Better (?) default left margin

copienator/commands/cutleft.py

@@ -26,6 +26,7 @@ from copienator import (
 from copienator.filesystem import staged_files
 from copienator.copy_errors import clear_copy_error, mark_copy_error, marked_copy_paths
 from copienator.configuration import SHOW_PERSONAL_STEPS
+from copienator.left_margin import analyse_margins
 
 DELIMITER_WIDTH = 5
 DELIMITER_COLOR = (0, 0, 0)
@@ -34,6 +35,7 @@ CROP_SHIFT_STEP = 50
 CROP_WIDTH_STEP = 50
 RENDER_DPI = 200
 NARROW_PAGE_RATIO = 0.8
+AUTO_STRIP_WIDTH_FACTOR = 1.10
 pdf_cache_lock = threading.Lock()
 
 # The requested spelling is kept as the public switch for this alternate layout.
@@ -116,6 +118,9 @@ def process_single_pdf(
     shift_offset: int = 0,
     max_per_file: int = 5,
     width_offset: int = 0,
+    *,
+    automatic_margin: bool = True,
+    margin_analysis: dict | None = None,
 ) -> tuple[Image.Image, list[Image.Image], dict[str, object]] | None:
     """Convert one PDF into a preview, full-resolution splits and metadata."""
     try:
@@ -124,7 +129,10 @@ def process_single_pdf(
         if not images:
             return None
         reference_width = normal_page_width(pdf_path, images)
+        if automatic_margin and max_per_file != 1 and margin_analysis is None:
+            margin_analysis = analyse_margins(images, RENDER_DPI, reference_width)
         full_width_pages = []
+        page_crops = []
         for page_index, image in enumerate(images):
             width, height = image.size
             if max_per_file == 1:
@@ -140,8 +148,17 @@ def process_single_pdf(
                     width,
                     width // 3 + 100 + shift_offset + max(0, width_offset),
                 )
+            if automatic_margin and max_per_file != 1:
+                left = min(width - 1, max(0, margin_analysis['starts_px'][page_index] + shift_offset))
+                strip_width = round(
+                    ((reference_width if width < NARROW_PAGE_RATIO * reference_width else width) // 3)
+                    * AUTO_STRIP_WIDTH_FACTOR
+                )
+                right = min(width, left + strip_width + max(0, width_offset))
             if right > left:
                 cropped_images.append(image.crop((left, 0, right, height)))
+                page_crops.append({'page': page_index + 1, 'left_px': left, 'right_px': right,
+                                   'width_px': width, 'height_px': height})
                 if left == 0 and right == width:
                     full_width_pages.append(page_index + 1)
         if not cropped_images:
@@ -165,6 +182,8 @@ def process_single_pdf(
             "number_of_files": len(split_images),
             "columns_per_file": distribution,
             "full_width_pages": full_width_pages,
+            "page_crops": page_crops,
+            "auto_margin": margin_analysis if automatic_margin and max_per_file != 1 else None,
         }
         return preview, split_images, schema
     except Exception as exc:  # noqa: BLE001 - interactive item failure
@@ -208,6 +227,7 @@ class ImageReviewer:
         files: list[Path],
         output_dir: Path,
         default_max_per_file: int = 5,
+        automatic_margin: bool = True,
     ) -> None:
         self.files = files
         self.output_dir = output_dir
@@ -216,6 +236,8 @@ class ImageReviewer:
         self.had_errors = False
         self.stop_prefetch = threading.Event()
         self.current_result = None
+        self.automatic_margin = automatic_margin
+        self.margin_analysis = None
         self.index = 0
         self.current_shift = 0
         self.current_width_offset = 0
@@ -283,6 +305,7 @@ class ImageReviewer:
         self.current_shift = 0
         self.current_width_offset = 0
         self.current_result = None
+        self.margin_analysis = None
         self.trigger_processing(self.files[self.index], self.current_shift)
 
     def trigger_processing(self, pdf_path: Path, shift: int) -> None:
@@ -299,6 +322,8 @@ class ImageReviewer:
                     shift,
                     self.current_max_per_file,
                     self.current_width_offset,
+                    automatic_margin=self.automatic_margin,
+                    margin_analysis=self.margin_analysis,
                 )
             )
 
