'commit'

Apps/AiTeJiYiChong/ReadImageKit.py

@@ -8,6 +8,7 @@ import aiohttp
 import logging
 import base64
 import cv2
+import uuid
 from openai import AsyncOpenAI, BadRequestError
 from Config.Config import (
     ALY_LLM_API_KEY, VL_MODEL_NAME, VL_MODEL_NAME_AD
@@ -16,7 +17,6 @@ from Apps.AiTeJiYiChong.Config.Setting import (
     SAFE_EXCLUDE_RATIO, FALLBACK_WIDTH, FALLBACK_HEIGHT,
     BOTTOM_SAFE_EXCLUDE_RATIO
 )
-from Util.PaddleOCRKit import get_ocr_kit
 from Apps.AiTeJiYiChong import Kit
 
 logger = logging.getLogger(__name__)
@@ -290,51 +290,107 @@ class ReadImageKit:
             logger.error(f"Error calling VL model for detail: {e}")
             return None
 
-    @classmethod
-    async def get_stations_hybrid(cls, image_path: str, device_info=None):
+    @staticmethod
+    async def recognize_card_with_vl(image_path: str):
         """
-        混合识别模式：图形学切片 + 本地 PaddleOCR 识别
+        对卡片切片进行 VL 识别
         """
         if not os.path.exists(image_path):
-            logger.error(f"Image not found: {image_path}")
-            return []
+            return None
+
+        with open(image_path, "rb") as f:
+            encoded = base64.b64encode(f.read()).decode("utf-8")
+            img_input = {"url": f"data:image/jpeg;base64,{encoded}"}
+
+        prompt = (
+            "请识别图片中的充电站卡片信息，并以 JSON 格式输出：\n"
+            "{\n"
+            "  \"station_name\": \"场站名称\",\n"
+            "  \"price\": \"价格(数字)\",\n"
+            "  \"piles\": \"空闲/总数(如 4/4)\",\n"
+            "  \"distance\": \"距离(如 2.5km)\",\n"
+            "  \"parking\": \"停车描述\"\n"
+            "}\n"
+            "只输出 JSON，不要有其他文字。"
+        )
+
+        try:
+            resp = await ReadImageKit._client.chat.completions.create(
+                model=VL_MODEL_NAME,
+                messages=[
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "text", "text": prompt},
+                            {"type": "image_url", "image_url": img_input}
+                        ]
+                    }
+                ],
+                max_tokens=300,
+                temperature=0.01
+            )
+            content = resp.choices[0].message.content.strip()
+            raw = ReadImageKit._extract_json(content)
+            return json.loads(raw)
+        except Exception as e:
+            logger.error(f"VL recognize_card_with_vl error: {e}")
+            return None
+
+    @staticmethod
+    async def get_stations_hybrid(image_path, device_info=None):
+        """
+        混合识别模式：图形学切片 + VL 模型识别
+        """
+        if device_info is None:
+            device_info = ReadImageKit._FALLBACK_DEVICE_INFO
 
         # 1. 使用 Kit 中的图形学算法识别卡片区域
-        # Kit.crop_cards_from_image 会生成 .json, _flag.jpg, _vl.jpg
-        # 我们主要需要它返回的 json_data
         json_data = Kit.crop_cards_from_image(image_path, save_debug=True)
         if not json_data or not json_data.get("cards"):
             logger.warning("No cards detected by graphical slicing.")
             return []
 
+        # 2. 对每个卡片区域进行 VL 识别
+        tasks = []
+        card_metas = []
+        
+        # 为了使用 VL，我们需要将每个卡片切出来保存到临时文件
+        temp_dir = os.path.join(os.path.dirname(image_path), "Patches")
+        os.makedirs(temp_dir, exist_ok=True)
+        
         img = Image.open(image_path).convert("RGB")
-        ocr_kit = get_ocr_kit()
         
-        final_stations = []
-        
-        # 2. 对每个卡片区域进行 OCR 识别
-        # 注意：PaddleOCR 识别过程较快，且通常不涉及网络请求，可以根据需要选择并行或串行
-        # 这里使用串行以保证日志输出整齐，如果追求极致性能可改用 asyncio.to_thread 并行
         for card in json_data["cards"]:
             rect = card["rect"] # [x1, y1, x2, y2]
-            # 裁剪卡片
+            patch_path = os.path.join(temp_dir, f"patch_{card['id']}_{uuid.uuid4().hex[:8]}.jpg")
+            
+            # 裁剪并保存
             patch = img.crop((rect[0], rect[1], rect[2], rect[3]))
+            patch.save(patch_path)
             
-            # 转换为 ndarray 供 PaddleOCR 使用
-            patch_cv = cv2.cvtColor(np.array(patch), cv2.COLOR_RGB2BGR)
+            card_metas.append({
+                "meta": card,
+                "patch_path": patch_path
+            })
+            tasks.append(ReadImageKit.recognize_card_with_vl(patch_path))
+
+        # 并行执行 VL 识别
+        results = await asyncio.gather(*tasks)
+        
+        final_stations = []
+        for i, res in enumerate(results):
+            meta = card_metas[i]["meta"]
+            patch_path = card_metas[i]["patch_path"]
             
-            # OCR 识别
-            logger.info(f"正在识别卡片 {card['id']}: {rect}")
-            res = ocr_kit.recognize(patch_cv)
-            
-            if res and res.get("station_name"):
+            if res and isinstance(res, dict) and res.get("station_name"):
                 # 注入点击坐标和原始区域信息
-                res["uia_center_x"] = card["click_point"][0]
-                res["uia_center_y"] = card["click_point"][1]
-                res["rect"] = rect
+                res["uia_center_x"] = meta["click_point"][0]
+                res["uia_center_y"] = meta["click_point"][1]
+                res["rect"] = meta["rect"]
                 
-                # 转换 bounds 到 0-1000 空间（保持与 VL 模式兼容）
+                # 转换 bounds 到 0-1000 空间
                 w, h = img.size
+                rect = meta["rect"]
                 res["bounds"] = [
                     int(rect[0] * 1000 / w),
                     int(rect[1] * 1000 / h),
@@ -343,9 +399,16 @@ class ReadImageKit:
                 ]
                 
                 final_stations.append(res)
-                logger.info(f"卡片 {card['id']} 识别成功: {res['station_name']}")
+                logger.info(f"卡片 {meta['id']} 识别成功: {res['station_name']}")
             else:
-                logger.warning(f"卡片 {card['id']} 识别失败或无名称")
+                logger.warning(f"卡片 {meta['id']} 识别失败或无名称")
+            
+            # 清理临时切片文件
+            try:
+                if os.path.exists(patch_path):
+                    os.remove(patch_path)
+            except:
+                pass
 
         return final_stations