outputs": proposed,
         })
 
         self.current_page_index += 1
@@ -442,12 +491,18 @@ class PDFPreviewer:
     def finish_and_process(self) -> bool:
         """Render and transactionally install the processed PDF."""
         try:
-            self.split_pdf()
-            self.reorder_pdfs()
-            self.concate_files()
-            output = commit_processed_pdf(self.workspace, self.pdf_path, self.final_file)
-            save_provenance(self.workspace, output, self.output_origins)
+            if all(s['keep']=='none' for s in self.page_settings):
+                self.doc.close()
+                output = commit_processed_pdf(self.workspace, self.pdf_path, None)
+                provenance_path(self.workspace, output.stem).unlink(missing_ok=True)
+            else:
+                self.split_pdf()
+                self.reorder_pdfs()
+                self.concate_files()
+                output = commit_processed_pdf(self.workspace, self.pdf_path, self.final_file)
+                save_provenance(self.workspace, output, self.output_origins)
             clear_copy_error(self.workspace, output, source='orientation')
+            clear_copy_error(self.workspace, output, source='page_splitter')
             report_path(self.workspace, output.stem).unlink(missing_ok=True)
         except Exception as exc:  # noqa: BLE001 - interactive boundary
             self.failed = True
@@ -514,15 +569,27 @@ class PDFPreviewer:
             line_x = settings['line_x']
             rotation_settings = settings['rotation']
             keep = settings['keep']
-            # Bake source rotation and the preview's extra rotation into content,
-            # so all cut coordinates refer to the displayed page.
-            page.remove_rotation()
-            page.set_rotation((rotation_settings + self.file_rotation + self.global_rotation) % 360)
-            page.remove_rotation()
+            # Normalize the displayed page without losing an offset CropBox.
+            # This matches the automatic export's coordinate system.
+            source_rotation = page.rotation
             width, height = page.rect.width, page.rect.height
+            normalized = pymupdf.open()
+            flat = normalized.new_page(width=width, height=height)
+            page.set_rotation(0)
+            if page.get_contents():
+                flat.show_pdf_page(flat.rect, self.doc, i, rotate=-source_rotation)
+            flat.set_rotation((rotation_settings+self.file_rotation+self.global_rotation) % 360)
+            flat.remove_rotation()
             sides = []
             if keep == "as_is":
                 sides.append((0, width, self.split_filename_left(i)))
+            elif settings.get('automatic_outputs') is not None:
+                for output in settings['automatic_outputs']:
+                    side = output['side']
+                    if keep == 'both' or keep == side:
+                        x0,_,x1,_ = output['crop_points']
+                        filename = self.split_filename_left(i) if side=='left' else self.split_filename_right(i)
+                        sides.append((x0,x1,filename))
             else:
                 if keep in {"both", "left"} and line_x > 0:
                     sides.append((0, line_x, self.split_filename_left(i)))
@@ -539,10 +606,12 @@ class PDFPreviewer:
                     for x0, x1 in zip(boundaries, boundaries[1:]):
                         clip = pymupdf.Rect(x0, 0, x1, height)
                         target = output.new_page(width=clip.width, height=clip.height)
-                        target.show_pdf_page(target.rect, self.doc, i, clip=clip)
+                        if flat.get_contents():
+                            target.show_pdf_page(target.rect, normalized, 0, clip=clip)
                         origins.append({'origin': origin, 'crop_points': list(clip)})
                     output.save(output_path)
                 self.fragment_origins[str(output_path)] = origins
+            normalized.close()
 
         self.doc.close()
         print(f"\nProcessing complete. Files are in '{self.split_dir}' directory.")
@@ -564,7 +633,7 @@ class PDFPreviewer:
             psk = ps[i]['keep']
 
             # Si c'est une copie double (on s'assure qu'on a bien 2 pages consécutives modifiables)
-            if psk in ["both", "right", "left", "none"] and i < len(ps)-1 and ps[i+1]['keep'] in ["both", "right", "left", "none"]:
+            if not getattr(self, 'automatic_review', False) and psk in ["both", "right", "left", "none"] and i < len(ps)-1 and ps[i+1]['keep'] in ["both", "right", "left", "none"]:
 
                 # 1. Page de garde (Extérieur Droit)
                 if ps[i]['keep'] in ["both", "right"]:
@@ -678,11 +747,17 @@ def _selected_inputs(
         ]
     else:
         candidates = list_pdf_files(directory, workspace)
+        if target == workspace.root and not candidates:
+            candidates = list_pdf_files(workspace.original_copies_dir, workspace)
     return list(reversed(candidates))
 
