init

image/image_qa.py

@@ -0,0 +1,189 @@
+"""
+图片处理结果质检模块
+
+处理完成后，用视觉 AI 对比原图和结果图，判断是否符合客户需求。
+评分 0-100，低于阈值则判定不合格，触发重试或人工跟进。
+"""
+import base64
+import os
+import time
+import asyncio
+from typing import Optional
+from dotenv import load_dotenv
+
+load_dotenv()
+
+_QA_PASS_SCORE = int(os.getenv("QA_PASS_SCORE", "70"))  # 合格分数线，默认70
+
+QA_PROMPT_TEMPLATE = """\
+你是一名专业的图片处理质检员，需要评估处理结果是否满足要求。
+
+【处理类型】{proc_type}
+【客户需求/Gemini提示词】{gemini_prompt}
+【原图描述】主体：{subject}，类型：{proc_type}，质量：{quality}
+
+请对比左图（原图）和右图（处理结果），从以下维度打分（每项0-25分）：
+
+1. 内容完整性：主体图案/内容是否完整保留，有无缺失、截断
+2. 畸变去除：褶皱/透视变形/背景是否已被清除
+3. 细节还原：颜色、线条、纹理等细节与原图的匹配程度
+4. 输出干净度：背景是否干净，有无多余内容、AI幻觉、模糊块
+
+输出格式（严格按照此格式，每行一个字段）：
+完整性: <0-25>
+畸变: <0-25>
+细节: <0-25>
+干净: <0-25>
+总分: <0-100>
+结论: <pass|fail>
+问题: <简述主要问题，不超过30字，无问题填"无">
+建议: <如果fail，给出重试改进建议，不超过40字，pass则填"无">
+"""
+
+
+class ImageQA:
+    """处理结果质检器"""
+
+    def __init__(self):
+        self.api_key  = os.getenv("OPENAI_API_KEY")
+        self.base_url = os.getenv("OPENAI_BASE_URL", "https://open.bigmodel.cn/api/paas/v4")
+        self.model    = os.getenv("VISION_MODEL", "glm-4v-flash")
+        self.pass_score = _QA_PASS_SCORE
+
+    def _to_base64(self, path: str) -> Optional[str]:
+        try:
+            with open(path, "rb") as f:
+                return base64.b64encode(f.read()).decode("utf-8")
+        except Exception as e:
+            print(f"[ImageQA] 读取图片失败 {path}: {e}")
+            return None
+
+    def _parse(self, text: str) -> dict:
+        def p(key):
+            for line in text.splitlines():
+                line = line.strip()
+                for k in [f"{key}:", f"{key}："]:
+                    if line.startswith(k):
+                        return line[len(k):].strip()
+            return ""
+
+        try:
+            score = int(p("总分"))
+        except ValueError:
+            score = 0
+
+        conclusion = p("结论").lower()
+        if conclusion not in ("pass", "fail"):
+            conclusion = "pass" if score >= self.pass_score else "fail"
+
+        return {
+            "score":       score,
+            "pass":        conclusion == "pass",
+            "issue":       p("问题"),
+            "suggestion":  p("建议"),
+            "detail": {
+                "completeness": p("完整性"),
+                "distortion":   p("畸变"),
+                "detail":       p("细节"),
+                "clean":        p("干净"),
+            },
+            "raw": text,
+        }
+
+    async def check(
+        self,
+        original_path: str,
+        result_path: str,
+        proc_type: str = "",
+        subject: str = "",
+        quality: str = "",
+        gemini_prompt: str = "",
+    ) -> dict:
+        """
+        质检处理结果。
+
+        Args:
+            original_path: 原图本地路径
+            result_path:   处理结果本地路径
+            proc_type:     处理类型（印花提取 / 高清修复等）
+            subject:       主体描述
+            quality:       原图质量
+            gemini_prompt: 传给 Gemini 的提示词（体现客户需求）
+
+        Returns:
+            {
+                "score": int,       # 0-100
+                "pass": bool,       # 是否合格
+                "issue": str,       # 主要问题
+                "suggestion": str,  # 重试改进建议
+                "detail": dict,     # 各维度分数
+            }
+        """
+        if not self.api_key:
+            print("[ImageQA] 未配置 API Key，跳过质检，默认通过")
+            return {"score": 80, "pass": True, "issue": "无", "suggestion": "无", "detail": {}}
+
+        orig_b64   = self._to_base64(original_path)
+        result_b64 = self._to_base64(result_path)
+        if not orig_b64 or not result_b64:
+            print("[ImageQA] 图片读取失败，跳过质检")
+            return {"score": 75, "pass": True, "issue": "质检图片读取失败", "suggestion": "无", "detail": {}}
+
+        prompt = QA_PROMPT_TEMPLATE.format(
+            proc_type=proc_type or "图片处理",
+            subject=subject or "未知",
+            quality=quality or "未知",
+            gemini_prompt=gemini_prompt or "按标准处理",
+        )
+
+        start = time.monotonic()
+        try:
+            from openai import AsyncOpenAI
+            client = AsyncOpenAI(base_url=self.base_url, api_key=self.api_key)
+
+            response = await client.responses.create(
+                model=self.model,
+                input=[
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "input_image",
+                                "image_url": f"data:image/jpeg;base64,{orig_b64}",
+                            },
+                            {
+                                "type": "input_image",
+                                "image_url": f"data:image/jpeg;base64,{result_b64}",
+                            },
+                            {
+                                "type": "input_text",
+                                "text": prompt,
+                            },
+                        ],
+                    }
+                ],
+            )
+            content = response.output_text
+            elapsed = time.monotonic() - start
+            result = self._parse(content)
+            result["elapsed"] = round(elapsed, 1)
+
+            status = "✓ 合格" if result["pass"] else "✗ 不合格"
+            print(f"[ImageQA] {status} | 得分: {result['score']}/100 | 问题: {result['issue']} | 耗时: {elapsed:.1f}s")
+            if not result["pass"]:
+                print(f"[ImageQA] 改进建议: {result['suggestion']}")
+            try:
+                from utils.api_cost_tracker import record
+                record("gemini_vision", count=1)
+            except Exception:
+                pass
+            return result
+
+        except Exception as e:
+            elapsed = time.monotonic() - start
+            print(f"[ImageQA] 质检失败 ({elapsed:.1f}s): {e}")
+            return {"score": 75, "pass": True, "issue": f"质检异常: {e}", "suggestion": "无", "detail": {}}
+
+
+# 全局实例
+image_qa = ImageQA()