Better (?) default left margin

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]}