 
-def run(workspace: EvaluationWorkspace, target: Path, *, marked: bool = False) -> ExitCode:
+def run(workspace: EvaluationWorkspace, target: Path, *, marked: bool = False,
+        automatic: bool = False, workers: int = 4) -> ExitCode:
     inputs = list(reversed(marked_copy_paths(workspace, originals=True))) if marked else _selected_inputs(workspace, target)
+    if automatic:
+        from copienator.commands.automatic_page_splitter import run as run_automatic
+        return run_automatic(workspace, list(reversed(inputs)), workers=workers)
     if not inputs:
         print(f"No PDF files found in {target}")
         return ExitCode.SUCCESS
@@ -695,6 +770,8 @@ def run(workspace: EvaluationWorkspace, target: Path, *, marked: bool = False) -
 def build_parser() -> argparse.ArgumentParser:
     parser = target_parser("Interactively split and reorder scanned PDF pages")
     parser.add_argument("--marked", action="store_true", help="Process only copies flagged during margin review")
+    parser.add_argument("--automatic", action="store_true", help="Split automatically and save editable proposals without reordering")
+    parser.add_argument("--workers", type=int, default=4, help="Copies analysed simultaneously in automatic mode")
     return parser
 
 
@@ -703,7 +780,7 @@ def main(argv: Sequence[str] | None = None) -> int:
 
     def handle(args: argparse.Namespace) -> ExitCode:
         workspace, target = workspace_from_target(args)
-        return run(workspace, target, marked=args.marked)
+        return run(workspace, target, marked=args.marked, automatic=args.automatic, workers=args.workers)
 
     return execute(parser, argv, handle)
 

copienator/copy_errors.py

@@ -20,7 +20,7 @@ def _validate(value) -> dict[str, dict[str, str]]:
         if isinstance(flags, str):
             flags = {"cutleft": flags}
         if (not isinstance(flags, dict) or not flags
-                or any(step not in {"cutleft", "plotting", "orientation"} or not isinstance(reason, str)
+                or any(step not in {"cutleft", "plotting", "orientation", "page_splitter"} or not isinstance(reason, str)
                        for step, reason in flags.items())):
             raise CliError("Invalid copy_errors.json: invalid step flags")
         normalized[name] = dict(flags)
@@ -35,7 +35,7 @@ def copy_errors(workspace: EvaluationWorkspace) -> dict[str, str]:
 
 def mark_copy_error(workspace: EvaluationWorkspace, pdf: Path, reason: str,
                     *, source: str = "cutleft") -> None:
-    if source not in {"cutleft", "plotting", "orientation"}:
+    if source not in {"cutleft", "plotting", "orientation", "page_splitter"}:
         raise ValueError(f"Unknown copy flag source: {source}")
     def update(errors):
         normalized = _validate(errors)
@@ -47,7 +47,7 @@ def mark_copy_error(workspace: EvaluationWorkspace, pdf: Path, reason: str,
 
 def clear_copy_error(workspace: EvaluationWorkspace, pdf: Path,
                      *, source: str = "cutleft") -> None:
-    if source not in {"cutleft", "plotting", "orientation"}:
+    if source not in {"cutleft", "plotting", "orientation", "page_splitter"}:
         raise ValueError(f"Unknown copy flag source: {source}")
     if not _path(workspace).exists():
         return

