320 lines
12 KiB
Python
320 lines
12 KiB
Python
"""
|
||
图片分析服务 - 后台静默分析图片，用于数据标定
|
||
使用智谱 GLM-4V 视觉模型分析客户发来的图片
|
||
"""
|
||
import os
|
||
import asyncio
|
||
import base64
|
||
import time
|
||
import json
|
||
import logging
|
||
from typing import Optional, Dict, Any
|
||
from openai import AsyncOpenAI
|
||
from dotenv import load_dotenv
|
||
import aiohttp
|
||
from PIL import Image
|
||
from io import BytesIO
|
||
|
||
load_dotenv()
|
||
logger = logging.getLogger("cs_agent")
|
||
|
||
|
||
ANALYSIS_PROMPT = """你是一个电商图片处理评估专家。
|
||
请仔细分析这张图片，输出以下字段，每行一个，不要多余内容：
|
||
|
||
敏感内容: <yes|no>
|
||
平整度: <flat|mild|rough>
|
||
含文字: <yes|no|partial>
|
||
文字量: <none|少量|大量>
|
||
含人脸: <yes|no>
|
||
人脸清晰度: <无|清晰|模糊|遮挡>
|
||
阴影: <yes|no>
|
||
复杂度: <simple|normal|complex|hard>
|
||
原因: <15字以内，说明复杂度判断依据>
|
||
主体: <图片核心内容，如：印花图案/logo/人物/产品/老照片/风景/文字/其他>
|
||
类型: <处理类型，如：印花提取/高清修复/去背景/老照片修复/logo提取/人像修复/其他>
|
||
质量: <原图质量，如：清晰/轻微模糊/严重模糊/低分辨率/截图/扫描件>
|
||
分辨率评估: <足够|偏低|严重不足>
|
||
色彩: <单色|双色|多色|渐变|全彩>
|
||
主色调: <白/黑/红/蓝/绿/黄/棕/灰/多彩等>
|
||
细节层级: <简约|中等|精细|极精细>
|
||
边缘清晰度: <清晰|模糊|毛糙>
|
||
背景: <纯色|简单|复杂|透明>
|
||
可做: <yes|partial|no>
|
||
风险: <none|low|high>
|
||
透视: <no|mild|strong>
|
||
比例: <从以下选一个最合适的：1:1 / 9:16 / 16:9 / 3:4 / 4:3 / 3:2 / 2:3 / 5:4 / 4:5>
|
||
预估工时: <5分钟/15分钟/30分钟/1小时/2小时以上>
|
||
难点: <具体说明处理难点，如：细节多/透视矫正/文字提取等，15字内>
|
||
建议方案: <AI处理/人工精修/AI+人工/建议重拍>
|
||
提示词: <为 Gemini 写处理指令，中文，80字以内，要详细具体>
|
||
备注: <给客服AI的特别提示，没有则填无>
|
||
|
||
判断规则：
|
||
- 平整度 flat：画面平整、无褶皱、无透视 → 便宜
|
||
- 平整度 mild：轻微褶皱/透视 → 中等
|
||
- 平整度 rough：有褶皱/透视/曲面 → 贵
|
||
- 含文字 yes：有小字要精细保留
|
||
- 含人脸 yes：有人脸需处理 → 加价
|
||
- 敏感内容=yes 时，可做必须填 no
|
||
- 细节层级影响工时和价格
|
||
"""
|
||
|
||
|
||
class ImageAnalyzerService:
|
||
"""图片分析服务 - 后台静默运行，不影响主流程"""
|
||
|
||
_CACHE_TTL_SECONDS = 300
|
||
_analysis_cache: Dict[str, tuple] = {}
|
||
|
||
PRICE_MAP = {
|
||
"simple": (10, 15),
|
||
"normal": (15, 20),
|
||
"complex": (20, 25),
|
||
"hard": (25, 30),
|
||
}
|
||
|
||
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.vision_model = os.getenv("VISION_MODEL", "glm-4v-flash")
|
||
|
||
def _is_url(self, path: str) -> bool:
|
||
return path.startswith("http://") or path.startswith("https://")
|
||
|
||
async def _get_image_size(self, image_path: str) -> tuple:
|
||
"""获取图片尺寸"""
|
||
try:
|
||
if self._is_url(image_path):
|
||
timeout = aiohttp.ClientTimeout(total=10)
|
||
async with aiohttp.ClientSession(timeout=timeout) as session:
|
||
async with session.get(image_path) as resp:
|
||
if resp.status != 200:
|
||
return (0, 0)
|
||
data = await resp.read()
|
||
with Image.open(BytesIO(data)) as img:
|
||
return img.size
|
||
else:
|
