ai通过坐标处理，增加通过跳过ocr直接用ai处理的开关

ccs_base/wizard/batch_get_pod_info_wizard_views.xml

@@ -13,6 +13,8 @@
                         <group>
                             <field name="remove_specified_text" readonly="1" widget="boolean_toggle"
                                    attrs="{'invisible': [('pdf_file', '!=', False)]}"/>
+                            <field name="skip_ocr_direct_ai" readonly="0" widget="boolean_toggle"
+                                   attrs="{'invisible': [('pdf_file', '!=', False)]}"/>
                         </group>
                         <group>
                             <field name="sync_last_mile_pod" widget="boolean_toggle"

ccs_base/wizard/image-to-coordinate.py

@@ -6,7 +6,8 @@ import base64
 import fitz  # PyMuPDF
 import json
 from PIL import Image, ImageDraw
-
+import time
+begin_time = time.time()
 client = OpenAI(
     # 若没有配置环境变量，请用阿里云百炼API Key将下行替换为：api_key="sk-xxx",
     # 新加坡和北京地域的API Key不同。获取API Key：https://help.aliyun.com/zh/model-studio/get-api-key
@@ -15,7 +16,7 @@ client = OpenAI(
     base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
 )
 
-pdf_path = "C:/Users/Administrator/Desktop/43610281036.pdf"
+pdf_path = "./43610272216.pdf"
 
 def pdf_to_images(pdf_path, output_dir='./pdf_pages', dpi=150):
     """
@@ -69,10 +70,16 @@ def safe_extract_json(text: str):
     """从模型返回文本中尽可能鲁棒地提取JSON对象。"""
     # 直接尝试解析
     try:
-        return json.loads(text)
+        if text.startswith("```json"):
+            text = text[7:-3].strip()
+        obj = json.loads(text)
+        if isinstance(obj, list):
+            return {'rects': obj}
+        return obj
     except Exception:
         pass
 
+    print(text)
     # 尝试提取首尾花括号之间的内容
     start = text.find('{')
     end = text.rfind('}')
@@ -169,6 +176,88 @@ def draw_debug_boxes(image_path: str, coords_map: dict, save_path: str):
     img.save(save_path)
     print(f"调试框已生成: {save_path}")
 