@@ -325,6 +350,8 @@ class ImageReviewer:
     ) -> None:
         self.current_preview = result[0]
         self.current_result = result
+        if result[2].get('auto_margin') is not None:
+            self.margin_analysis = result[2]['auto_margin']
         self.update_display(pdf_path.name, result[2])
 
     def update_display(self, filename: str, schema: dict[str, object]) -> None:
@@ -338,6 +365,17 @@ class ImageReviewer:
             if personnal_binding
             else ""
         )
+        analysis = schema.get('auto_margin')
+        margin_info = ''
+        if analysis:
+            offsets = sorted({round(crop['left_px'] * 25.4 / RENDER_DPI, 1)
+                              for crop in schema['page_crops']})
+            needs_review = any(group['status'] in {'inconsistent', 'uncertain', 'insufficient-ink'}
+                               for group in analysis['groups'])
+            margin_info = (f"Auto margin: {offsets} mm from left"
+                           + (' | Check uncertain pages' if needs_review else '') + '\n')
+            if analysis.get('no_margin_from_page') is not None:
+                margin_info += f"No margin from page {analysis['no_margin_from_page']} onwards\n"
         self.label_info.configure(
             text=(
                 f"[{self.index + 1}/{len(self.files)}] {filename} | "
@@ -345,6 +383,7 @@ class ImageReviewer:
                 f"Extra width: {self.current_width_offset}px\n"
                 f"Files: {schema['number_of_files']} | "
                 f"Cols: {schema['columns_per_file']}\n"
+                f"{margin_info}"
                 "Arrows: left/right move by 50px | up/down widen/narrow by 50px\n"
                 "Enter: Save and next | !: flag error and skip | 1: use full pages"
                 f"{personal_instructions}"
@@ -428,6 +467,7 @@ def run(
     fullpage: bool = False,
     no_visual_check: bool = False,
     marked: bool = False,
+    automatic_margin: bool = True,
 ) -> ExitCode:
     files = marked_copy_paths(workspace) if marked else _selected_files(workspace, target)
     if not files:
@@ -463,6 +503,7 @@ def run(
         files,
         workspace.cutleft_dir,
         default_max_per_file=1 if fullpage else 5,
+        automatic_margin=automatic_margin,
     )
     if not reviewer.completed:
         return ExitCode.INTERRUPTED
@@ -472,6 +513,10 @@ def run(
 def build_parser() -> argparse.ArgumentParser:
     parser = target_parser("Interactively crop the label margin from PDF copies")
     parser.add_argument("--marked", action="store_true", help="Review flagged copies and clear only margin-cropping flags after saving with Enter")
+    parser.add_argument(
+        '--manual-margin', action='store_true',
+        help='Use the previous fixed margin instead of automatically estimating body-text positions',
+    )
     parser.add_argument(
         "--fullpage",
         action="store_true",
@@ -496,6 +541,7 @@ def main(argv: Sequence[str] | None = None) -> int:
             fullpage=args.fullpage,
             no_visual_check=args.no_visual_check,
             marked=args.marked,
+            automatic_margin=not args.manual_margin,
         )
 
     return execute(parser, argv, handle)

