init: 商品图合规检测工具（豆包初筛 + DeepSeek 复检级联）

- src 标准布局：config/providers/pipeline/report + CLI/Tk GUI 双入口 - 级联省钱：豆包全量初筛，仅无违规/违规不明图进 DeepSeek 复检（含两票复核） - 输出：时间戳目录 + 分类文件夹图片归档 + Excel 报表 - 30 个单元测试（tests/，测试图片不入库）
2026-09-02 15:28:51 +08:00
commit 08d26710b4
20 changed files with 2513 additions and 0 deletions
@@ -0,0 +1,178 @@
+# -*- coding: utf-8 -*-
+"""视觉模型提供方：豆包（火山方舟）与 DeepSeek，均为 OpenAI 兼容 chat/completions。
+
+统一的返回结构：
+    {"file", "provider", "status"(ok/parse_fail/error), "category", "attr",
+     "logic", "raw", "usage": {"prompt", "completion", "cached"}, "finish_reason"}
+"""
+import asyncio
+import base64
+import json
+import os
+import re
+
+import aiohttp
+
+from .config import AppConfig
+
+MIME = {".png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg",
+        ".webp": "image/webp", ".gif": "image/gif", ".bmp": "image/bmp",
+        ".tiff": "image/tiff"}
+IMAGE_EXTS = set(MIME)
+
+# 17 类标准名称：按分类号归一化模型输出的写法差异
+CANON = {
+    1: "1. 烟草、毒品、赌博、犹太", 2: "2. 酒精", 3: "3. 色情", 4: "4. LGBT",
+    5: "5. 负向敏感类信息（血腥暴力）、战争（含动漫里的战争）、武器类、纳粹",
+    6: "6. 政治相关（包含军队组织等相关内容）", 7: "7. 种族歧视和文化差异",
+    8: "8. 旗帜和徽章", 9: "9. 宗教 - 其他教", 10: "10. 宗教 - 伊斯兰教",
+    11: "11. 宗教 - 基督教", 12: "12. 侵权 - 除人物外的其他侵权",
+    13: "13. 侵权 - 人物相关（如肖像权、姓名权等）",
+    14: "14. 脏话、侮辱性、谩骂攻击类文字图案（包含冒犯和争议性内容）",
+    15: "15. 图案消极低俗或者暴力（包含冒犯和争议性内容）",
+    16: "16. 欧区国家禁售", 17: "17. 冒犯和争议性内容",
+}
+
+
+def encode_image(path: str) -> str:
+    ext = os.path.splitext(path)[1].lower()
+    if ext not in MIME:
+        raise ValueError(f"不支持的图片格式: {path}")
+    with open(path, "rb") as f:
+        b64 = base64.b64encode(f.read()).decode("utf-8")
+    return f"data:{MIME[ext]};base64,{b64}"
+
+
+def norm_category(cat: str) -> str:
+    c = str(cat).strip()
+    m = re.match(r"^(\d+)", c)
+    return CANON.get(int(m.group(1)), c) if m else c
+
+
+def parse_judgment(raw: str):
+    """解析模型返回的三字段 JSON，返回 (attr, logic, category) 或 None。"""
+    if not raw:
+        return None
+    text = raw.strip()
+    m = re.search(r"```(?:json)?\s*(.*?)```", text, re.S)
+    if m:
+        text = m.group(1).strip()
+    s, e = text.find("{"), text.rfind("}")
+    if s == -1 or e <= s:
+        return None
+    try:
+        obj = json.loads(text[s:e + 1])
+    except json.JSONDecodeError:
+        return None
+    keys = ["文字/图形属性", "侵权/违规逻辑", "违规分类"]
+    if not all(k in obj for k in keys):
+        return None
+    return (str(obj["文字/图形属性"]).strip(), str(obj["侵权/违规逻辑"]).strip(),
+            str(obj["违规分类"]).strip())
+
+
+def _norm_usage(u: dict) -> dict:
+    u = u or {}
+    return {"prompt": u.get("prompt_tokens"), "completion": u.get("completion_tokens"),
+            "cached": (u.get("prompt_tokens_details") or {}).get("cached_tokens", 0)}
+
+
+def _result(fname, provider, content, data, path_for_err=""):
+    ch = (data.get("choices") or [{}])[0]
+    parsed = parse_judgment(content)
+    base = {"file": fname, "provider": provider, "finish_reason": ch.get("finish_reason"),
+            "usage": _norm_usage(data.get("usage"))}
+    if parsed:
+        attr, logic, cat = parsed
+        base.update({"status": "ok", "attr": attr, "logic": logic,
+                     "category": norm_category(cat), "raw": content})
+    else:
+        base.update({"status": "parse_fail", "attr": "", "logic": "",
+                     "category": "违规不明", "raw": content or ""})
+    return base
+
+
+async def detect_ark(session, sem, cfg: AppConfig, prompt: str, path: str, stop=None):
+    """豆包（火山方舟）视觉检测：图片在前、文本在后，强制 JSON 输出。"""
+    fname = os.path.basename(path)
+    url = cfg.ark_base_url.rstrip("/") + "/chat/completions"
