Fix watermark detection + S3 upload + NVENC color + task download on S3 failure

- smart_detect.py: rewrite detect_static_regions using saturation-signature method
  (high-sat logo frames vs no-logo frames diff) to handle animated/periodic watermarks
  that break temporal-std approach; merge top-region boxes; add print debugging
- web_app.py: fix S3 upload (async chunked streaming, error visible in task log);
  s3_failed flag keeps task downloadable even when S3 upload fails; add s3_failed
  column to DB; download/stream endpoints allow s3_failed tasks
- web/index.html: show play/download buttons for s3_failed tasks; task sort by
  created_at desc; watermark mode label update; isLocalDone() helper
- video_io.py: add -vf format=yuv420p before h264_nvenc to fix green video

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

.env.example

@@ -0,0 +1,13 @@
+# VSR Web 配置示例 — 复制为 .env 后修改
+# 参考 df-autotest-platform 的 .env 结构
+
+# 管理员密码（留空则不需要登录）
+ADMIN_PASSWORD=
+
+# 存储目录（默认 ./web_cache）
+# VSR_STORAGE=./web_cache
+
+# 服务监听端口（默认 8001）
+# PORT=8001
+
+# Platform Worker 连接信息在「系统设置」页面配置，无需在此设置

.gitignore

@@ -375,3 +375,5 @@ test*_no_sub*.mp4
 /backend/tools/train/release_model/
 inference.onnx
 /config/config.json
+web_cache/
+.env

README.md

@@ -3,6 +3,9 @@
 <div align="center">
   <img src="design/icon_1024.PNG" alt="VSR Logo" width="128" height="128">
 </div>
+<div align="center">
+<a href="https://trendshift.io/repositories/9120" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9120" alt="YaoFANGUK%2Fvideo-subtitle-remover | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</div>
 
 ## 项目简介
 
@@ -15,6 +18,7 @@ Video-subtitle-remover (VSR) 是一款基于AI技术，将视频中的硬字幕
 主要实现了以下功能：
 - **无损分辨率**将视频中的硬字幕去除，生成去除字幕后的文件
 - 通过超强AI算法模型，对去除字幕文本的区域进行填充（非相邻像素填充与马赛克去除）
+- 提取原视频字幕，可配合：[video-subtitle-extractor (VSE)](https://github.com/YaoFANGUK/video-subtitle-extractor)
 - 支持自定义字幕位置，仅去除定义位置中的字幕（传入位置）
 - 支持全视频自动去除所有文本（不传入位置）
 - 支持多选图片批量去除水印文本
@@ -260,6 +264,10 @@ LAMA_SUPER_FAST = False  # 保证效果
 
 解决方案：升级7-zip解压程序到最新版本
 
+4. Mac版本运行报错：Error "bad CPU type in executable"
+
+解决方案：打开控制台输入`softwareupdate --install-rosetta` 安装rosetta
+
 
 ## 赞助
 

README_en.md

@@ -3,6 +3,9 @@
 <div align="center">
   <img src="design/icon_1024.PNG" alt="VSR Logo" width="128" height="128">
 </div>
+<div align="center">
+<a href="https://trendshift.io/repositories/9120" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9120" alt="YaoFANGUK%2Fvideo-subtitle-remover | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</div>
 
