'commit'

Apps/TeLaiDian/ReadImageKit.py

@@ -245,12 +245,14 @@ class ReadImageKit:
         cv_bboxes = detect_cards_cv(image_path, top_ratio=SAFE_EXCLUDE_RATIO, bottom_ratio=BOTTOM_SAFE_EXCLUDE_RATIO)
         
         if cv_bboxes:
-            draw_rectangles(image_path, bboxes=cv_bboxes)
+            img = read_image(image_path)
+            h, w = img.shape[:2] if img is not None else (2400, 1080)
+            y_threshold = h * 0.58
+            # 在最终绘制阶段再生成可视化，避免早期绘制产生误导
             prompt = f"""
-            图片中已经用绿色矩形框标记了 {len(cv_bboxes)} 个可能的充电站卡片。
-            请按从上到下的顺序，识别每个绿色框内的场站信息。
+            图片下半部分可能包含多个充电站卡片，请按从上到下识别这些卡片的名称与地址。
             
-            输出格式为 JSON 数组，长度必须为 {len(cv_bboxes)}。
+            输出格式为 JSON 数组。
             每个对象包含：
             - "name": 场站名称
             - "address": 场站地址
@@ -281,23 +283,51 @@ class ReadImageKit:
                 img = read_image(image_path)
                 h, w = img.shape[:2] if img is not None else (2400, 1080)
                 y_threshold = h * 0.58
-                
-                for i, res in enumerate(vlm_results):
-                    if i < len(cv_bboxes):
-                        bbox = cv_bboxes[i]
-                        # 物理坐标硬过滤：即使 CV 识别到了，y 坐标在 58% 以上的也统统干掉
-                        center_y = (bbox[1] + bbox[3]) // 2
-                        if center_y < y_threshold:
-                            logger.warning(f"CV 误报过滤 (y={center_y} < {y_threshold}): 可能是顶部营销卡片")
-                            continue
-                            
-                        if res and (res.get("is_valid") is True or (res.get("name") and res.get("is_valid") is not False)):
-                            final_stations.append({
-                                "name": res.get("name"),
-                                "address": res.get("address"),
-                                "point": [(bbox[0] + bbox[2]) // 2, (bbox[1] + bbox[3]) // 2],
-                                "bbox": bbox
-                            })
+                bottom_threshold = int(h * (1 - BOTTOM_SAFE_EXCLUDE_RATIO))
+                use_bboxes = []
+                for b in cv_bboxes:
+                    cy = (b[1] + b[3]) // 2
+                    if cy >= y_threshold:
+                        use_bboxes.append(b)
+                for res in vlm_results:
+                    if not res:
+                        continue
+                    p = res.get("point")
+                    actual_p = None
+                    if p and len(p) == 2:
+                        actual_p = p
+                        if p[0] <= 1000 and p[1] <= 1000:
+                            actual_p = [int(p[0] * w / 1000), int(p[1] * h / 1000)]
+                    # 顶部与底部安全过滤
+                    if actual_p and (actual_p[1] < y_threshold or actual_p[1] > bottom_threshold):
+                        logger.warning(f"过滤掉可能的误点击点 (y={actual_p[1]} < {y_threshold}): {res.get('name')}")
+                        continue
+                    matched_bbox = None
+                    if actual_p:
+                        for b in use_bboxes:
+                            if b[0] <= actual_p[0] <= b[2] and b[1] <= actual_p[1] <= b[3]:
+                                matched_bbox = b
+                                break
+                    # 如果没有匹配到 CV 框，则根据点击点生成一个虚拟框
+                    if matched_bbox is None and actual_p is not None:
+                        half_w = int(w * 0.4)
+                        half_h = 90
+                        x1 = max(0, actual_p[0] - half_w)
+                        x2 = min(w, actual_p[0] + half_w)
+                        y1 = max(0, actual_p[1] - half_h)
+                        y2 = min(h, actual_p[1] + half_h)
+                        matched_bbox = [x1, y1, x2, y2]
+                    # 如果既没有 point 也没有匹配框，跳过
+                    if actual_p is None and matched_bbox is None:
+                        continue
+                    if actual_p is None and matched_bbox:
+                        actual_p = [(matched_bbox[0] + matched_bbox[2]) // 2, (matched_bbox[1] + matched_bbox[3]) // 2]
+                    final_stations.append({
+                        "name": res.get("name"),
+                        "address": res.get("address"),
+                        "point": actual_p,
+                        "bbox": matched_bbox
+                    })
             elif not cv_bboxes:
                 vlm_list = vlm_results if isinstance(vlm_results, list) else []
                 w, h = 0, 0