Benjamin_Boenisch/breakpilot-lehrer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: add 3px cell padding + upscale small RapidOCR crops + diagnostic logging

Browse Source 
- Add 3px padding around cell crops to avoid clipping edge characters
  (parentheses in "Tanz(veranstaltung)", descenders, etc.)
- Upscale small BGR crops for RapidOCR, same as Tesseract path
- Add info-level diagnostic logging to _ocr_cell_crop for debugging

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-03-04 16:45:59 +01:00
parent e4bdb3cc24
commit 113a1c10e5
1 changed files with 44 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
klausur-service/backend/cv_vocab_pipeline.py
View File

@@ -4703,11 +4703,16 @@ def _ocr_cell_crop(
disp_w = col.width
disp_h = row.height
# Crop boundaries (clamped to image)
cx = max(0, disp_x)
cy = max(0, disp_y)
cw = min(disp_w, img_w - cx)
ch = min(disp_h, img_h - cy)
# Crop boundaries: add small internal padding (3px each side) to avoid
# clipping characters near column/row edges (e.g. parentheses, descenders).
# Stays within image bounds but may extend slightly beyond strict cell.
_PAD = 3
cx = max(0, disp_x - _PAD)
cy = max(0, disp_y - _PAD)
cx2 = min(img_w, disp_x + disp_w + _PAD)
cy2 = min(img_h, disp_y + disp_h + _PAD)
cw = cx2 - cx
ch = cy2 - cy
empty_cell = {
'cell_id': f"R{row_idx:02d}_C{col_idx}",
@@ -4727,7 +4732,7 @@ def _ocr_cell_crop(
}
if cw <= 0 or ch <= 0:
logger.debug("_ocr_cell_crop R%02d_C%d: zero-size crop (%dx%d)", row_idx, col_idx, cw, ch)
logger.info("_ocr_cell_crop R%02d_C%d: zero-size crop (%dx%d)", row_idx, col_idx, cw, ch)
return empty_cell
# --- Pixel-density check: skip truly empty cells ---
@@ -4736,8 +4741,8 @@ def _ocr_cell_crop(
if crop.size > 0:
dark_ratio = float(np.count_nonzero(crop < 180)) / crop.size
if dark_ratio < 0.005:
logger.debug("_ocr_cell_crop R%02d_C%d: skip empty (dark_ratio=%.4f, crop=%dx%d)",
row_idx, col_idx, dark_ratio, cw, ch)
logger.info("_ocr_cell_crop R%02d_C%d: skip empty (dark_ratio=%.4f, crop=%dx%d)",
row_idx, col_idx, dark_ratio, cw, ch)
return empty_cell
# --- Prepare crop for OCR ---
@@ -4755,8 +4760,31 @@ def _ocr_cell_crop(
cell_region = PageRegion(type=col.type, x=cx, y=cy, width=cw, height=ch)
words = ocr_region_lighton(img_bgr, cell_region)
elif engine_name == "rapid" and img_bgr is not None:
cell_region = PageRegion(type=col.type, x=cx, y=cy, width=cw, height=ch)
words = ocr_region_rapid(img_bgr, cell_region)
# Upscale small BGR crops for RapidOCR (same as Tesseract path)
bgr_crop = img_bgr[cy:cy + ch, cx:cx + cw]
if bgr_crop.size == 0:
words = []
else:
min_dim = 64
scale = 1.0
if ch < min_dim or cw < min_dim:
scale = max(min_dim / max(ch, 1), min_dim / max(cw, 1), 2.0)
bgr_crop = cv2.resize(bgr_crop, None, fx=scale, fy=scale,
interpolation=cv2.INTER_CUBIC)
up_h, up_w = bgr_crop.shape[:2]
tmp_region = PageRegion(type=col.type, x=0, y=0, width=up_w, height=up_h)
words = ocr_region_rapid(bgr_crop, tmp_region)
# Remap positions back to original image coords
if words and scale != 1.0:
for w in words:
w['left'] = int(w['left'] / scale) + cx
w['top'] = int(w['top'] / scale) + cy
w['width'] = int(w['width'] / scale)
w['height'] = int(w['height'] / scale)
elif words:
for w in words:
w['left'] += cx
w['top'] += cy
else:
# Tesseract: upscale tiny crops for better recognition
if ocr_img is not None:
@@ -4790,11 +4818,11 @@ def _ocr_cell_crop(
y_tol = max(15, ch)
text = _words_to_reading_order_text(words, y_tolerance_px=y_tol)
avg_conf = round(sum(w['conf'] for w in words) / len(words), 1)
logger.debug("_ocr_cell_crop R%02d_C%d: OCR raw text=%r conf=%.1f nwords=%d crop=%dx%d psm=%s engine=%s",
row_idx, col_idx, text, avg_conf, len(words), cw, ch, psm, engine_name)
logger.info("_ocr_cell_crop R%02d_C%d: OCR raw text=%r conf=%.1f nwords=%d crop=%dx%d psm=%s engine=%s",
row_idx, col_idx, text, avg_conf, len(words), cw, ch, psm, engine_name)
else:
logger.debug("_ocr_cell_crop R%02d_C%d: OCR returned NO words (crop=%dx%d psm=%s engine=%s)",
row_idx, col_idx, cw, ch, psm, engine_name)
logger.info("_ocr_cell_crop R%02d_C%d: OCR returned NO words (crop=%dx%d psm=%s engine=%s)",
row_idx, col_idx, cw, ch, psm, engine_name)
# --- PSM 7 fallback for still-empty Tesseract cells ---
if not text.strip() and engine_name == "tesseract" and ocr_img is not None:
@@ -4819,8 +4847,8 @@ def _ocr_cell_crop(
pre_filter = text
text = _clean_cell_text_lite(text)
if not text:
logger.debug("_ocr_cell_crop R%02d_C%d: _clean_cell_text_lite REMOVED %r",
row_idx, col_idx, pre_filter)
logger.info("_ocr_cell_crop R%02d_C%d: _clean_cell_text_lite REMOVED %r",
row_idx, col_idx, pre_filter)
avg_conf = 0.0
result = dict(empty_cell)