 ## Project Introduction
 

backend/tools/smart_detect.py

@@ -0,0 +1,507 @@
+"""
+智能水印检测：通过帧差分找静态区域，用SAM2精确分割。
+"""
+import sys
+import os
+import numpy as np
+import cv2
+
+SAM2_PATH = "/opt/sam2"
+SAM2_CKPT = "/opt/sam2/checkpoints/sam2.1_hiera_tiny.pt"
+SAM2_CFG  = "configs/sam2.1/sam2.1_hiera_t.yaml"
+
+
+def detect_static_regions(video_path: str, sample_frames: int = 30, threshold: float = 20.0,
+                          max_scan_seconds: float = None):
+    """
+    饱和度签名 + 跨帧对比法检测水印/Logo。
+    采样策略：全程每秒一帧，分析时缩放到 1/4 分辨率提速，坐标还原回原始尺寸。
+    """
+    cap = cv2.VideoCapture(video_path)
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
+    orig_frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    orig_frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    duration_s = total / fps
+
+    # 全程每秒采一帧
+    step = max(1, int(fps))
+    positions = list(range(0, total, step))
+
+    # 分析用 1/4 分辨率加速
+    SW, SH = max(1, orig_frame_w // 4), max(1, orig_frame_h // 4)
+
+    bgr_frames = []    # 小尺寸帧，用于分析
+    gray_frames = []
+    actual_pos_s = []  # 每帧对应的时间戳（秒）
+
+    for pos in positions:
+        cap.set(cv2.CAP_PROP_POS_FRAMES, pos)
+        ret, frame = cap.read()
+        if not ret:
+            continue
+        small = cv2.resize(frame, (SW, SH), interpolation=cv2.INTER_AREA)
+        bgr_frames.append(small)
+        gray_frames.append(cv2.cvtColor(small, cv2.COLOR_BGR2GRAY).astype(np.float32))
+        actual_pos_s.append(pos / fps)
+
+    cap.release()
+
+    print(f"[SAM2-detect] 采样 {len(bgr_frames)} 帧（全程每秒一帧，共{duration_s:.1f}s，分析尺寸{SW}x{SH}）", flush=True)
+
+    if len(bgr_frames) < 4:
+        print(f"[SAM2-detect] 帧数不足4帧，返回空", flush=True)
+        return []
+
+    # ── 阶段1：高饱和度识别含logo帧 ─────────────────────────────────────
+    top_h = SH // 5
+    logo_sat_thresh = 200
+    logo_min_pixels = top_h * SW * 0.003
+
+    is_logo = []
+    for i, bgr in enumerate(bgr_frames):
+        hsv = cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)
+        high_sat = int(np.sum(hsv[:top_h, :, 1] > logo_sat_thresh))
+        has_logo = high_sat >= logo_min_pixels
+        is_logo.append(has_logo)
+        mark = '✓' if has_logo else '✗'
+        print(f"[SAM2-detect]   {mark} {actual_pos_s[i]:.1f}s: 高饱和={high_sat}px", flush=True)
+
+    logo_bgr_frames  = [bgr_frames[i] for i, v in enumerate(is_logo) if v]
+    nlogo_bgr_frames = [bgr_frames[i] for i, v in enumerate(is_logo) if not v]
+    logo_gray_frames = [gray_frames[i] for i, v in enumerate(is_logo) if v]
+
+    print(f"[SAM2-detect] logo含帧: {len(logo_bgr_frames)}/{len(bgr_frames)}", flush=True)
+
+    # 分析时使用小尺寸坐标空间
+    w, h = SW, SH
+
+    # ── 路径A：彩色水印（高饱和度）检测 ──────────────────────────────────
+    sat_mask = np.zeros((h, w), dtype=np.uint8)
+    if len(logo_gray_frames) >= 2 and nlogo_bgr_frames:
+        sat_acc = np.zeros((h, w), dtype=np.float32)
+        for bgr in logo_bgr_frames:
+            sat_acc += cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)[:, :, 1].astype(np.float32) / 255.0
+        sat_mean_logo = sat_acc / len(logo_bgr_frames)
+
+        sat_acc_nl = np.zeros((h, w), dtype=np.float32)
+        for bgr in nlogo_bgr_frames[:6]:
+            sat_acc_nl += cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)[:, :, 1].astype(np.float32) / 255.0
+        sat_mean_nlogo = sat_acc_nl / min(len(nlogo_bgr_frames), 6)
+        sat_diff = sat_mean_logo - sat_mean_nlogo
+
+        sat_thresh = min(0.6, np.percentile(sat_mean_logo, 92))
+        diff_thresh = max(0.25, sat_diff.mean() + sat_diff.std())
+        sat_mask = ((sat_mean_logo >= sat_thresh) & (sat_diff >= diff_thresh)).astype(np.uint8) * 255
+        print(f"[SAM2-detect] 彩色路径: sat_thresh={sat_thresh:.3f} diff_thresh={diff_thresh:.3f} 候选={np.sum(sat_mask>0)}", flush=True)
+
+    # ── 路径B：白色/灰色水印（亮度差分）检测 ────────────────────────────────
+    # "sjhs.pw" 类白色文字水印：饱和度≈0，但亮度比背景高，且只出现在少数帧
+    # 策略：把前 N 秒的帧与后续帧做亮度差分，找只出现在片头的高亮区域
+    bright_mask = np.zeros((h, w), dtype=np.uint8)
+    early_cutoff = min(len(gray_frames), max(5, int(10 * fps / step)))  # 前10秒
+    if early_cutoff >= 2 and len(gray_frames) > early_cutoff:
+        early_stack = np.stack(gray_frames[:early_cutoff], axis=0)
+        rest_stack  = np.stack(gray_frames[early_cutoff:], axis=0)
+        early_mean = early_stack.mean(axis=0)
+        rest_mean  = rest_stack.mean(axis=0)
+        # 片头比后续亮得多 → 水印像素（值越大越亮）
+        bright_diff = early_mean - rest_mean
+        # 阈值：差值超过 20 灰度级的区域
+        bright_mask_raw = (bright_diff > 20).astype(np.uint8) * 255
+        # 只保留四角区域（水印通常在角落），避免误报视频内容
+        corner_mask = np.zeros((h, w), dtype=np.uint8)
+        ch, cw = h // 4, w // 4   # 角落占画面 25%
+        corner_mask[:ch, :cw] = 1          # 左上
+        corner_mask[:ch, -cw:] = 1         # 右上
+        corner_mask[-ch:, :cw] = 1         # 左下
+        corner_mask[-ch:, -cw:] = 1        # 右下
+        bright_mask = cv2.bitwise_and(bright_mask_raw, corner_mask * 255)
