import cv2 import numpy as np from PIL import Image, ImageDraw from pathlib import Path from multiprocessing import Pool, cpu_count import time INPUT_DIR = Path(r"C:\allgifted\mathapi11v2\docs\past year questions\Old Images") OUTPUT_DIR = Path(r"C:\allgifted\mathapi11v2\docs\past year questions\Only Images") BORDER = 3 PAD = 36 # was 22 -- recover clipped axis labels on charts D0ACAC = (208, 172, 172) MIN_AREA = 600 # filters body text; keeps illustration shapes HEADER_SKIP = 880 # only applied to tall full-page scans (see height gate) HEADER_H_GATE = 1500 # below this height, skip no header (small per-question scans) DILATE_K = 22 # was 25 -- slightly tighter grouping, still holds charts together MERGE_OVERLAP = 0.35 # min horizontal overlap to merge vertically-adjacent regions MERGE_GAP = 12 # max vertical gap to consider merging INNER_FRAC = 0.20 # margin trimmed from each side for density test (inner 60%) MIN_DENSITY = 0.008 # drop crops whose inner region is <0.8% ink (hollow boxes) def extract(image_path): pil = Image.open(image_path).convert("RGB") img = np.array(pil) h, w = img.shape[:2] gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY) _, binary = cv2.threshold(gray, 210, 255, cv2.THRESH_BINARY_INV) # Height-aware school-header blanking: only on tall full-page scans. skip = HEADER_SKIP if h > HEADER_H_GATE else 0 if skip: binary[:skip, :] = 0 # Blank yellow section-header bands R, G, B = img[:,:,0], img[:,:,1], img[:,:,2] yellow = (R.astype(int)>200) & (G.astype(int)>200) & (B.astype(int)<180) for y in np.where(yellow.mean(axis=1) > 0.2)[0]: binary[max(0,y-2):min(h,y+28), :] = 0 # Keep large components; strip body text (small) and full-width thin rules # (answer blanks / separators). Rule strip is chart-safe: only near-full-width # hairlines are removed, not in-plot chart gridlines. n_lbl, labels, stats, _ = cv2.connectedComponentsWithStats(binary) illus = np.zeros_like(binary) rule_w = int(0.70 * w) for i in range(1, n_lbl): area = stats[i, cv2.CC_STAT_AREA] if area < MIN_AREA: continue cw = stats[i, cv2.CC_STAT_WIDTH] ch = stats[i, cv2.CC_STAT_HEIGHT] if ch <= 10 and cw >= rule_w: # full-width thin rule / answer blank continue illus[labels == i] = 255 kernel = np.ones((DILATE_K, DILATE_K), np.uint8) dilated = cv2.dilate(illus, kernel) contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) regions = [] for c in contours: x, y, bw, bh = cv2.boundingRect(c) if bh < 40 or bw < 80 or bw * bh < 8000: continue if bh < 45 and bw > 400: # thin horizontal separator band continue roi = illus[y:y+bh, x:x+bw] ys, xs = np.where(roi > 0) if len(xs) == 0: continue x1 = max(0, x + int(xs.min()) - PAD) y1 = max(0, y + int(ys.min()) - PAD) x2 = min(w, x + int(xs.max()) + PAD) y2 = min(h, y + int(ys.max()) + PAD) regions.append([x1, y1, x2, y2]) regions.sort(key=lambda b: b[1]) # Overlap-aware vertical merge: only merge stacked regions that actually # share horizontal extent (a split figure), NOT a figure + an offset # answer/working block sitting below it. def hoverlap(a, b): inter = max(0, min(a[2], b[2]) - max(a[0], b[0])) minw = max(1, min(a[2]-a[0], b[2]-b[0])) return inter / minw merged = [] for b in regions: if (merged and b[1] <= merged[-1][3] + MERGE_GAP and hoverlap(merged[-1], b) >= MERGE_OVERLAP): merged[-1][0] = min(merged[-1][0], b[0]) merged[-1][1] = min(merged[-1][1], b[1]) merged[-1][2] = max(merged[-1][2], b[2]) merged[-1][3] = max(merged[-1][3], b[3]) else: merged.append(b[:]) # Re-tighten + density filter (drop hollow boxes / near-empty frames) final = [] for b in merged: roi = illus[b[1]:b[3], b[0]:b[2]] ys, xs = np.where(roi > 0) if len(xs) == 0: continue x1 = max(0, b[0] + int(xs.min()) - PAD) y1 = max(0, b[1] + int(ys.min()) - PAD) x2 = min(w, b[0] + int(xs.max()) + PAD) y2 = min(h, b[1] + int(ys.max()) + PAD) if (y2-y1) < 40 or (x2-x1) < 60: continue # interior ink-density test on inner 60% mx = int((x2-x1) * INNER_FRAC); my = int((y2-y1) * INNER_FRAC) inner = binary[y1+my:y2-my, x1+mx:x2-mx] if inner.size == 0 or (inner > 0).mean() < MIN_DENSITY: continue final.append((x1, y1, x2, y2)) return pil, final def process_one(img_path_str): cv2.setNumThreads(1) img_path = Path(img_path_str) try: pil, regions = extract(img_path) except Exception as e: return (img_path.name, -1, 0, str(e)) stem = img_path.stem saved = 0 for i, (x1, y1, x2, y2) in enumerate(regions, 1): crop = pil.crop((x1, y1, x2, y2)) bw, bh = crop.size bordered = Image.new("RGB", (bw + BORDER*2, bh + BORDER*2), (255,255,255)) draw = ImageDraw.Draw(bordered) draw.rectangle([0, 0, bw+BORDER*2-1, bh+BORDER*2-1], outline=D0ACAC, width=BORDER) bordered.paste(crop, (BORDER, BORDER)) bordered.save(OUTPUT_DIR / f"{stem}_illus_{i:02d}.png") saved += 1 return (img_path.name, len(regions), saved, None) if __name__ == "__main__": OUTPUT_DIR.mkdir(parents=True, exist_ok=True) exts = {".png", ".jpg", ".jpeg", ".bmp"} files = sorted(str(f) for f in INPUT_DIR.iterdir() if f.suffix.lower() in exts) workers = max(1, min(cpu_count() - 2, len(files))) print(f"Found {len(files)} paper images. Using {workers} parallel workers.", flush=True) t0 = time.time() results = [] with Pool(workers) as pool: for res in pool.imap_unordered(process_one, files): results.append(res) if res[3]: print(f" ERROR {res[0]}: {res[3]}", flush=True) elapsed = time.time() - t0 results.sort(key=lambda r: r[0]) log_lines, total_saved, errors, zero = [], 0, 0, 0 for name, nreg, saved, err in results: if err: log_lines.append(f"{name} | ERROR | {err}"); errors += 1 else: log_lines.append(f"{name} | {nreg} illustrations | {saved} saved") total_saved += saved if saved == 0: zero += 1 (OUTPUT_DIR / "_extraction_log.txt").write_text("\n".join(log_lines), encoding="utf-8") print(f"\nDONE in {elapsed:.1f}s", flush=True) print(f" images processed : {len(results)}", flush=True) print(f" crops saved : {total_saved}", flush=True) print(f" images w/ 0 crops: {zero}", flush=True) print(f" errors : {errors}", flush=True)