copienator/left_margin.py

@@ -0,0 +1,151 @@
+"""Conservative body-text margin suggestions for the cutleft reviewer."""
+from __future__ import annotations
+
+import cv2
+import numpy as np
+from PIL import Image
+
+from copienator.ink_detection import detect_bounds
+
+ANALYSIS_DPI = 100
+WIDTH_TOLERANCE = .08
+# Copie01's name/class block reaches about 27% of the full page height.
+# Start below it with clearance, including on later pages with repeated names.
+BODY_TOP = .30
+BODY_BOTTOM = .82
+DISCARD_LEFT_FRACTION = .10
+STANDARD_MARGIN_MM = 40
+MIN_MARGIN_MM = STANDARD_MARGIN_MM / 2
+
+
+def trimmed_left_edge(starts: list[float]) -> float | None:
+    """Discard the leftmost 10% of occupied bands; retain the next minimum."""
+    if not starts:
+        return None
+    ordered = sorted(starts)
+    return round(ordered[int(len(ordered) * DISCARD_LEFT_FRACTION)], 1)
+
+
+def handwriting_mask(rgb: np.ndarray) -> np.ndarray:
+    """Intersect rognage's foreground with a directional ruling filter."""
+    px = ANALYSIS_DPI / 25.4
+    gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY)
+    darkness = 255 - rgb.min(axis=2)
+    length = round(5.2 * px) | 1
+    horizontal = cv2.morphologyEx(darkness, cv2.MORPH_OPEN, np.ones((1, length), np.uint8))
+    vertical = cv2.morphologyEx(darkness, cv2.MORPH_OPEN, np.ones((length, 1), np.uint8))
+    residual = cv2.subtract(darkness, np.maximum(horizontal, vertical))
+    binary = ((residual > 32) & (gray < 200)).astype(np.uint8)
+    count, labels, stats, _ = cv2.connectedComponentsWithStats(binary, 8)
+    keep = np.zeros(count, bool)
+    for index, (_x, _y, width, height, area) in enumerate(stats[1:], 1):
+        keep[index] = (area >= max(5, .35 * px * px) and height >= .5 * px
+                       and width >= .25 * px and height <= 6 * width and width <= 15 * height)
+    diagnostics = {}
+    detect_bounds(rgb, ANALYSIS_DPI, diagnostics=diagnostics)
+    return keep[labels] & diagnostics['masks']['foreground']
+
+
+def body_edge(mask: np.ndarray) -> dict:
+    """Estimate a left edge from occupied 4 mm bands within 30–82% height."""
+    height = mask.shape[0]
+    band_height = max(1, round(4 * ANALYSIS_DPI / 25.4))
+    bottom = round(BODY_BOTTOM * height)
+    starts = []
+    for y in range(round(BODY_TOP * height), bottom, band_height):
+        band = mask[y:min(y + band_height, bottom)]
+        xs = np.flatnonzero(band.sum(axis=0) >= 2)
+        if band.sum() >= 40 and len(xs) >= 12:
+            starts.append(float(xs[0]) * 25.4 / ANALYSIS_DPI)
+    return {
+        'edge_mm': trimmed_left_edge(starts) if len(starts) >= 6 else None,
+        'bands': len(starts),
+        'band_starts_mm': starts,
+    }
+
+
+def estimate_page(image: Image.Image, render_dpi: int) -> dict:
+    scale = ANALYSIS_DPI / render_dpi
+    small = image.convert('RGB').resize(
+        (max(1, round(image.width * scale)), max(1, round(image.height * scale))),
+        Image.Resampling.LANCZOS,
+    )
+    return body_edge(handwriting_mask(np.asarray(small)))
+
+
+def format_groups(images: list[Image.Image], reference_width: float) -> list[list[int]]:
+    """Group similar widths; keep secondary narrow subpages in separate groups."""
+    groups: list[list[int]] = []
+    for index, image in enumerate(images):
+        narrow = image.width < .8 * reference_width
+        for group in groups:
+            width = float(np.median([images[i].width for i in group]))
+            if (images[group[0]].width < .8 * reference_width) == narrow and abs(image.width - width) <= width * WIDTH_TOLERANCE:
+                group.append(index)
+                break
+        else:
+            groups.append([index])
+    return groups
+
+
+def analyse_margins(images: list[Image.Image], render_dpi: int, reference_width: float) -> dict:
+    """Pool occupied bands per format, trim outliers, and crop at half the edge."""
+    starts = [0] * len(images)
+    groups = []
+    page_estimates = {}
+
+    def page_estimate(index):
+        if index not in page_estimates:
+            page_estimates[index] = estimate_page(images[index], render_dpi)
+        return page_estimates[index]
+
+    for indices in format_groups(images, reference_width):
+        # Spread samples through the copy rather than relying on the header page.
+        sample = sorted({indices[0], indices[min(1, len(indices) - 1)],
+                         indices[len(indices) // 2], indices[-1]})
+        estimates = {index: page_estimate(index) for index in sample}
+        values = [item['edge_mm'] for item in estimates.values() if item['edge_mm'] is not None]
+        # Trim across the copy's format group, not separately on each page:
+        # one marginal comment must not veto the other pages' shared margin.
+        band_starts = [start for item in estimates.values() for start in item['band_starts_mm']]
+        status = 'insufficient-ink'
+        edge = trimmed_left_edge(band_starts)
+        left_mm = 0.0
+        if len(values) >= min(2, len(indices)):
+            status = 'near-edge' if edge < MIN_MARGIN_MM else 'margin'
+            left_mm = 0.0 if edge < MIN_MARGIN_MM else edge / 2
+        # Secondary subpages can start in the middle of an answer. Preserve
+        # their left edge even when the only visible writing is indented.
+        if images[indices[0]].width < .8 * reference_width:
+            status = 'narrow-page'
+            left_mm = 0.0
+        start = round(left_mm * render_dpi / 25.4)
+        for index in indices:
+            starts[index] = start
+        groups.append({
+            'pages': [index + 1 for index in indices],
+            'status': status,
+            'edge_mm': edge,
+            'left_mm': round(left_mm, 1),
+            'bands': len(band_starts),
+            'discarded_bands': int(len(band_starts) * DISCARD_LEFT_FRACTION),
+            'samples': [{'page': index + 1, **item} for index, item in estimates.items()],
+        })
+    # Inspect actual neighbours, including pages outside the format samples.
+    # Blank/uncertain pages do not establish or extend a no-margin run.
+    no_margin_run = 0
+    no_margin_from_page = None
+    for index in range(len(images)):
+        edge = page_estimate(index)['edge_mm']
+        no_margin_run = no_margin_run + 1 if edge is not None and edge < MIN_MARGIN_MM else 0
+        if no_margin_run >= 2:
+            run_start = index - 1
+            no_margin_from_page = run_start + 1
+            starts[run_start:] = [0] * (len(images) - run_start)
+            break
+
+    return {'version': 5, 'starts_px': starts, 'groups': groups,
+            'no_margin_from_page': no_margin_from_page,
+            'discard_left_fraction': DISCARD_LEFT_FRACTION, 'cut_fraction_of_margin': .5,
+            'standard_margin_mm': STANDARD_MARGIN_MM, 'minimum_margin_mm': MIN_MARGIN_MM,
+            'body_height_fraction': [BODY_TOP, BODY_BOTTOM]}