+        print(f"[SAM2-detect] 亮度差分路径: bright_diff_max={bright_diff.max():.1f} 角落候选={np.sum(bright_mask>0)}", flush=True)
+
+    # ── 合并两条路径 ──────────────────────────────────────────────────────
+    static_mask = cv2.bitwise_or(sat_mask, bright_mask)
+    if np.sum(static_mask) == 0:
+        # 两条路径都没找到，回退到全局时序std法
+        print(f"[SAM2-detect] 两条路径均无结果，回退到时序std法", flush=True)
+        stack = np.stack(gray_frames, axis=0)
+        std_map = np.std(stack, axis=0)
+        static_mask = (std_map < 8.0).astype(np.uint8) * 255
+    print(f"[SAM2-detect] 合并后候选像素数={np.sum(static_mask>0)}", flush=True)
+
+    # 仅排除纯黑区域
+    mean_gray = np.mean(np.stack(gray_frames, axis=0), axis=0)
+    content_mask = (mean_gray > 8).astype(np.uint8) * 255
+    static_mask = cv2.bitwise_and(static_mask, content_mask)
+    print(f"[SAM2-detect] 过滤纯黑后候选像素数={np.sum(static_mask > 0)}", flush=True)
+
+    # 形态学处理：去噪 + 膨胀
+    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
+    static_mask = cv2.morphologyEx(static_mask, cv2.MORPH_OPEN, kernel)
+    static_mask = cv2.dilate(static_mask, kernel, iterations=2)
+
+    # 找轮廓
+    contours, _ = cv2.findContours(static_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+
+    regions = []
+    min_area = (w * h) * 0.0005  # 至少占画面 0.05%
+    max_area = (w * h) * 0.25    # 最多占画面 25%
+
+    print(f"[SAM2-detect] 找到 {len(contours)} 个轮廓，min_area={min_area:.0f} max_area={max_area:.0f}", flush=True)
+    for cnt in contours:
+        area = cv2.contourArea(cnt)
+        x, y, bw, bh = cv2.boundingRect(cnt)
+        ratio = bw / bh if bh > 0 else 0
+        if min_area < area < max_area:
+            if 0.1 < ratio < 20:
+                regions.append((x, y, x + bw, y + bh))
+                print(f"[SAM2-detect]   ✓ 区域 ({x},{y})-({x+bw},{y+bh}) 面积={area:.0f} 比例={ratio:.2f}", flush=True)
+            else:
+                print(f"[SAM2-detect]   ✗ 比例过滤: ({x},{y})-({x+bw},{y+bh}) 面积={area:.0f} 比例={ratio:.2f}", flush=True)
+        else:
+            print(f"[SAM2-detect]   ✗ 面积过滤: ({x},{y})-({x+bw},{y+bh}) 面积={area:.0f}", flush=True)
+
+    # 若检测到多个位于帧顶部的区域，合并成一个大框覆盖整个logo
+    # logo通常有高饱和的亮色区域（检测到）和低饱和的暗色区域（未检测到）
+    # 合并+适当扩展可确保完整覆盖
+    if len(regions) >= 2:
+        top_regions = [r for r in regions if r[1] < h * 0.3]  # 只合并帧顶30%内的区域
+        if len(top_regions) >= 2:
+            x1 = max(0, min(r[0] for r in top_regions) - 30)
+            y1 = max(0, min(r[1] for r in top_regions) - 15)
+            x2 = min(w, max(r[2] for r in top_regions) + 30)
+            y2 = min(h, max(r[3] for r in top_regions) + 20)
+            other = [r for r in regions if r[1] >= h * 0.3]
+            merged = [(x1, y1, x2, y2)] + other
+            print(f"[SAM2-detect] 合并顶部 {len(top_regions)} 个区域为 ({x1},{y1})-({x2},{y2})", flush=True)
+            regions = merged
+
+    # ── 坐标从分析尺寸（SW×SH）还原到原始分辨率 ──────────────────────────
+    scale_x = orig_frame_w / SW
+    scale_y = orig_frame_h / SH
+    regions = [(int(x1*scale_x), int(y1*scale_y), int(x2*scale_x), int(y2*scale_y))
+               for x1, y1, x2, y2 in regions]
+
+    print(f"[SAM2-detect] detect_static_regions 返回 {len(regions)} 个区域（已还原到{orig_frame_w}x{orig_frame_h}）", flush=True)
+    return regions
+
+
+def refine_with_sam2(image: np.ndarray, boxes: list, gpu_id: int = 0):
+    """
+    用 SAM2 对检测到的区域精确分割，返回合并后的 mask。
+    image: BGR numpy array
+    boxes: [(x1,y1,x2,y2), ...]
+    返回 binary mask (H, W) uint8
+    """
+    if not boxes:
+        return None
+
+    try:
+        if SAM2_PATH not in sys.path:
+            sys.path.insert(0, SAM2_PATH)
+
+        import torch
+        from sam2.build_sam import build_sam2
+        from sam2.sam2_image_predictor import SAM2ImagePredictor
+
+        device = f"cuda:{gpu_id}" if torch.cuda.is_available() else "cpu"
+        model = build_sam2(SAM2_CFG, SAM2_CKPT, device=device)
+        predictor = SAM2ImagePredictor(model)
+
+        rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        predictor.set_image(rgb)
+
+        h, w = image.shape[:2]
+        combined_mask = np.zeros((h, w), dtype=np.uint8)
+
+        for (x1, y1, x2, y2) in boxes:
+            box = np.array([x1, y1, x2, y2])
+            masks, scores, _ = predictor.predict(
+                box=box[None],
+                multimask_output=False,
+            )
+            if masks is not None and len(masks) > 0:
+                m = masks[0].astype(np.uint8) * 255
+                combined_mask = cv2.bitwise_or(combined_mask, m)
+
+        return combined_mask
+
+    except Exception as e:
+        print(f"[SAM2] 分割失败: {e}", flush=True)
+        return None
+
+
+def _frame_text_candidates(gray, tophat_kernel, conn_kernel, th_val=22):
+    """单帧：white top-hat 提取细笔画亮文字 → 候选文字框 (xmin,xmax,ymin,ymax)。
+    top-hat 只响应比周围亮、且比核(15px)细的结构 → 文字笔画命中，大块皮肤/丝袜不命中。
+    再用「连通组件数」区分文字(多个字母,>=3)和实心亮物(卡扣/内衣,1-2块)。"""
+    tophat = cv2.morphologyEx(gray, cv2.MORPH_TOPHAT, tophat_kernel)
+    th = cv2.threshold(tophat, th_val, 255, cv2.THRESH_BINARY)[1]