copienator/fast_page_split.py

@@ -0,0 +1,276 @@
+"""Standard coarse-to-fine splitting: cheap positive content gate, precise fold.
+
+The whole-page gate only establishes obvious writing. All uncertain/sparse pages
+use the reference blank detector. Full-resolution fold processing is restricted
+to a central strip; edge-reaching overflow and ambiguous ownership fall back to
+the complete reference analysis. Full-spread analysis remains an internal fallback.
+"""
+
+from __future__ import annotations
+import cv2
+import numpy as np
+import pymupdf
+from copienator.auto_page_split import analyse_ink, _analyse_page_full, plan_masks
+
+
+def quick_page_content(rgb, dpi, target_dpi=75, *, native=True, _fallback=None):
+    """Positive-only coarse gate. Ambiguous/sparse pages use full detection."""
+    h, w = rgb.shape[:2]
+    small = cv2.resize(
+        rgb,
+        (round(w * target_dpi / dpi), round(h * target_dpi / dpi)),
+        interpolation=cv2.INTER_AREA,
+    )
+    px = target_dpi / 25.4
+    lo = small.min(axis=2)
+    ch = small.max(axis=2) - lo
+    gray = cv2.cvtColor(small, cv2.COLOR_RGB2GRAY)
+    strong = (ch > 100) | (gray < 95)
+    n, lab, stats, _ = cv2.connectedComponentsWithStats(strong.astype("uint8"), 8)
+    ok = np.zeros(n, bool)
+    x, y, bw, bh, area = stats[1:].T
+    # Only compact interior pen components can establish an obvious written page.
+    ok[1:] = (
+        (area >= 0.25 * px * px)
+        & (bw >= 0.35 * px)
+        & (bh >= 0.5 * px)
+        & (bw < 20 * px)
+        & (bh < 15 * px)
+        & (x > 0.09 * small.shape[1])
+        & (x + bw < 0.91 * small.shape[1])
+    )
+    selected = ok[lab]
+    grouped = cv2.dilate(
+        selected.astype("uint8"),
+        np.ones((max(1, round(px)), max(1, round(3 * px))), np.uint8),
+    )
+    ng, groups, gs, _ = cv2.connectedComponentsWithStats(grouped, 8)
+    comp_group = np.zeros(n, np.int32)
+    ys, xs = np.nonzero(selected)
+    comp_group[lab[ys, xs]] = groups[ys, xs]
+    counts = np.bincount(comp_group[ok], minlength=ng)
+    amounts = np.bincount(groups[selected], minlength=ng)
+    good = (
+        (counts >= 4)
+        & (amounts >= 3 * px * px)
+        & (gs[:, 2] >= 8 * px)
+        & (gs[:, 3] >= 2 * px)
+    )
+    good[0] = False
+    if good.any():
+        # Retain all strong ink for ownership support; the gate alone is interior-only.
+        return (
+            strong
+            if native
+            else cv2.resize(
+                strong.astype("uint8"), (w, h), interpolation=cv2.INTER_NEAREST
+            )
+            > 0
+        ), True
+    return (_fallback() if _fallback is not None else analyse_ink(rgb, dpi)[0]), False
+
+
+def _a4_plan(mask, gate, h, w, dpi, gate_dpi):
+    count = int(mask.sum())
+    return dict(
+        format="a4",
+        size_px=[w, h],
+        outputs=[
+            dict(
+                side="whole",
+                rect_px=[0, 0, w, h],
+                state="content" if count else "blank",
+                keep=bool(count),
+                ink_pixels=count,
+                content_dpi=gate_dpi if gate else dpi,
+            )
+        ],
+        crossings=[],
+        review_reasons=[],
+        extensions_mm=[0, 0],
+        analysis=dict(
+            mode="fast",
+            early_written=[gate],
+            fold_fraction=0,
+            gate_dpi=gate_dpi,
+            fallback=None,
+        ),
+    )
+
+
+def analyse_page_fast(rgb, dpi, kind, *, fold_fraction=0.1, gate_dpi=75):
+    """Fast analysis for an existing raster image."""
+    if kind not in {"a3", "a4"}:
+        return _analyse_page_full(rgb, dpi, kind)
+    gate_dpi = min(gate_dpi, dpi)
+    h, w = rgb.shape[:2]
+    if kind == "a4":
+        mask, gate = quick_page_content(rgb, dpi, gate_dpi, native=True)
+        return _a4_plan(mask, gate, h, w, dpi, gate_dpi)
+    middle = w // 2
+    halves = []
+    early = []
+    for a, b in ((0, middle), (middle, w)):
+        mask, gate = quick_page_content(rgb[:, a:b], dpi, gate_dpi, native=True)
+        halves.append(mask)
+        early.append(gate)
+    return _assemble_fast(
+        halves,
+        early,
+        h,
+        w,
+        dpi,
+        fold_fraction,
+        gate_dpi,
+        lambda x0, x1: rgb[:, x0:x1],
+        lambda: _analyse_page_full(rgb, dpi, kind),
+    )
+
+
+def _assemble_fast(
+    halves, early, h, w, dpi, fold_fraction, gate_dpi, render_roi, reference
+):
+    middle = w // 2
+    half = max(1, round(w * fold_fraction / 2))
+    x0 = max(0, middle - half)
+    x1 = min(w, middle + half)
+    roi = render_roi(x0, x1)
+    content = np.zeros(roi.shape[:2], bool)
+    for i, (a, b) in enumerate(((0, middle), (middle, w))):
+        lo = max(a, x0)
+        hi = min(b, x1)
+        mask = halves[i]
+        ys = np.arange(h) * mask.shape[0] // h
+        xs = (np.arange(lo, hi) - a) * mask.shape[1] // (b - a)
