首次提交

tools/analyze_ocr_result.py

@@ -0,0 +1,100 @@
+# -*- coding: utf-8 -*-
+"""
+分析一个 ocr_result.jsonl，判断"冗余文字过多"是真冗余还是误杀。
+复现过滤逻辑：看文字的垂直位置分布(distinct_bands)、每帧文字框数、有没有稳定字幕行。
+
+用法:  python temp/analyze_ocr_result.py "ocr_result (1).jsonl"
+"""
+import json
+import sys
+from collections import defaultdict
+
+PATH = sys.argv[1] if len(sys.argv) > 1 else "ocr_result.jsonl"
+
+
+def parse(path):
+    """解析 'pic_path: X' + json 交替格式，返回 [(frame, detections), ...]。"""
+    pairs = []
+    cur_frame = None
+    with open(path, encoding="utf-8") as f:
+        for line in f:
+            line = line.rstrip("\n")
+            if line.startswith("pic_path:"):
+                cur_frame = line.split("pic_path:", 1)[1].strip()
+            elif line.strip():
+                try:
+                    data = json.loads(line)
+                    pairs.append((cur_frame, data.get("TextDetections", [])))
+                except json.JSONDecodeError:
+                    pass
+    return pairs
+
+
+def center_y(det):
+    poly = det.get("Polygon", [])
+    ys = [p.get("Y", 0) for p in poly]
+    return sum(ys) / len(ys) if ys else 0
+
+
+def main():
+    try:
+        pairs = parse(PATH)
+    except FileNotFoundError:
+        print(f"找不到文件: {PATH}")
+        return
+    if not pairs:
+        print("没解析到数据(格式不对?)")
+        return
+
+    n_frames = len(pairs)
+    boxes_per_frame = [len(d) for _, d in pairs]
+    total_boxes = sum(boxes_per_frame)
+    frames_with_text = sum(1 for b in boxes_per_frame if b)
+
+    # 所有文字框的中心 Y
+    all_cy = []
+    for _, dets in pairs:
+        for det in dets:
+            all_cy.append(center_y(det))
+
+    print(f"文件: {PATH}")
+    print("=" * 60)
+    print(f"帧/图数:        {n_frames}")
+    print(f"有文字的:       {frames_with_text}")
+    print(f"文字框总数:     {total_boxes}")
+    print(f"平均每帧框数:   {total_boxes / n_frames:.1f}")
+    print(f"最多一帧框数:   {max(boxes_per_frame) if boxes_per_frame else 0}")
+
+    # distinct vertical bands(复现 merge.py 的冗余判定：中心Y间隔>80算一个新垂直簇)
+    cys = sorted(all_cy)
+    bands = 1 if cys else 0
+    for i in range(1, len(cys)):
+        if cys[i] - cys[i - 1] > 80:
+            bands += 1
+    print(f"不同垂直位置簇: {bands}   (>=8 会被判冗余)")
+
+    # Y 分布直方图(每 100px 一档)，看文字集中在哪还是散布全屏
+    hist = defaultdict(int)
+    for y in all_cy:
+        hist[int(y // 100) * 100] += 1
+    print("\n文字垂直分布(Y每100px一档，看文字集中度):")
+    if hist:
+        mx = max(hist.values())
+        for y in sorted(hist):
+            bar = "#" * int(hist[y] / mx * 40)
+            print(f"  y={y:>5}-{y+99:<5} {hist[y]:>6}  {bar}")
+
+    # 判定
+    print("\n" + "=" * 60)
+    if bands >= 8:
+        print(f"❌ 真冗余：文字散布在 {bands} 个不同垂直位置(满屏文字/弹幕/文档类)。")
+    elif total_boxes / n_frames > 10:
+        print(f"⚠️ 偏冗余：平均每帧 {total_boxes/n_frames:.0f} 个文字框，画面文字很多。")
+    else:
+        topy = max(hist, key=hist.get) if hist else 0
+        print(f"✅ 可能误杀：文字主要集中在 y={topy}~{topy+99}，像是有固定字幕行的。")
+        print("   若确实有正常字幕，可放宽 merge.py 的 distinct_bands 阈值。")
+
+
+if __name__ == "__main__":
+    main()