+    closed = cv2.morphologyEx(th, cv2.MORPH_CLOSE, conn_kernel, iterations=2)
+    cnts, _ = cv2.findContours(closed, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    boxes = []
+    for c in cnts:
+        x, y, bw, bh = cv2.boundingRect(c)
+        area = bw * bh
+        ar = bw / max(bh, 1)
+        # 文字特征：宽>高、有一定宽度、不至于太小
+        if not (area > 350 and 1.3 < ar < 15 and bw > 35 and bh > 6):
+            continue
+        # 框内细笔画（未闭合的原始 top-hat）的独立组件数：文字=多字母(>=3)，实心物=1-2
+        roi = th[y:y+bh, x:x+bw]
+        n, _, stats, _ = cv2.connectedComponentsWithStats(roi, connectivity=8)
+        comp = sum(1 for i in range(1, n) if stats[i, cv2.CC_STAT_AREA] >= 8)
+        if comp < 3:
+            continue
+        boxes.append((x, x + bw, y, y + bh))   # (xmin,xmax,ymin,ymax)
+    return boxes
+
+
+def detect_color_logo(video_path: str, sample_n: int = 40):
+    """检测彩色不透明 logo 水印（如「黑料网」橙色横幅）。
+    判据：高饱和(>150) AND 时序低方差(静止) → 排除会动的皮肤、柔和的背景画。
+    返回 [(xmin,xmax,ymin,ymax), ...] 静止框（应用到所有帧）。"""
+    cap = cv2.VideoCapture(video_path)
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    positions = np.linspace(0, total - 1, min(sample_n, total)).astype(int)
+    gray_stack, sat_acc, cnt = [], np.zeros((h, w), np.float32), 0
+    for p in positions:
+        cap.set(cv2.CAP_PROP_POS_FRAMES, int(p))
+        ret, f = cap.read()
+        if not ret:
+            continue
+        gray_stack.append(cv2.cvtColor(f, cv2.COLOR_BGR2GRAY).astype(np.float32))
+        sat_acc += cv2.cvtColor(f, cv2.COLOR_BGR2HSV)[:, :, 1].astype(np.float32)
+        cnt += 1
+    cap.release()
+    if cnt < 4:
+        return []
+    std = np.std(np.stack(gray_stack), axis=0)
+    sat = sat_acc / cnt
+    mask = ((sat > 150) & (std < 12)).astype(np.uint8) * 255
+    k = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 15))
+    mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, k, iterations=3)
+    mask = cv2.dilate(mask, k, iterations=1)
+    cnts, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    raw = []
+    for c in cnts:
+        x, y, bw, bh = cv2.boundingRect(c)
+        if bw * bh > 2000:
+            raw.append((x, y, x + bw, y + bh))
+    if not raw:
+        return []
+    # 合并相互靠近的区域（同一 logo 的不同彩色部件）
+    merged = []
+    used = [False] * len(raw)
+    for i in range(len(raw)):
+        if used[i]:
+            continue
+        x1, y1, x2, y2 = raw[i]
+        used[i] = True
+        changed = True
+        while changed:
+            changed = False
+            for j in range(len(raw)):
+                if used[j]:
+                    continue
+                bx1, by1, bx2, by2 = raw[j]
+                # 间距 < 80px 视为同一 logo
+                if (bx1 < x2 + 80 and bx2 > x1 - 80 and by1 < y2 + 80 and by2 > y1 - 80):
+                    x1, y1 = min(x1, bx1), min(y1, by1)
+                    x2, y2 = max(x2, bx2), max(y2, by2)
+                    used[j] = True
+                    changed = True
+        # pad 扩展确保覆盖整个 logo（含低饱和的白字/黑字部分）
+        pad = 18
+        x1, y1 = max(0, x1 - pad), max(0, y1 - pad)
+        x2, y2 = min(w, x2 + pad), min(h, y2 + pad)
+        # 角落吸附：logo 通常贴边，若框已靠近某边(<160px)则延伸到该边，
+        # 确保覆盖低饱和的边缘文字（如「黑料网」的黑字）
+        if x1 < 160:
+            x1 = 0
+        if x2 > w - 160:
+            x2 = w
+        if y1 < 160:
+            y1 = 0
+        if y2 > h - 160:
+            y2 = h
+        merged.append((x1, x2, y1, y2))
+    print(f"[ColorLogo] 检测到 {len(merged)} 个彩色 logo: {merged}", flush=True)
+    return merged
+
+
+def detect_watermark_per_frame(video_path: str, sample_step: int = None, pad: int = 6):
+    """逐帧检测半透明/移动文字水印，返回 {frame_no: [(xmin,xmax,ymin,ymax), ...]}。
+
+    适用于 OCR 检测不到的淡色水印（白字、半透明）以及垂直滚动水印。
+    采样：每秒一帧（cap.set 跳帧，长视频也快）。
+    判据：基于轨迹自身（绝对出现秒数 + 出现时段内密度），不依赖视频总长，
+          因此「只在部分时段出现的滚动水印」「贯穿全程偶有漏检的固定水印」都能被保留。
+    """
+    cap = cv2.VideoCapture(video_path)
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+
+    tophat_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15))
+    conn_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (25, 9))
+
+    # 阶段1：每秒一帧采样（cap.set 跳帧；frame_no 用 1-based 与 inpaint 管线一致）
+    step = max(1, int(round(fps)))      # 每秒对应的帧步长
+    sample_frames = list(range(0, total, step))
+    frame_boxes = {}                    # sample_idx -> [box]
+    sidx_to_fno = {}                    # sample_idx -> 1-based 帧号
+    for si, pos in enumerate(sample_frames):
+        cap.set(cv2.CAP_PROP_POS_FRAMES, pos)
+        ret, frame = cap.read()
+        if not ret:
+            continue
+        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+        frame_boxes[si] = _frame_text_candidates(gray, tophat_kernel, conn_kernel)
+        sidx_to_fno[si] = pos + 1
+    cap.release()
+
+    n_sampled = len(frame_boxes)
+    if n_sampled == 0:
+        return {}
+
+    # 阶段2：轨迹连接（按「采样序号」相邻，间断容忍 3 个采样点=3秒）
+    sampled_sidx = sorted(frame_boxes.keys())
+