+        content[:, lo - x0 : hi - x0] = mask[ys[:, None], xs[None, :]]
+    ink, _ = analyse_ink(roi, dpi, spread=True)
+    remaining = {}
+    plan = plan_masks(
+        ink, dpi, "a3", content_ink=content, rgb=roi, _remaining_content=remaining
+    )
+    # If a proposed overflow reaches the strip's edge or ownership remains
+    # ambiguous, use the complete reference analysis rather than truncate it.
+    margin = round(2 * dpi / 25.4)
+    if (
+        plan["review_reasons"]
+        or plan["outputs"][0]["rect_px"][2] >= x1 - x0 - margin
+        or plan["outputs"][1]["rect_px"][0] <= margin
+    ):
+        plan = reference()
+        plan["analysis"] = dict(mode="fast", fallback="full-spread")
+        return plan
+    for i, (a, b) in enumerate(((0, middle), (middle, w))):
+        mask = halves[i]
+        side = "left" if i == 0 else "right"
+        if not early[i]:
+            lo = max(a, x0)
+            hi = min(b, x1)
+            mask[:, lo - a : hi - a] = remaining[side]
+            count = int(mask.sum())
+        else:
+            # Exact count of the implied nearest-neighbour full-resolution mask,
+            # without allocating it. Coarse masks establish obvious content only.
+            wy = np.diff(
+                np.ceil(np.arange(mask.shape[0] + 1) * h / mask.shape[0]).astype(
+                    "int64"
+                )
+            )
+            wx = np.diff(
+                np.ceil(np.arange(mask.shape[1] + 1) * (b - a) / mask.shape[1]).astype(
+                    "int64"
+                )
+            )
+            count = int(np.sum(mask * wy[:, None] * wx[None, :]))
+        o = plan["outputs"][i]
+        o["ink_pixels"] = count
+        o["keep"] = count > 0
+        o["state"] = "content" if count else "blank"
+        o["rect_px"] = (
+            [0, 0, o["rect_px"][2] + x0, h]
+            if i == 0
+            else [o["rect_px"][0] + x0, 0, w, h]
+        )
+    for c in plan["crossings"]:
+        c["box"][0] += x0
+        c["box"][2] += x0
+        for box in c.get("anchor_boxes", []):
+            if box is not None:
+                box[0] += x0
+                box[2] += x0
+    plan["size_px"] = [w, h]
+    plan["analysis"] = dict(
+        mode="fast",
+        gate_dpi=gate_dpi,
+        fold_fraction=fold_fraction,
+        early_written=early,
+        fallback=None,
+    )
+    return plan
+
+
+def analyse_pdf_page_fast(page, dpi, kind, *, fold_fraction=0.1, gate_dpi=75):
+    """Render a small overview and the precise fold directly from the PDF.
+
+    Full-resolution halves are rendered only when the positive overview gate
+    cannot establish writing. Raster clips are aligned to the same integer
+    pixel grid as a complete reference render, including rotated CropBoxes.
+    """
+    gate_dpi = min(gate_dpi, dpi)
+    bounds = (page.rect * pymupdf.Matrix(dpi / 72, dpi / 72)).irect
+    h, w = bounds.height, bounds.width
+    coarse_pix = page.get_pixmap(
+        dpi=round(gate_dpi), colorspace=pymupdf.csRGB, alpha=False
+    )
+    coarse = np.frombuffer(coarse_pix.samples, np.uint8).reshape(
+        coarse_pix.height, coarse_pix.width, 3
+    )
+
+    def render(x0, x1):
+        rect = pymupdf.Rect(x0 * 72 / dpi, 0, x1 * 72 / dpi, page.rect.height)
+        pix = page.get_pixmap(
+            dpi=round(dpi), clip=rect, colorspace=pymupdf.csRGB, alpha=False
+        )
+        data = np.frombuffer(pix.samples, np.uint8).reshape(pix.height, pix.width, 3)
+        out = np.full((h, x1 - x0, 3), 255, np.uint8)
+        left = max(x0, pix.x)
+        top = max(0, pix.y)
+        right = min(x1, pix.x + pix.width)
+        bottom = min(h, pix.y + pix.height)
+        out[top:bottom, left - x0 : right - x0] = data[
+            top - pix.y : bottom - pix.y, left - pix.x : right - pix.x
+        ]
+        return out
+
+    def reference():
+        return _analyse_page_full(render(0, w), dpi, kind)
+
+    if kind not in {"a3", "a4"}:
+        return reference(), coarse
+    if kind == "a4":
+        mask, gate = quick_page_content(
+            coarse,
+            gate_dpi,
+            gate_dpi,
+            native=True,
+            _fallback=lambda: analyse_ink(render(0, w), dpi)[0],
+        )
+        return _a4_plan(mask, gate, h, w, dpi, gate_dpi), coarse
+    middle = w // 2
+    cm = coarse.shape[1] // 2
+    halves = []
+    early = []
+    for a, b, ca, cb in ((0, middle, 0, cm), (middle, w, cm, coarse.shape[1])):
+        mask, gate = quick_page_content(
+            coarse[:, ca:cb],
+            gate_dpi,
+            gate_dpi,
+            native=True,
+            _fallback=lambda a=a, b=b: analyse_ink(render(a, b), dpi)[0],
+        )
+        halves.append(mask)
+        early.append(gate)
+    plan = _assemble_fast(
+        halves, early, h, w, dpi, fold_fraction, gate_dpi, render, reference
+    )
+    plan["analysis"]["rendering"] = "direct-pdf-roi"
+    return plan, coarse