+def convert_ai_json_to_coords_map(result, img_w: int, img_h: int) -> dict:
+    """
+    将AI返回的JSON统一转换为 {label: [x1,y1,x2,y2]} 形式，兼容多种结构：
+    1) {"rects":[{"text":"AGN","bbox_norm":{x1,y1,x2,y2},"bbox_px":{x1,y1,x2,y2}}]}
+    2) {"AGN":[x1,y1,x2,y2], "UCLINK":[...], ...}
+    3) {"rects":[{"label":"AGN","bbox":[x1,y1,x2,y2]}]}
+    4) {"rects":[{"text":"AGN","x1":...,"y1":...,"x2":...,"y2":...}]}
+    返回值可以包含像素或归一化坐标，后续由 normalize_bbox 统一处理。
+    """
+    coords_map: dict = {}
+
+    def dict_to_list(b):
+        if isinstance(b, dict):
+            return [b.get("x1"), b.get("y1"), b.get("x2"), b.get("y2")]
+        return b
+
+    try:
+        # 情形A：顶层是dict
+        if isinstance(result, dict):
+            # A1：包含 rects 列表
+            if "rects" in result and isinstance(result["rects"], list):
+                for i, item in enumerate(result["rects"]):
+                    if not isinstance(item, dict):
+                        continue
+                    label = item.get("text") or item.get("label") or item.get("word") or f"rect_{i}"
+                    idx = item.get("occurrence_index")
+                    key = f"{label}#{idx}" if isinstance(idx, int) and idx > 0 else label
+
+                    bbox_px = dict_to_list(item.get("bbox_px") or item.get("bbox_pixels"))
+                    bbox_norm = dict_to_list(item.get("bbox_norm"))
+                    bbox_generic = dict_to_list(item.get("bbox"))
+
+                    chosen = None
+                    # 如果同时存在像素和归一化，做一致性校验
+                    if isinstance(bbox_px, (list, tuple)) and len(bbox_px) == 4 and isinstance(bbox_norm, (list, tuple)) and len(bbox_norm) == 4:
+                        try:
+                            px_from_norm = [int(round(float(bbox_norm[0]) * img_w)),
+                                            int(round(float(bbox_norm[1]) * img_h)),
+                                            int(round(float(bbox_norm[2]) * img_w)),
+                                            int(round(float(bbox_norm[3]) * img_h))]
+                            diff = sum(abs(px_from_norm[j] - int(round(float(bbox_px[j])))) for j in range(4))
+                            chosen = bbox_px if diff <= 4 else bbox_norm
+                        except Exception:
+                            chosen = bbox_px
+                    elif isinstance(bbox_px, (list, tuple)) and len(bbox_px) == 4:
+                        chosen = bbox_px
+                    elif isinstance(bbox_norm, (list, tuple)) and len(bbox_norm) == 4:
+                        chosen = bbox_norm
+                    elif isinstance(bbox_generic, (list, tuple)) and len(bbox_generic) == 4:
+                        chosen = bbox_generic
+                    else:
+                        # 直接字段 x1,y1,x2,y2
+                        if all(k in item for k in ("x1", "y1", "x2", "y2")):
+                            chosen = [item.get("x1"), item.get("y1"), item.get("x2"), item.get("y2")]
+
+                    if isinstance(chosen, (list, tuple)) and len(chosen) == 4:
+                        coords_map[key] = list(chosen)
+                    else:
+                        print(f"跳过无法解析的rect: {item}")
+            else:
+                # A2：简单键值对形式
+                for k, v in result.items():
+                    if isinstance(v, (list, tuple)) and len(v) == 4:
+                        coords_map[k] = list(v)
+
+        # 情形B：顶层是list
+        elif isinstance(result, list):
+            for i, item in enumerate(result):
+                if not isinstance(item, dict):
+                    continue
+                label = item.get("text") or item.get("label") or item.get("word") or f"rect_{i}"
+                bbox = item.get("bbox_px") or item.get("bbox_norm") or item.get("bbox")
+                bbox = dict_to_list(bbox)
+                if isinstance(bbox, (list, tuple)) and len(bbox) == 4:
+                    coords_map[label] = list(bbox)
+        else:
+            print("AI返回的JSON结构未知，无法解析。")
+    except Exception as e:
+        print(f"解析AI JSON时发生错误: {e}")
+
+    return coords_map
+
 def images_to_pdf(image_paths, output_pdf):
     os.makedirs(os.path.dirname(output_pdf), exist_ok=True)
     pil_images = [Image.open(p).convert('RGB') for p in image_paths]
@@ -179,6 +268,22 @@ def images_to_pdf(image_paths, output_pdf):
     first.save(output_pdf, save_all=True, append_images=rest)
     print(f"已生成PDF: {output_pdf}")
 