+    def cx_of(b): return (b[0] + b[1]) / 2
+    def cy_of(b): return (b[2] + b[3]) / 2
+    def wid_of(b): return b[1] - b[0]
+
+    GAP_TOL = 5          # 允许漏检 5 个采样点（5秒，滚动水印经过浅背景会暂时漏检）
+    tracks = []          # {"boxes": {sidx:box}, "last_si":, "last_box":}
+    for si in sampled_sidx:
+        for b in frame_boxes[si]:
+            best, best_d = None, 1e9
+            for tr in tracks:
+                if si - tr["last_si"] > GAP_TOL:
+                    continue
+                lb = tr["last_box"]
+                gap = si - tr["last_si"]
+                dx = abs(cx_of(b) - cx_of(lb))
+                dw = abs(wid_of(b) - wid_of(lb))
+                dy = abs(cy_of(b) - cy_of(lb))
+                # 同一水印：X几乎不变、宽度接近；Y 可滚动（按间隔放宽，每秒最多移动~60px）
+                if dx < 45 and dw < 28 and dy < 60 * max(1, gap):
+                    d = dx * 2 + dw * 2 + dy * 0.2
+                    if d < best_d:
+                        best, best_d = tr, d
+            if best is None:
+                tracks.append({"boxes": {si: b}, "last_si": si, "last_box": b})
+            else:
+                best["boxes"][si] = b
+                best["last_si"] = si
+                best["last_box"] = b
+
+    # 阶段3：筛选有效轨迹 —— 不依赖视频总长
+    #   ① 至少出现 MIN_SECONDS 秒（绝对阈值，滤掉一闪而过的噪点）
+    #   ② 在其出现时段 [first,last] 内密度高（持续存在，不是零散巧合）
+    #   ③ 宽度稳定（文字水印宽度恒定，人体亮区会抖）
+    MIN_SECONDS = 4
+    valid = []   # 有效轨迹的 boxes dict 列表
+    for tr in tracks:
+        boxes = tr["boxes"]
+        n_pts = len(boxes)
+        if n_pts < MIN_SECONDS:
+            continue
+        sids = sorted(boxes.keys())
+        span = sids[-1] - sids[0] + 1
+        density = n_pts / span
+        if density < 0.5:
+            continue
+        widths = [wid_of(b) for b in boxes.values()]
+        wmean = sum(widths) / len(widths)
+        wstd = (sum((x - wmean) ** 2 for x in widths) / len(widths)) ** 0.5
+        if wmean > 0 and wstd / wmean > 0.4:
+            continue
+        valid.append(boxes)
+
+    print(f"[WM-perframe] 每秒一帧采样{n_sampled}帧 轨迹{len(tracks)}条 有效{len(valid)}条", flush=True)
+
+    # 阶段4：每条有效轨迹在其 [first,last] 时段内逐采样点线性插值，
+    #   补全中间漏检的采样点（滚动水印经过浅背景漏检时也能被覆盖），
+    #   再把每个采样点的框展开到对应的 1 秒帧区间，y 方向外扩覆盖滚动移动。
+    ROLL_PAD = 16
+    full = {}
+
+    def lerp_box(ba, bb, t):
+        return tuple(int(round(ba[i] + (bb[i] - ba[i]) * t)) for i in range(4))
+
+    for boxes in valid:
+        sids = sorted(boxes.keys())
+        first, last = sids[0], sids[-1]
+        for si in range(first, last + 1):
+            if si in boxes:
+                b = boxes[si]
+            else:
+                # 线性插值：前后最近的已知采样点
+                prev = max(s for s in sids if s < si)
+                nxt = min(s for s in sids if s > si)
+                b = lerp_box(boxes[prev], boxes[nxt], (si - prev) / (nxt - prev))
+            x1, x2, y1, y2 = b
+            bx = (max(0, x1 - pad), min(w, x2 + pad),
+                  max(0, y1 - ROLL_PAD), min(h, y2 + ROLL_PAD))
+            # 该采样点覆盖 [本帧, 下一采样帧) 的所有原始帧
+            f_start = si * step + 1
+            f_end = (si + 1) * step + 1
+            for f in range(f_start, min(f_end, total + 1)):
+                full.setdefault(f, [])
+                full[f].append(bx)
+
+    # 阶段5：彩色 logo（高饱和+静止）→ 静止框，注入所有帧
+    color_logos = detect_color_logo(video_path)
+    if color_logos:
+        for f in range(1, total + 1):
+            full.setdefault(f, [])
+            full[f] = full[f] + list(color_logos)
+
+    total_boxes = sum(len(v) for v in full.values())
+    print(f"[WM-perframe] 输出 {len(full)} 帧 共 {total_boxes} 个水印框（文字+彩色logo）", flush=True)
+    return full
+
+
+def smart_detect_watermark(video_path: str, sample_frames: int = 30, gpu_id: int = 0):
+    """
+    完整流程：帧差分（前30秒每秒一帧）→ SAM2 精确分割。
+    限制扫描前30秒：H.264压缩导致长视频里即使静态logo的像素std也会随时间累积超过阈值。
+    返回 [(x1, y1, x2, y2), ...] 像素坐标列表。
+    """
+    print(f"[SAM2] smart_detect_watermark 开始: {video_path}", flush=True)
+    # 1. 静态区域检测：跨时间段帧对比法（全视频）
+    boxes = detect_static_regions(video_path, sample_frames)
+    print(f"[SAM2] detect_static_regions 返回 {len(boxes)} 个候选区域: {boxes}", flush=True)
+    if not boxes:
+        return []
+
+    # 2. 读取代表帧用于 SAM2
+    cap = cv2.VideoCapture(video_path)
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    cap.set(cv2.CAP_PROP_POS_FRAMES, total // 2)
+    ret, frame = cap.read()
+    cap.release()
+    if not ret:
+        return boxes  # SAM2 失败时直接返回粗检测结果
+
+    # 3. SAM2 精确分割
+    mask = refine_with_sam2(frame, boxes, gpu_id=gpu_id)
+    if mask is None:
+        return boxes
+
+    # 4. 从 mask 提取最终 bounding boxes
+    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    refined = []
+    for cnt in contours:
+        if cv2.contourArea(cnt) > 100:
+            x, y, bw, bh = cv2.boundingRect(cnt)
+            refined.append((x, y, x + bw, y + bh))
+
+    result = refined if refined else boxes
+    print(f"[SAM2] smart_detect_watermark 最终返回 {len(result)} 个区域: {result}", flush=True)
+    return result

