| """Crop photographs to the zero-shot inscription box (data/ground_4b.jsonl) with a 20% margin and |
| rescale each crop to about 0.64 MP, so that runes occupy more of the recognizer's input. |
| Boxes are re-parsed from the raw model output (the "box" field of runs before the parse_box fix is wrong). |
| Images without a usable box are kept whole. Writes crops/ and splits/ocr_crop_{train,val,test}.jsonl. |
| Runs as the first step of the ocr4b_crop job (exp_v4.sh).""" |
| import json |
| import os |
| import re |
|
|
| from PIL import Image |
|
|
| R = "/home/jovyan/humonen/runic_access" |
| TARGET = 640_000 |
| MARGIN = 0.2 |
|
|
|
|
| def parse_box(text, w, h): |
| m = re.search(r"\[([^\[\]]*)\]", text) |
| nums = re.findall(r"-?\d+(?:\.\d+)?", m.group(1) if m else "") |
| if len(nums) < 4: |
| return None |
| x1, y1, x2, y2 = [float(v) for v in nums[:4]] |
| sx, sy = (w, h) if max(x1, y1, x2, y2) <= 1.0 else (w / 1000.0, h / 1000.0) |
| x1, x2 = sorted((x1 * sx, x2 * sx)) |
| y1, y2 = sorted((y1 * sy, y2 * sy)) |
| x1, y1, x2, y2 = max(0, x1), max(0, y1), min(w, x2), min(h, y2) |
| if x2 - x1 < 0.05 * w or y2 - y1 < 0.05 * h: |
| return None |
| return [x1, y1, x2, y2] |
|
|
|
|
| def crop_one(src, raw_text, dst): |
| """Crop src to the parsed box with a margin, rescale to ~TARGET pixels, save to dst; return the crop box or None.""" |
| im = Image.open(src).convert("RGB") |
| box = parse_box(raw_text, im.width, im.height) |
| if box: |
| x1, y1, x2, y2 = box |
| mx, my = MARGIN * (x2 - x1), MARGIN * (y2 - y1) |
| box = [max(0, x1 - mx), max(0, y1 - my), min(im.width, x2 + mx), min(im.height, y2 + my)] |
| im = im.crop(tuple(round(v) for v in box)) |
| s = (TARGET / (im.width * im.height)) ** 0.5 |
| im = im.resize((max(28, round(im.width * s)), max(28, round(im.height * s))), Image.BICUBIC) |
| im.save(dst + ".tmp.jpg", quality=92) |
| os.replace(dst + ".tmp.jpg", dst) |
| return box |
|
|
|
|
| if __name__ == "__main__": |
| raw = {} |
| for l in open(f"{R}/data/ground_4b.jsonl"): |
| r = json.loads(l) |
| raw[r["image"]] = r["raw"] |
| os.makedirs(f"{R}/crops", exist_ok=True) |
| stats = {"cropped": 0, "whole": 0} |
| boxes_out = open(f"{R}/data/crop_boxes.jsonl", "w") |
| for split in ("train", "val", "test"): |
| with open(f"{R}/splits/ocr_crop_{split}.jsonl", "w") as f: |
| for l in open(f"{R}/splits/ocr_real_{split}.jsonl"): |
| r = json.loads(l) |
| dst_rel = "crops/" + os.path.basename(r["image"]) |
| box = crop_one(f"{R}/{r['image']}", raw.get(r["image"], ""), f"{R}/{dst_rel}") |
| stats["cropped" if box else "whole"] += 1 |
| boxes_out.write(json.dumps({"image": r["image"], "crop_box": box}) + "\n") |
| f.write(json.dumps(dict(r, image=dst_rel, image_full=r["image"]), ensure_ascii=False) + "\n") |
| print(split, stats, flush=True) |
| boxes_out.close() |
| print("DONE", stats) |
|
|