copienator/ink_detection.py

@@ -45,12 +45,15 @@ def _paper_blur(gray: np.ndarray, sigma: float) -> np.ndarray:
     return blurred.astype(np.uint8)
 
 
-def _large_blank_ink(gray: np.ndarray, clean: np.ndarray, dpi: float):
+def _large_blank_ink(gray: np.ndarray, clean: np.ndarray, dpi: float, *,
+                     content_only: bool = False):
     """Refine confirmed ruling, only accepting substantial extra blank margins.
 
     Short directional openings tolerate locally bent/broken paper lines. A
     physical component-size threshold rejects their remaining tiny fragments.
-    This is deliberately limited to already-confirmed ruled paper.
+    Crop refinement is limited to already-confirmed ruled paper. Content-only
+    analysis also uses this stroke filter on seedless plain pages and can return
+    an empty mask; the normal crop path deliberately preserves its input then.
     """
     px = dpi / 25.4
     height, width = gray.shape
@@ -120,6 +123,14 @@ def _large_blank_ink(gray: np.ndarray, clean: np.ndarray, dpi: float):
         overlap = np.bincount(lab[strong_ruling > 0], minlength=count)
         bars &= overlap < st[:, 4]*.5
     refined |= bars[lab]
+    if content_only:
+        # A crop must preserve its input when there is no reliable content.
+        # A blank-page classifier needs that empty result instead. Recover the
+        # same nearby original strokes without restoring remote paper artifacts.
+        radius = max(1, round(2*px))
+        nearby = cv2.dilate(refined.astype(np.uint8),
+            cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(2*radius+1,2*radius+1))) > 0
+        return refined | ((clean > 0) & nearby), len(holes), True
     ys = np.flatnonzero(np.any(refined, axis=1))
     original = np.flatnonzero(np.any(clean, axis=1))
     if not len(ys) or not len(original):
@@ -138,8 +149,171 @@ def _large_blank_ink(gray: np.ndarray, clean: np.ndarray, dpi: float):
     return result, len(holes), changed
 