backend/tools/subtitle_detect.py

@@ -21,9 +21,10 @@ class SubtitleDetect:
     # 采样间隔，根据视频帧率在 _init_sample_step 中自适应设置
     SAMPLE_STEP = 3
 
-    def __init__(self, video_path, sub_areas=[]):
+    def __init__(self, video_path, sub_areas=[], gpu_id=0):
         self.video_path = video_path
         self.sub_areas = sub_areas
+        self.gpu_id = gpu_id
         self._init_sample_step()
 
     def _init_sample_step(self):
@@ -46,10 +47,12 @@ class SubtitleDetect:
         hardware_accelerator = HardwareAccelerator.instance()
         onnx_providers = hardware_accelerator.onnx_providers
         model_config = ModelConfig()
+        import torch as _torch
+        _device = f"gpu:{self.gpu_id}" if _torch.cuda.is_available() else "cpu"
         return TextDetection(
             model_name=model_config.DET_MODEL_NAME,
             model_dir=model_config.DET_MODEL_DIR,
-            device="cpu",
+            device=_device,
             enable_hpi=len(onnx_providers) > 0,
         )
 
@@ -81,34 +84,97 @@ class SubtitleDetect:
                             break
         return temp_list
 
-    def find_subtitle_frame_no(self, sub_remover=None):
-        video_cap = cv2.VideoCapture(get_readable_path(self.video_path))
-        frame_count = video_cap.get(cv2.CAP_PROP_FRAME_COUNT)
-        tbar = tqdm(total=int(frame_count), unit='frame', position=0, file=sys.__stdout__, desc='Subtitle Finding')
-        current_frame_no = 0
-        # 阶段1：采样检测，仅对每隔 sample_step 帧执行 OCR
-        sampled_results = {}  # frame_no -> temp_list
-        if sub_remover:
-            sub_remover.append_output(tr['Main']['ProcessingStartFindingSubtitles'])
-        while video_cap.isOpened():
-            ret, frame = video_cap.read()
-            # 如果读取视频帧失败（视频读到最后一帧）
+    def _scan_range(self, start_frame, end_frame, ab_sections, results, lock,
+                    progress_counter, frame_count, sub_remover, gpu_id=None):
+        """单线程扫描指定帧范围。
+        每个线程在自己内部初始化 OCR 模型（CUDA 上下文归该线程所有），避免跨线程 CUDA 问题。
+        gpu_id: >=0 用 GPU，None 用 CPU。
+        """
+        # 在本线程内创建并初始化 OCR 模型，CUDA 上下文属于本线程
+        if gpu_id is not None and gpu_id >= 0:
+            import torch as _t
+            if _t.cuda.is_available():
+                _local_det = SubtitleDetect(self.video_path, self.sub_areas, gpu_id=gpu_id)
+                _ = _local_det.text_detector  # 在本线程建立 CUDA 上下文
+            else:
+                _local_det = SubtitleDetect(self.video_path, self.sub_areas)
+        else:
+            _local_det = SubtitleDetect(self.video_path, self.sub_areas)
+
+        cap = cv2.VideoCapture(get_readable_path(self.video_path))
+        cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
+        frame_no = start_frame
+        while frame_no < end_frame:
+            ret, frame = cap.read()
             if not ret:
                 break
-            # 读取视频帧成功
-            current_frame_no += 1
-            if not is_frame_number_in_ab_sections(current_frame_no - 1, sub_remover.ab_sections):
-                tbar.update(1)
+            frame_no += 1
+            if not is_frame_number_in_ab_sections(frame_no - 1, ab_sections):
                 continue
-            # 仅对采样帧执行 OCR 推理
-            if (current_frame_no - 1) % self.SAMPLE_STEP == 0 or self.SAMPLE_STEP <= 1:
-                temp_list = self.detect_subtitle(frame)
-                if len(temp_list) > 0:
-                    sampled_results[current_frame_no] = temp_list
-            tbar.update(1)
-            if sub_remover:
-                sub_remover.progress_total = (100 * float(current_frame_no) / float(frame_count)) // 2
+            if (frame_no - 1) % self.SAMPLE_STEP == 0 or self.SAMPLE_STEP <= 1:
+                temp_list = _local_det.detect_subtitle(frame)
+                if temp_list:
+                    with lock:
+                        results[frame_no] = temp_list
+            with lock:
+                progress_counter[0] += 1
+                if sub_remover:
+                    sub_remover.progress_total = (100 * progress_counter[0] / frame_count) // 2
+                    sub_remover.notify_progress_listeners()
+        cap.release()
+
+    def find_subtitle_frame_no(self, sub_remover=None, num_workers=4):
+        import threading
+        video_cap = cv2.VideoCapture(get_readable_path(self.video_path))
+        frame_count = int(video_cap.get(cv2.CAP_PROP_FRAME_COUNT))
         video_cap.release()
+        tbar = tqdm(total=frame_count, unit='frame', position=0, file=sys.__stdout__, desc='Subtitle Finding')
+        # 阶段1：并行采样检测
+        sampled_results = {}
+        if sub_remover:
+            sub_remover.append_output(tr['Main']['ProcessingStartFindingSubtitles'])
+        ab_sections = sub_remover.ab_sections if sub_remover else []
+        lock = threading.Lock()
+        import torch as _torch
+        cuda_count = _torch.cuda.device_count() if _torch.cuda.is_available() else 0
+        # 每个线程在自己内部初始化模型，CUDA 上下文属于该线程，避免跨线程推理失败
+        # GPU0 用于 OCR，GPU1 留给 STTN/LaMa（若可用）；无 GPU 则 CPU
+        _ocr_gpu_id = 0 if cuda_count >= 1 else None
+        print(f"[OCR] {'GPU' + str(_ocr_gpu_id) if _ocr_gpu_id is not None else 'CPU'} 推理，{num_workers} 线程并行（每线程独立模型）", flush=True)
+        progress_counter = [0]
+        chunk = max(1, frame_count // num_workers)
+        ranges = [(i * chunk, min((i + 1) * chunk, frame_count)) for i in range(num_workers)]
+        threads = []
+        for i, (s, e) in enumerate(ranges):
+            t = threading.Thread(
+                target=self._scan_range,
+                args=(s, e, ab_sections, sampled_results, lock, progress_counter, frame_count, sub_remover, _ocr_gpu_id),
+                daemon=True,
+            )
+            threads.append(t)
+            t.start()
+        # tqdm 进度跟踪（单独线程更新）
+        import time as _time
+        last = 0
+        while any(t.is_alive() for t in threads):
+            with lock:
+                cur = progress_counter[0]
+            tbar.update(cur - last)
+            last = cur
+            _time.sleep(0.2)
+        for t in threads:
+            t.join()
+        tbar.update(frame_count - last)
+        tbar.close()
+        # 线程退出后其局部 OCR 模型自动被 GC，触发 GPU 显存释放
+        import gc as _gc
+        _gc.collect()
+        try:
+            import torch as _t2
+            if _t2.cuda.is_available():
+                _t2.cuda.empty_cache()
+        except Exception:
+            pass
         # 阶段2：插值填充 — 两个采样帧之间都有字幕时，中间帧也标记为有字幕
         subtitle_frame_no_box_dict = {}
         detected_nos = sorted(sampled_results.keys())