+text = f"""（仅归一化坐标，严格 JSON）
+你是一名版面定位助手。请在下图中定位并分别框出以下四个单词：AGN、UCLINK、LOGISITICS、LTD。
+
+坐标系与输出要求：
+
+- 图像尺寸：宽 {img_w} 像素，高 {img_h} 像素。
+- 原点位于图像左上角；x 向右增大，y 向下增大。
+- 为每个目标词返回它的最小外接矩形框，边界紧贴字形，不要添加额外边距。
+- 返回坐标为相对宽高的归一化浮点数，范围 [0,1]，保留 4 位小数；保证 0 ≤ x1 < x2 ≤ 1，0 ≤ y1 < y2 ≤ 1。
+- 禁止任何图片预处理（裁剪、缩放、加边距、重采样）；坐标必须对应原始图像。
+- 严格只输出下面的压缩的 JSON，不要附加解释或其他文本。
+- JSON中不要出现不在实例中的参数,例如bbox_2d
+输出 JSON 格式（示例为格式演示，实际数值请识别后填充）："""
+
+text += '[{"text":"AGN","bbox_norm":{"x1":0.0000,"y1":0.0000,"x2":0.0000,"y2":0.0000}},{"text":"UCLINK","bbox_norm":{"x1":0.0000,"y1":0.0000,"x2":0.0000,"y2":0.0000}},{"text":"LOGISITICS","bbox_norm":{"x1":0.0000,"y1":0.0000,"x2":0.0000,"y2":0.0000}},{"text":"LTD","bbox_norm":{"x1":0.0000,"y1":0.0000,"x2":0.0000,"y2":0.0000}}]'
+
 completion = client.chat.completions.create(
     model="qwen3-vl-plus", # 此处以qwen3-vl-plus为例，可按需更换模型名称。模型列表：https://help.aliyun.com/zh/model-studio/models
     messages=[
@@ -191,13 +296,15 @@ completion = client.chat.completions.create(
                         "url": image_base64
                     },
                 },
-                {"type": "text", "text": f"图像分辨率为{img_w}x{img_h}像素。坐标系定义：以原始图像左上角为原点(0,0)，x向右增加，y向下增加；不要使用任何预处理(缩放或加黑边)产生的坐标。请仅返回这两个文本的矩形框坐标，且必须是归一化到[0,1]的浮点数(相对于原始图像宽高)，返回格式严格为压缩JSON、无任何解释：{{\"AGN\": [x1_rel, y1_rel, x2_rel, y2_rel], \"UCLINK LOGISITICS LTD\": [x3_rel, y3_rel, x4_rel, y4_rel]}}。"},
+                {"type": "text", "text": text},
             ],
         },
     ],
+    temperature=0.1,
 )
 
 raw_text = completion.choices[0].message.content
+# raw_text = '```json[{"bbox_norm": {"x1": 0.1028, "y1": 0.1934, "x2": 0.1325, "y2": 0.2006}, "text": "AGN", "occurrence_index": 0},{"bbox_norm": {"x1": 0.1028, "y1": 0.2057, "x2": 0.1608, "y2": 0.2165}, "text": "UCLINK", "occurrence_index": 0},{"bbox_norm": {"x1": 0.1677, "y1": 0.2057, "x2": 0.2657, "y2": 0.2165}, "text": "LOGISITICS", "occurrence_index": 0},{"bbox_norm": {"x1": 0.2726, "y1": 0.2057, "x2": 0.3023, "y2": 0.2165}, "text": "LTD", "occurrence_index": 0}]```'
 print(raw_text)
 result = safe_extract_json(raw_text)
 if result is None or not isinstance(result, dict):
@@ -207,10 +314,16 @@ if result is None or not isinstance(result, dict):
 cleaned_dir = os.path.join("./output")
 cleaned_first = os.path.join(cleaned_dir, "cleaned_page_1.png")
 debug_first = os.path.join(cleaned_dir, "debug_page_1.png")
-draw_debug_boxes(image_paths[0], result, debug_first)
-erase_regions_on_image(image_paths[0], result, cleaned_first)
+coords_map = convert_ai_json_to_coords_map(result, img_w, img_h)
+if not coords_map:
+    raise RuntimeError("无法从AI返回中提取矩形框坐标，请检查输出格式或提示词。")
+print(f"解析并统一后的坐标字典: {coords_map}")
+draw_debug_boxes(image_paths[0], coords_map, debug_first)
+erase_regions_on_image(image_paths[0], coords_map, cleaned_first)
 
 # 合成PDF：第一页使用清理后的图片，其余页沿用原图
 final_images = [cleaned_first] + image_paths[1:]
 images_to_pdf(final_images, os.path.join(cleaned_dir, "cleaned.pdf"))
+end_time = time.time()
+print(f"耗时: {end_time - begin_time} 秒")