 
+def _blank_content_support(content: np.ndarray, dpi: float) -> tuple[np.ndarray, dict]:
+    """Require a glyph, neighbouring strokes, or a substantial diagram.
+
+    A few remote dots and a short loose curve do not establish a written page.
+    Grouping protects fragmented pencil letters without a page-wide ink quota.
+    This evidence filter is for blank classification, never margin cropping.
+    """
+    px = dpi/25.4
+    n,labels,stats,_ = cv2.connectedComponentsWithStats(content.astype(np.uint8),8)
+    if n == 1:
+        return content,dict(content_supported_groups=0,content_unsupported_pixels=0)
+    radius = max(1,round(1.5*px))
+    joined = cv2.dilate(content.astype(np.uint8),
+        cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(2*radius+1,2*radius+1)))
+    _,groups = cv2.connectedComponents(joined,8)
+    ys,xs = np.nonzero(content)
+    component_groups = np.zeros(n,np.int32)
+    component_groups[labels[ys,xs]] = groups[ys,xs]
+    area = stats[:,4]
+    bw,bh = stats[:,2],stats[:,3]
+    density = area/np.maximum(1,bw*bh)
+    letter = ((area >= .8*px*px) & (np.minimum(bw,bh) >= .6*px)
+              & (density >= .18))
+    # A dense sub-2 mm ink spot is not a glyph on its own. Keep it eligible
+    # for a supported group (decimal points, dots over letters, punctuation).
+    compact_spot = np.zeros(n,bool)
+    spot_candidates = np.flatnonzero((density > .55) & (np.maximum(bw,bh) < 2*px))
+    for i in spot_candidates:
+        if i == 0:
+            continue
+        x,y,ww,hh,_ = stats[i]
+        yy,xx = np.nonzero(labels[y:y+hh,x:x+ww] == i)
+        hull_area = cv2.contourArea(cv2.convexHull(np.c_[xx,yy].astype(np.int32)))
+        # Small x's have empty wedges inside their convex hull; solid spots
+        # fill it. Pixel-centre hull area can be smaller than the pixel count.
+        compact_spot[i] = area[i] >= .9*max(1.,hull_area)
+    letter &= ~compact_spot
+    substantial = (area >= 4*px*px) & (np.maximum(bw,bh) >= 5*px)
+    # A lone short smear at the physical top/bottom edge can pass both area
+    # tests. Require its pixels to be almost collinear before discounting it;
+    # normal digits, letters and closed diagrams have broader stroke geometry.
+    # Neighbouring meaningful strokes still support it through the group rule.
+    near_edge = (stats[:,1] < 15*px) | (stats[:,1]+bh > content.shape[0]-15*px)
+    edge_smear = np.zeros(n,bool)
+    candidates = np.flatnonzero(near_edge & (area < 6*px*px)
+                               & (np.maximum(bw,bh) < 8*px)
+                               & (np.minimum(bw,bh) > 2*px) & (density < .35))
+    for i in candidates:
+        if i == 0:
+            continue
+        x,y,ww,hh,_ = stats[i]
+        yy,xx = np.nonzero(labels[y:y+hh,x:x+ww] == i)
+        values = np.linalg.eigvalsh(np.cov(np.c_[xx,yy].T))
+        edge_smear[i] = values[0] < .05*values[1]
+    letter &= ~edge_smear
+    substantial &= ~edge_smear
+    credible = area >= .2*px*px
+    credible[0] = False
+    counts = np.bincount(component_groups[credible],minlength=int(groups.max())+1)
+    totals = np.bincount(component_groups,weights=area,minlength=len(counts))
+    supported = (counts >= 2) & (totals >= 1.2*px*px)
+    supported[component_groups[letter | substantial]] = True
+    supported[0] = False
+    retained = supported[groups] & content
+    return retained,dict(content_supported_groups=int(supported.sum()),
+                         content_unsupported_pixels=int(content.sum()-retained.sum()),
+                         content_compact_spots=int(compact_spot[1:].sum()),
+                         content_edge_smears=int(edge_smear.sum()))
+
+
+def detect_page_content(rgb: np.ndarray, dpi: float = 200) -> tuple[np.ndarray, dict]:
+    """Expose rognage's content evidence on a single logical page.
+
+Normal crop behaviour stays unchanged. On confirmed ruled paper, use its
+large-blank refinement even when that refinement finds no content, instead of
+the crop-specific fallback that keeps the original page. No separate weak-mark
+detector is used. The returned mask is in the input image coordinates.
+"""
+    content_masks = {}
+    paper_geometry = {}
+    result = detect_bounds(rgb,dpi,content_masks=content_masks,_paper_geometry=paper_geometry)
+    content = content_masks['foreground']
+    gray = cv2.cvtColor(rgb,cv2.COLOR_RGB2GRAY)
+    # Use the same deskew and paper-family confirmation as the crop pipeline.
+    h,w = gray.shape
+    if paper_geometry:
+        matrix,size,deskewed = (paper_geometry[key] for key in ('matrix','size','deskewed'))
+    else:
+        angle = _skew(gray,angle_step=.5)
+        matrix = cv2.getRotationMatrix2D((w/2,h/2),angle,1)
+        corners = cv2.transform(np.array([[[0,0],[w,0],[0,h],[w,h]]],float),matrix)[0]
+        origin = np.floor(corners.min(axis=0))
+        size = np.ceil(corners.max(axis=0)-origin).astype(int)
+        matrix[:,2] -= origin
+        deskewed = cv2.warpAffine(gray,matrix,tuple(size),borderValue=255)
+    binary = cv2.adaptiveThreshold(deskewed,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
+        cv2.THRESH_BINARY_INV,max(15,int(dpi/5)|1),9)
+    horizontal,nh = _ruling(binary,True)
+    vertical,nv = _ruling(binary,False)
+    confirmed = ((nh and _edge_connected_family(binary,horizontal,True)) or
+                 (nv and _edge_connected_family(binary,vertical,False)))
+    if not confirmed:
+        # Slight local warp breaks one-pixel pale grid lines. Confirm the same
+        # periodic, edge-reaching family with a small physical-width tolerance.
+        tolerance = max(3,round(.3*dpi/25.4)|1)
+        tolerant = cv2.dilate(binary,np.ones((tolerance,tolerance),np.uint8))
+        th,tnh = _ruling(tolerant,True)
+        tv,tnv = _ruling(tolerant,False)
+        confirmed = ((tnh and _edge_connected_family(tolerant,th,True)) or
+                     (tnv and _edge_connected_family(tolerant,tv,False)))
+        result['content_tolerant_ruling'] = bool(confirmed)
+    # With no seeds, the same coherent-stroke filter can recover pencil-only
+    # writing on plain paper. Arbitrary weak pixels alone are not a veto.
+    refine = bool(confirmed) or not content.any()
+    result['content_refinement'] = refine
+    result['content_ruling_confirmed'] = bool(confirmed)
+    if refine:
+        clean = cv2.warpAffine(content.astype(np.uint8),matrix,tuple(size),
+                              flags=cv2.INTER_NEAREST,borderValue=0)
+        refined,holes,_ = _large_blank_ink(deskewed,clean,dpi,content_only=True)
+        content = cv2.warpAffine(refined.astype(np.uint8),cv2.invertAffineTransform(matrix),
+            (w,h),flags=cv2.INTER_NEAREST,borderValue=0)>0
+        result['content_holes'] = holes
+    result['content_pixels'] = int(content.sum())
+    if not confirmed and not content.any():
+        # The directional refinement can erase very thin pencil letters on
+        # plain paper. Reuse the crop pipeline's adaptive foreground instead,
+        # requiring coherent letter-sized components rather than arbitrary
+        # weak pixels. This fallback is confined to seedless/plain pages.
+        weak = _foreground(gray,dpi,9,np.zeros_like(gray))
+        weak,_ = _content_mask(weak,dpi)
+        count,labels,stats,_ = cv2.connectedComponentsWithStats(weak,8)
+        keep = np.zeros(count,bool)
+        px = dpi/25.4
+        keep[1:] = ((stats[1:,4] >= .5*px*px)
+                    & (stats[1:,2] >= px) & (stats[1:,3] >= px))
+        # Faint scanner borders can be one huge connected component. They are
+        # not coherent writing merely because their bounding box is large.
+        x,y,bw,bh,area = stats[1:].T
+        sparse = area / np.maximum(1,bw*bh) < .1
+        edge_frame = (((bw > .9*w) & ((y < px) | (y+bh > h-px)))
+                      | ((bh > .9*h) & ((x < px) | (x+bw > w-px))))
+        edge_frame |= ((bw > .8*w) & ((x < 2*px) | (x+bw > w-2*px)))
+        keep[1:] &= ~(sparse & edge_frame)
+        # A faint narrow scanner-edge band is not a pencil letter. Require
+        # both edge placement and pale neutral appearance, not position alone.
+        for i in np.flatnonzero(keep[1:])+1:
+            xx,yy,ww,hh,_ = stats[i]
+            if ww < 3*px and (xx < 2*px or xx+ww > w-2*px) and hh > ww:
+                values = rgb[yy:yy+hh,xx:xx+ww][labels[yy:yy+hh,xx:xx+ww]==i]
+                if np.median(values.min(axis=1)) > 205 and np.median(np.ptp(values.astype(np.int16),axis=1)) < 30:
+                    keep[i] = False
+        content = keep[labels]
+        result['content_pixels'] = int(content.sum())
+        result['content_pencil_recovery'] = bool(content.any())
+    content,support = _blank_content_support(content,dpi)
+    result.update(support)
+    result['content_pixels'] = int(content.sum())
+    result['content_state'] = 'content' if content.any() else 'blank'
+    return content,result
+
+
 def _neutral_paper_foreground(gray: np.ndarray, chroma: np.ndarray, dpi: float,
-                              *, diagnostics: dict | None = None):
+                              *, diagnostics: dict | None = None,
+                              geometry: dict | None = None):
     """Clean confirmed dark ruling before it can seed whole pages.
 