backend/tools/video_io.py

@@ -25,7 +25,13 @@ class FramePrefetcher:
     def _read_loop(self):
         while not self._stopped:
             ret, frame = self.cap.read()
-            self._buffer.put((ret, frame))
+            # 用带超时的 put，避免队列满时永久阻塞
+            while not self._stopped:
+                try:
+                    self._buffer.put((ret, frame), timeout=0.5)
+                    break
+                except queue.Full:
+                    pass
             if not ret:
                 break
 
@@ -39,11 +45,12 @@ class FramePrefetcher:
     def stop(self):
         """停止预读取，不释放底层 video_cap。"""
         self._stopped = True
-        try:
-            while not self._buffer.empty():
+        # 排空队列，解除后台线程可能阻塞的 put()
+        while True:
+            try:
                 self._buffer.get_nowait()
-        except queue.Empty:
-            pass
+            except queue.Empty:
+                break
         self._thread.join(timeout=5)
 
     def release(self):
@@ -53,52 +60,65 @@ class FramePrefetcher:
 
 class FFmpegVideoWriter:
     """
-    通过 FFmpeg 管道写入帧，使用 libx264 编码。
-    接口兼容 cv2.VideoWriter（write/release）。
+    独立 I/O 线程写帧到 FFmpeg stdin，直接输出 H.264。
+    主线程只往队列里放帧，永不阻塞在 I/O 上。
     """
 
     def __init__(self, output_path, fps, size):
         w, h = size
-        cmd = [
-            FFmpegCLI.instance().ffmpeg_path,
-            '-y',
-            '-f', 'rawvideo',
-            '-vcodec', 'rawvideo',
-            '-s', f'{w}x{h}',
-            '-pix_fmt', 'bgr24',
-            '-r', str(fps),
-            '-i', '-',
-            '-c:v', 'libx264',
-            '-pix_fmt', 'yuv420p',
-            '-crf', '18',
-            '-preset', 'fast',
-            '-loglevel', 'error',
-            output_path
-        ]
-        self._process = subprocess.Popen(
-            cmd,
-            stdin=subprocess.PIPE,
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
+        self._queue = queue.Queue(maxsize=60)
+        self._error = None
+        # 优先用系统 FFmpeg + NVENC（GPU 编码），回退到 libx264
+        import shutil as _shutil
+        _sys_ffmpeg = '/usr/bin/ffmpeg'
+        _nvenc_ok = (
+            _shutil.which(_sys_ffmpeg) is not None and
+            subprocess.run(
+                [_sys_ffmpeg, '-hide_banner', '-encoders'],
+                capture_output=True, text=True
+            ).stdout.find('h264_nvenc') >= 0
         )
+        ffmpeg_bin = _sys_ffmpeg if _nvenc_ok else FFmpegCLI.instance().ffmpeg_path
+        codec_args = (
+            ['-vf', 'format=yuv420p', '-vcodec', 'h264_nvenc', '-cq', '18', '-preset', 'p4']
+            if _nvenc_ok else
+            ['-vcodec', 'libx264', '-crf', '18', '-preset', 'ultrafast']
+        )
+        cmd = [
+            ffmpeg_bin, '-y',
+            '-f', 'rawvideo', '-vcodec', 'rawvideo',
+            '-s', f'{w}x{h}', '-pix_fmt', 'bgr24',
+            '-r', str(fps), '-i', 'pipe:0',
+            *codec_args,
+            '-loglevel', 'error', output_path,
+        ]
+        self._proc = subprocess.Popen(cmd, stdin=subprocess.PIPE)
+        self._thread = threading.Thread(target=self._io_loop, daemon=True)
+        self._thread.start()
+
+    def _io_loop(self):
+        try:
+            while True:
+                frame = self._queue.get()
+                if frame is None:   # sentinel
+                    break
+                self._proc.stdin.write(frame.tobytes())
+        except Exception as e:
+            self._error = e
+        finally:
+            try:
+                self._proc.stdin.close()
+            except Exception:
+                pass
 
