# 按帧号跳帧采样，裁剪底部 18% 字幕区域
frame_no = 0
while frame_no < total_frames:
    cap.set(cv2.CAP_PROP_POS_FRAMES, frame_no)
    ret, frame = cap.read()
    if not ret:
        break

    # Crop subtitle region (bottom 18% of frame)
    y1 = int(frame_height * subtitle_y_min)   # 0.78
    y2 = int(frame_height * subtitle_y_max)   # 0.96
    x1 = int(frame_width  * subtitle_x_min)   # 0.05
    x2 = int(frame_width  * subtitle_x_max)   # 0.95
    roi = frame[y1:y2, x1:x2]

    results = list(ocr.predict_iter(roi))     # PP-OCRv5
    if results:
        res = results[0]
        texts = [t for t, s in zip(res["rec_texts"], res["rec_scores"])
                 if float(s) > 0.5]           # 置信度过滤
        if texts:
            raw_results.append((frame_no, f"({x1},{y1},{x2},{y2})",
                                " ".join(texts)))
    frame_no += sample_interval               # 每 30 帧采样一次

def _dedup(self, content_list):
    """连续相似帧归组，取组内最长文本（消除单帧误识别）"""
    items = [RawInfo(str(c[0]), c[2]) for c in content_list]
    unique_list = []
    idx_i = 0
    n = len(items)

    while idx_i < n:
        current = items[idx_i]
        start_frame = current.no
        idx_j = idx_i
        # 向后扫描：相似度 >= 0.80 的连续帧归入同一组
        while idx_j < n:
            if idx_j + 1 == n or ratio(
                current.content.replace(" ", ""),
                items[idx_j + 1].content.replace(" ", ""),
            ) < self.threshold_similarity:          # 0.80
                group = items[idx_i : idx_j + 1]
                # 组内取最长文本作为最优识别结果
                stripped = [it.content.replace(" ", "") for it in group]
                best_idx, _ = max(enumerate(stripped), key=lambda x: len(x[1]))
                unique_list.append((start_frame, items[idx_j].no,
                                    group[best_idx].content))
                idx_i = idx_j + 1
                break
            else:
                idx_j += 1
    return unique_list

# Finger curl values: 0.0 = fully straight, 1.0 = fully curled (fist)
HS = {
    # "B" — 平掌，五指并拢伸直
    "flat":      {"thumb": 0.15, "index": 0.0, "middle": 0.0, "ring": 0.0, "pinky": 0.0},
    # "A" — 握拳，拇指贴侧
    "fist":      {"thumb": 0.3,  "index": 1.0, "middle": 1.0, "ring": 1.0, "pinky": 1.0},
    # "V" — 食指中指伸出（胜利/二）
    "victory":   {"thumb": 0.7,  "index": 0.0, "middle": 0.0, "ring": 1.0, "pinky": 1.0},
    # "L" — 拇指+食指伸直成 L 形
    "l_shape":   {"thumb": 0.0,  "index": 0.0, "middle": 1.0, "ring": 1.0, "pinky": 1.0},
    # 点赞 — 拇指竖起
    "thumbs_up": {"thumb": 0.0,  "index": 1.0, "middle": 1.0, "ring": 1.0, "pinky": 1.0},
    # ILY — 拇指+食指+小指伸出（"我爱你"）
    "ily":       {"thumb": 0.0,  "index": 0.0, "middle": 1.0, "ring": 1.0, "pinky": 0.0},
}

# 每词手势还含位置（8 个身体参照位）与移动序列：
# 肩/肘/腕三关节旋转 [rx, ry, rz] + 五指卷曲度 + 序列姿势（接近→主姿势→返回）

def text_to_gesture_sequence(text, fps=10, gesture_duration=0.8):
    # 1. jieba 分词（词库词汇自动注册为自定义词）
    for word in GESTURE_LIBRARY:
        if len(word) >= 2:
            jieba.add_word(word)
    words = list(jieba.cut(text))

    # 2. 逐词查库：精确 → 大小写 → 逐字分解 → 确定性回退
    def gesture_for_token(token):
        if token in GESTURE_LIBRARY:
            return GESTURE_LIBRARY[token]
        chars = list(token)
        if len(chars) > 1:
            return _merge_gestures([_gesture_for_char(c) for c in chars])
        return _gesture_for_char(chars[0])

    def _gesture_for_char(ch):
        # 未收录汉字：以 Unicode 码点为种子生成确定性手势
        seed_val = sum(ord(c) for c in ch)      # 同一文字永远同一手势
        rng = random.Random(seed_val)
        shapes = ["flat","five","point","fist","c_shape","pinch", ...]
        refs = [_FWD_CHEST, _CHEST, _HEAD, _FOREHEAD, _MOUTH, ...]
        return {"left": _rest(),
                "right": {**rng.choice(refs), "fingers": F[rng.choice(shapes)]}}

    # 3. 关键帧生成：保持帧 + 25% 过渡插值
    for i in range(len(gesture_frames) - 1):
        g_from, g_to = gesture_frames[i], gesture_frames[i + 1]
        for f in range(hold_frames):                       # 保持当前手势
            keyframes.append(_frame_dict(current_frame + f, g_from))
        current_frame += hold_frames
        for f in range(transition_frames):                 # 词间过渡
            t = (f + 1) / (transition_frames + 1)
            keyframes.append(_frame_dict(current_frame + f,
                             _interp_gesture(g_from, g_to, t)))  # 逐关节线性插值
        current_frame += transition_frames

POST /api/extract-subtitles   上传视频 → 字幕提取（VSE 引擎，失败自动降级简易 OCR）
POST /api/ocr-image           单帧图片 OCR（字幕区域识别）
POST /api/translate           中译英（googletrans，失败回退戏曲词典）
POST /api/segment-text        jieba 分词 + 词库命中/未命中统计
POST /api/sign-language       文本 → 手语关键帧序列（{frame, left, right}）
POST /api/csl-translate       中文词 → CSL 手语 ID 映射（精确→映射→逐字三级查找）
GET  /api/health              健康检查（含翻译服务可用状态）
GET  /                        前端页面（禁缓存）· /css /js /assets /avatar 静态挂载