     Returns None for ordinary ink components or unconfirmed paper geometry.
@@ -163,6 +337,8 @@ def _neutral_paper_foreground(gray: np.ndarray, chroma: np.ndarray, dpi: float,
     size = np.ceil(corners.max(axis=0)-origin).astype(int)
     matrix[:, 2] -= origin
     deskewed = cv2.warpAffine(gray, matrix, tuple(size), borderValue=255)
+    if geometry is not None:
+        geometry.update(matrix=matrix,size=size,deskewed=deskewed)
     background = _paper_blur(deskewed, dpi/8)
     normalized = cv2.divide(deskewed, np.maximum(background,1), scale=255)
     block = max(15, int(dpi/5) | 1)
@@ -220,7 +396,9 @@ def _seed_mask(mask: np.ndarray, px: float, *, paper_nearby=None,
 
 def detect_bounds(rgb: np.ndarray, dpi: float = 200, padding_mm: float = 6,
                   min_crop_mm: float = 5, *, diagnostics: dict | None = None,
-                  colour_cleanup: bool = False, edge_mark_cleanup: bool = True) -> dict:
+                  colour_cleanup: bool = False, edge_mark_cleanup: bool = True,
+                  content_masks: dict | None = None,
+                  _paper_geometry: dict | None = None) -> dict:
     """Locate strong ink, recover adjacent faint strokes, and retain padding.
 