     def write(self, frame):
-        """写入一帧（numpy BGR 数组）。"""
         if frame.dtype != np.uint8:
             frame = np.clip(frame, 0, 255).astype(np.uint8)
-        try:
-            self._process.stdin.write(frame.tobytes())
-        except BrokenPipeError:
-            pass
+        self._queue.put(frame)
 
     def release(self):
-        """关闭管道并等待编码完成。"""
-        try:
-            self._process.stdin.close()
-        except BrokenPipeError:
-            pass
-        try:
-            self._process.wait(timeout=600)
-        except subprocess.TimeoutExpired:
-            self._process.terminate()
-            self._process.wait(timeout=5)
+        self._queue.put(None)       # 通知 I/O 线程退出
+        self._thread.join()         # 等待所有帧写完
+        self._proc.wait()           # 等待 FFmpeg 完成编码
+        if self._error:
+            raise RuntimeError(f"FFmpegVideoWriter error: {self._error}")

qfluentwidgets/__init__.py

@@ -0,0 +1,155 @@
+"""
+Minimal qfluentwidgets stub for headless/server use — no Qt required.
+Provides just enough for backend/config.py to import and function correctly.
+"""
+
+import json
+import os
+
+
+class _Signal:
+    def connect(self, fn): pass
+    def disconnect(self, fn=None): pass
+    def emit(self, *args): pass
+
+
+class ConfigValidator:
+    def validate(self, v): return True
+    def correct(self, v): return v
+
+
+class BoolValidator(ConfigValidator):
+    def validate(self, v): return isinstance(v, bool)
+    def correct(self, v): return bool(v)
+
+
+class OptionsValidator(ConfigValidator):
+    def __init__(self, options):
+        try:
+            self._opts = [e for e in options]
+        except TypeError:
+            self._opts = []
+
+    def validate(self, v): return v in self._opts
+    def correct(self, v): return v if v in self._opts else (self._opts[0] if self._opts else v)
+
+
+class RangeValidator(ConfigValidator):
+    def __init__(self, min_val, max_val):
+        self.min_val = min_val
+        self.max_val = max_val
+
+    def validate(self, v): return self.min_val <= v <= self.max_val
+    def correct(self, v): return max(self.min_val, min(self.max_val, v))
+
+
+class EnumSerializer:
+    def __init__(self, enum_cls):
+        self.enum_cls = enum_cls
+
+    def serialize(self, v):
+        return v.value if hasattr(v, 'value') else str(v)
+
+    def deserialize(self, s):
+        for e in self.enum_cls:
+            if e.value == s:
+                return e
+        return s
+
+
+class ConfigItem:
+    def __init__(self, section, key, default, validator=None, restart=False):
+        self.section = section
+        self.key = key
+        self.value = default
+        self.validator = validator or ConfigValidator()
+        self.restart = restart
+        self.valueChanged = _Signal()
+
+
+class OptionsConfigItem(ConfigItem):
+    def __init__(self, section, key, default, validator=None, serializer=None, restart=False):
+        super().__init__(section, key, default, validator, restart)
+        self.serializer = serializer
+
+
+class RangeConfigItem(ConfigItem):
+    def __init__(self, section, key, default, validator=None):
+        super().__init__(section, key, default, validator)
+
+
+class QConfig:
+    appRestartSig = _Signal()
+
+    def set(self, item, value):
+        item.value = value
+        try:
+            item.valueChanged.emit(value)
+        except Exception:
+            pass
+
+
+class _QConfigLoader:
+    def load(self, filepath, config_obj):
+        if not os.path.exists(filepath):
+            return
+        try:
+            with open(filepath, 'r', encoding='utf-8') as f:
+                data = json.load(f)
+        except Exception:
+            return
+
+        cls = type(config_obj)
+        for attr_name in dir(cls):
+            if attr_name.startswith('_'):
+                continue
+            item = getattr(cls, attr_name, None)
+            if not isinstance(item, ConfigItem):
+                continue
+            section_data = data.get(item.section, {})
+            if item.key not in section_data:
+                continue
+            raw = section_data[item.key]
+            if isinstance(item, OptionsConfigItem) and item.serializer:
+                try:
+                    raw = item.serializer.deserialize(raw)
+                except Exception:
+                    pass
+            if item.validator.validate(raw):
+                item.value = raw
+
+
+qconfig = _QConfigLoader()
+
+# ── additional symbols referenced by gui code (safe no-ops) ──────────────
+class FluentIcon:
+    HOME = None; SETTING = None; FOLDER = None; PLAY = None
+
+class Theme:
+    DARK = 'dark'; LIGHT = 'light'
+
+class NavigationItemPosition:
+    BOTTOM = 'bottom'
+
+def setTheme(*a, **k): pass
+def setThemeColor(*a, **k): pass
+def isDarkTheme(): return True
+
+# Fluent widget stubs — only imported by GUI code, never used by backend
+class _Widget:
+    def __init__(self, *a, **k): pass
+
+CardWidget = _Widget
+PushButton = _Widget
+Slider = _Widget
+ProgressBar = _Widget
+PlainTextEdit = _Widget
+TextEdit = _Widget
+SubtitleLabel = _Widget
+HollowHandleStyle = _Widget
+InfoBar = _Widget
+SettingCardGroup = _Widget
+ComboBoxSettingCard = _Widget
+SwitchSettingCard = _Widget
+FluentWindow = _Widget
+NavigationWidget = _Widget