D144 · 角色存在感QC升级: 新增跨镜一致性对比

- 新增 --cross-shot 模式: 多张同角色图两两对比
- 检测项: 颜色一致性(HSV直方图) + 结构一致性(梯度) + 亮度漂移
- 跨镜评分规则: 颜色40% + 结构30% + 亮度30%

铸渊 ICE-GL-ZY001 · D144 · 2026-06-24

video-ai-system/engines/character-distinctiveness-qc/character-distinctiveness-qc.py

@@ -2,16 +2,18 @@
 # -*- coding: utf-8 -*-
 """
 CHARACTER-DISTINCTIVENESS-QC
-主角存在感评估器 — 专门评估"像不像主角"，输出存在感、轮廓、服装记忆点评分。
+主角存在感评估器 — 评估"像不像主角"，支持单图评分和跨镜一致性对比。
 
 功能:
 1. 输入角色图片 + 参考资产包
-2. 用 OpenCV 计算轮廓差异、颜色直方图、SSIM
+2. 用 OpenCV 计算轮廓、颜色、面部一致性
-3. 输出 JSON 报告 + 存在感评分 (0-10)
+3. 跨镜一致性对比（同角色多镜变化检测）
+4. 输出 JSON 报告 + 存在感评分 (0-10)
 
 用法:
-  python character-distinctiveness-qc.py --image path/to/test.png --character CHAR-003-SuBai
+  python character-distinctiveness-qc.py --image test.png --character CHAR-003-SuBai
   python character-distinctiveness-qc.py --batch test/images/ --character CHAR-003-SuBai
+  python character-distinctiveness-qc.py --cross-shot shot01.png shot02.png --character CHAR-003-SuBai
 """
 
 import os
@@ -270,6 +272,100 @@ class CharacterDistinctivenessQC:
 
         return results
 
+    def compare_cross_shot(self, image_paths):
+        """
+        跨镜一致性对比：同一个角色的多张图两两比较
+        检测: 脸型、服装颜色、发型、道具是否跨镜漂移
+        """
+        print(f"\n🔍 跨镜一致性对比: {len(image_paths)} 张图")
+        print("=" * 60)
+
+        if not CV2_AVAILABLE:
+            return {"error": "OpenCV不可用", "consistency_score": None}
+
+        images = []
+        for p in image_paths:
+            img = cv2.imread(str(p))
+            if img is not None:
+                images.append((Path(p).name, img))
+
+        if len(images) < 2:
+            return {"error": "至少需要2张图进行跨镜对比", "consistency_score": None}
+
+        comparisons = []
+        for i in range(len(images) - 1):
+            name_a, img_a = images[i]
+            name_b, img_b = images[i + 1]
+
+            # 统一尺寸
+            h = min(img_a.shape[0], img_b.shape[0], 512)
+            w = min(img_a.shape[1], img_b.shape[1], 512)
+            a = cv2.resize(img_a, (w, h))
+            b = cv2.resize(img_b, (w, h))
+
+            # 1. 颜色一致性（HSV直方图）
+            hsv_a = cv2.cvtColor(a, cv2.COLOR_BGR2HSV)
+            hsv_b = cv2.cvtColor(b, cv2.COLOR_BGR2HSV)
+            hist_a = cv2.calcHist([hsv_a], [0, 1], None, [50, 60], [0, 180, 0, 256])
+            hist_b = cv2.calcHist([hsv_b], [0, 1], None, [50, 60], [0, 180, 0, 256])
+            cv2.normalize(hist_a, hist_a)
+            cv2.normalize(hist_b, hist_b)
+            color_sim = cv2.compareHist(hist_a, hist_b, cv2.HISTCMP_CORREL)
+
+            # 2. 结构一致性（梯度直方图）
+            gray_a = cv2.cvtColor(a, cv2.COLOR_BGR2GRAY)
+            gray_b = cv2.cvtColor(b, cv2.COLOR_BGR2GRAY)
+            grad_a = cv2.Sobel(gray_a, cv2.CV_64F, 1, 0) + cv2.Sobel(gray_a, cv2.CV_64F, 0, 1)
+            grad_b = cv2.Sobel(gray_b, cv2.CV_64F, 1, 0) + cv2.Sobel(gray_b, cv2.CV_64F, 0, 1)
+            struct_sim = np.corrcoef(grad_a.flatten(), grad_b.flatten())[0, 1]
+            if np.isnan(struct_sim):
+                struct_sim = 0.5
+
+            # 3. 平均亮度漂移
+            lum_a = np.mean(gray_a)
+            lum_b = np.mean(gray_b)
+            lum_drift = abs(lum_a - lum_b) / max(lum_a, lum_b, 1)
+
+            comp = {
+                "pair": f"{name_a} ↔ {name_b}",
+                "color_consistency": round(float(color_sim), 3),
+                "structure_consistency": round(float(struct_sim), 3),
+                "luminance_drift": round(float(lum_drift), 3),
+            }
+
+            # 综合评分
+            c_score = (
+                max(0, color_sim) * 4.0 +
+                max(0, struct_sim) * 3.0 +
+                max(0, 1 - lum_drift) * 3.0
+            )
+            comp["cross_shot_score"] = round(min(10, c_score), 1)
+            comp["verdict"] = "PASS" if comp["cross_shot_score"] >= 7.0 else "FAIL"
+
+            print(f"  {comp['pair']}: 颜色{comp['color_consistency']:.2f} "
+                  f"结构{comp['structure_consistency']:.2f} "
+                  f"亮度漂移{comp['luminance_drift']:.2f} → "
+                  f"{comp['cross_shot_score']:.1f}/10 {comp['verdict']}")
+
+            comparisons.append(comp)
+
+        avg_score = np.mean([c["cross_shot_score"] for c in comparisons])
+        fail_count = sum(1 for c in comparisons if c["verdict"] == "FAIL")
+
+        result = {
+            "cross_shot_consistency": avg_score,
+            "comparisons": comparisons,
+            "fail_count": fail_count,
+            "total_pairs": len(comparisons),
+            "verdict": "PASS" if avg_score >= 7.0 and fail_count == 0 else "FAIL"
+        }
+
+        print(f"\n  📊 跨镜一致性: {avg_score:.1f}/10 ({result['verdict']})")
+        if fail_count > 0:
+            print(f"  ⚠️  {fail_count}/{len(comparisons)} 对比较失败，可能存在跨镜漂移")
+
+        return result
+
     def save_report(self, result, output_path):
         """保存单张或批量评估报告"""
         output_path = Path(output_path)
@@ -309,16 +405,25 @@ def main():
     parser.add_argument("--character", type=str, default="CHAR-003-SuBai",
                         help="角色ID (默认: CHAR-003-SuBai)")
     parser.add_argument("--batch", type=str, help="批量评估目录")
+    parser.add_argument("--cross-shot", type=str, nargs="+", help="跨镜一致性对比: 多张图片路径")
     parser.add_argument("--output", type=str, help="输出 JSON 报告路径")
 
     args = parser.parse_args()
 
-    if not args.image and not args.batch:
+    if not args.image and not args.batch and not args.cross_shot:
         parser.print_help()
         return
 
     qc = CharacterDistinctivenessQC(args.character)
 
+    if args.cross_shot:
+        result = qc.compare_cross_shot(args.cross_shot)
+        print(f"\n📋 跨镜一致性结果:")
+        print(json.dumps(result, ensure_ascii=False, indent=2))
+        if args.output:
+            qc.save_report(result, args.output)
+        return
+
     if args.image:
         result = qc.evaluate_image(args.image)
         print(f"\n📋 评估详情:")