+    headers = {"Authorization": f"Bearer {cfg.ark_api_key}",
+               "Content-Type": "application/json",
+               "x-is-encrypted": "true"}
+    payload = {
+        "model": cfg.ark_model,
+        "messages": [{
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": encode_image(path)}},
+                {"type": "text", "text": prompt},
+            ],
+        }],
+        "response_format": {"type": "json_object"},
+    }
+
+    last_err = ""
+    for attempt in range(1, cfg.retries + 1):
+        if stop and stop():
+            return {"file": fname, "provider": "doubao", "status": "error", "attr": "",
+                    "logic": "", "category": "检测异常", "raw": "用户停止", "usage": {}}
+        try:
+            async with sem:
+                async with session.post(url, json=payload, headers=headers) as r:
+                    if r.status == 200:
+                        data = await r.json()
+                        content = data["choices"][0]["message"]["content"]
+                        return _result(fname, "doubao", content, data)
+                    last_err = f"HTTP {r.status}: {(await r.text())[:300]}"
+        except Exception as e:  # noqa: BLE001
+            last_err = f"{type(e).__name__}: {e}"
+        if attempt < cfg.retries:
+            await asyncio.sleep(3 * attempt)
+    return {"file": fname, "provider": "doubao", "status": "error", "attr": "",
+            "logic": "", "category": "检测异常", "raw": last_err, "usage": {}}
+
+
+async def detect_deepseek(session, sem, cfg: AppConfig, prompt: str, path: str, stop=None):
+    """DeepSeek 视觉检测：temperature=0 + JSON 模式。
+    注意该模型为推理模型，reasoning 消耗 completion 预算，max_tokens 须留足。"""
+    fname = os.path.basename(path)
+    url = cfg.deepseek_base_url.rstrip("/") + "/chat/completions"
+    headers = {"Authorization": f"Bearer {cfg.deepseek_api_key}",
+               "Content-Type": "application/json"}
+    payload = {
+        "model": cfg.deepseek_model,
+        "messages": [{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": prompt},
+                {"type": "image_url", "image_url": {"url": encode_image(path)}},
+            ],
+        }],
+        "temperature": 0,
+        "response_format": {"type": "json_object"},
+        "max_tokens": cfg.max_tokens,
+    }
+
+    last_err = ""
+    for attempt in range(1, cfg.retries + 1):
+        if stop and stop():
+            return {"file": fname, "provider": "deepseek", "status": "error", "attr": "",
+                    "logic": "", "category": "检测异常", "raw": "用户停止", "usage": {}}
+        try:
+            async with sem:
+                async with session.post(url, json=payload, headers=headers) as r:
+                    if r.status == 200:
+                        data = await r.json()
+                        content = data["choices"][0]["message"]["content"]
+                        if content and content.strip():
+                            return _result(fname, "deepseek", content, data)
+                        # 推理烧尽预算会得到空正文，按可重试错误处理
+                        last_err = "empty content (reasoning exhausted max_tokens?)"
+                    else:
+                        last_err = f"HTTP {r.status}: {(await r.text())[:300]}"
+        except Exception as e:  # noqa: BLE001
+            last_err = f"{type(e).__name__}: {e}"
+        if attempt < cfg.retries:
+            await asyncio.sleep(3 * attempt)
+    return {"file": fname, "provider": "deepseek", "status": "error", "attr": "",
+            "logic": "", "category": "检测异常", "raw": last_err, "usage": {}}