     Neutral punched-hole shadows usually have neither sufficient chroma nor
@@ -237,6 +415,9 @@ def detect_bounds(rgb: np.ndarray, dpi: float = 200, padding_mm: float = 6,
     Optional diagnostics are written to a separate caller-owned dictionary.
     Its masks use the original image coordinates and never influence decisions
     or enter the JSON-serializable return value.
+
+    content_masks requests only foreground and weak-candidate masks, avoiding
+    diagnostic component explanations and intermediate paper-mask exports.
     """
     px = dpi/25.4
     h, w = rgb.shape[:2]
@@ -259,7 +440,8 @@ def detect_bounds(rgb: np.ndarray, dpi: float = 200, padding_mm: float = 6,
         diagnostics['masks'] = {}
         diagnostics['thresholds'] = []
     masks = diagnostics['masks'] if diagnostics is not None else None
-    cleaned, paper_info = _neutral_paper_foreground(gray, chroma, dpi, diagnostics=masks)
+    cleaned, paper_info = _neutral_paper_foreground(gray, chroma, dpi, diagnostics=masks,
+                                                  geometry=_paper_geometry)
     if cleaned is not None:
         # Adaptive neutral cleanup includes faint paper fragments. Such pixels
         # need local contrast or substantial darkness before they seed a crop.
@@ -279,6 +461,9 @@ def detect_bounds(rgb: np.ndarray, dpi: float = 200, padding_mm: float = 6,
         masks['isolated_edge_marks'] = np.zeros((h,w), bool)
     weak = ((chroma > 60) | ((gray < 175) & (residual > 25))).astype(np.uint8)
     weak[suppressed] = 0
+    if content_masks is not None:
+        content_masks.clear()
+        content_masks.update(foreground=np.zeros((h,w),bool),weak_candidates=weak>0)
     if masks is not None:
         masks['strong_seeds'] = np.zeros((h,w), bool)
         masks['foreground'] = np.zeros((h,w), bool)
@@ -329,6 +514,8 @@ def detect_bounds(rgb: np.ndarray, dpi: float = 200, padding_mm: float = 6,
                     masks['isolated_edge_marks'] |= discarded[labels]
                 seeded_labels = remaining
         retained = seeded_labels[labels]
+        if content_masks is not None:
+            content_masks['foreground'] |= retained
         if masks is not None:
             masks['strong_seeds'] |= seeds & retained
         ys = np.flatnonzero(np.any(retained, axis=1))

copienator/paper_ruling.py

@@ -93,7 +93,6 @@ def ruling_suppression(rgb: np.ndarray, dpi: float):
     matrix[:,2] -= origin
     image = cv2.warpAffine(rgb, matrix, tuple(size), borderValue=(255,255,255))
     dark = 255-image
-    dark16 = dark.astype(np.int16)
     darkness = _maximum(dark)
     binary = cv2.adaptiveThreshold(_minimum(image), 255,
         cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV,
@@ -119,7 +118,7 @@ def ruling_suppression(rgb: np.ndarray, dpi: float):
         support = _maximum(model) > 9
         if count:
             support &= bands > 0
-        close = _maximum(np.abs(dark16-model.astype(np.int16))) < 40
+        close = _maximum(cv2.absdiff(dark,model)) < 40
         sample = support & (darkness > 9) & close & ~thick
         # Neutral, red, green and blue line families are modelled separately so
         # a red margin rule cannot be mistaken for unusually dark blue writing.
@@ -128,7 +127,10 @@ def ruling_suppression(rgb: np.ndarray, dpi: float):
         classes[second <= third] = 1
         classes[(first <= second) & (first <= third)] = 0
         classes[cv2.subtract(_maximum(model),_minimum(model)) < 25] = 3
-        credible = np.zeros(binary.shape, bool)
+        # Rejected classes have ceiling -1, below every uint8 darkness value.
+        # Compute each class ceiling once, then test pixels by lookup instead
+        # of subtracting and splitting full RGB images for every colour class.
+        ceilings = np.full((4,3),-1,np.int16)
         for colour in range(4):
             selected = sample & (classes == colour)
             if selected.sum() < 100:
@@ -143,9 +145,10 @@ def ruling_suppression(rgb: np.ndarray, dpi: float):
             # An isolated long fraction bar must not define its own paper colour.
             # The upper quantile admits darker major rules within an established
             # family while rejecting excess ink on most of its pixels.
-            ceiling = np.floor(np.quantile(dark[selected],.99,axis=0)+25).astype(np.int16)
-            credible |= ((classes == colour)
-                         & (_maximum(dark16-ceiling) <= 0))
+            ceilings[colour] = np.floor(np.quantile(dark[selected],.99,axis=0)+25).astype(np.int16)
+        credible = ((dark[:,:,0] <= ceilings[classes,0])
+                    & (dark[:,:,1] <= ceilings[classes,1])
+                    & (dark[:,:,2] <= ceilings[classes,2]))
         match = close & credible & (darkness > 9) & ~thick
         removed |= support & match
         protected |= support & ~match & (binary > 0)