||
with Image.open(image_path) as img:
|
||
return img.size
|
||
except Exception as e:
|
||
logger.debug(f"[ImageAnalyzer] 获取尺寸失败: {e}")
|
||
return (0, 0)
|
||
|
||
async def analyze(self, image_url: str) -> Dict[str, Any]:
|
||
"""
|
||
异步分析图片，返回结构化结果
|
||
|
||
Returns:
|
||
{
|
||
"url": 图片URL,
|
||
"complexity": simple|normal|complex|hard,
|
||
"subject": 主体描述,
|
||
"proc_type": 处理类型,
|
||
"quality": 质量评估,
|
||
"flatness": flat|mild|rough,
|
||
"has_text": yes|no,
|
||
"has_face": yes|no,
|
||
"has_shadow": yes|no,
|
||
"risk": none|low|high,
|
||
"feasibility": yes|partial|no,
|
||
"perspective": no|mild|strong,
|
||
"aspect_ratio": 比例,
|
||
"gemini_prompt": 处理提示词,
|
||
"note": 备注,
|
||
"price_suggest": 建议价格,
|
||
"width": 宽度,
|
||
"height": 高度,
|
||
"analyzed_at": 分析时间,
|
||
"success": True/False
|
||
}
|
||
"""
|
||
if not self.api_key:
|
||
return self._fallback(image_url, "未配置 API Key")
|
||
|
||
# 缓存检查
|
||
cache_key = image_url
|
||
now = time.monotonic()
|
||
cached = self._analysis_cache.get(cache_key)
|
||
if cached:
|
||
result, cached_at = cached
|
||
if now - cached_at < self._CACHE_TTL_SECONDS:
|
||
logger.debug(f"[ImageAnalyzer] 缓存命中: {image_url[:50]}...")
|
||
return dict(result)
|
||
else:
|
||
del self._analysis_cache[cache_key]
|
||
|
||
start = time.monotonic()
|
||
|
||
try:
|
||
client = AsyncOpenAI(base_url=self.base_url, api_key=self.api_key)
|
||
|
||
response = await asyncio.wait_for(
|
||
client.chat.completions.create(
|
||
model=self.vision_model,
|
||
messages=[{
|
||
"role": "user",
|
||
"content": [
|
||
{"type": "image_url", "image_url": {"url": image_url}},
|
||
{"type": "text", "text": ANALYSIS_PROMPT}
|
||
]
|
||
}],
|
||
max_tokens=500
|
||
),
|
||
timeout=30
|
||
)
|
||
|
||
content = response.choices[0].message.content
|
||
elapsed = time.monotonic() - start
|
||
|
||
result = self._parse_result(image_url, content)
|
||
result["elapsed"] = round(elapsed, 2)
|
||
|
||
# 获取尺寸
|
||
w, h = await self._get_image_size(image_url)
|
||
result["width"] = w
|
||
result["height"] = h
|
||
|
||
# 缓存
|
||
self._analysis_cache[cache_key] = (result, now)
|
||
|
||
# 详细日志
|
||
log_parts = [
|
||
f"主体={result.get('subject', '?')}",
|
||
f"类型={result.get('proc_type', '?')}",
|
||
f"复杂度={result.get('complexity', '?')}",
|
||
f"色彩={result.get('color', '?')}",
|
||
f"细节={result.get('detail_level', '?')}",
|
||
f"工时={result.get('est_time', '?')}",
|
||
f"方案={result.get('suggest_method', '?')}",
|
||
f"¥{result.get('price_suggest', 0)}",
|
||
]
|
||
logger.info(f"[ImageAnalyzer] 分析完成: {' | '.join(log_parts)}")
|
||
if result.get('difficulty'):
|
||
logger.info(f"[ImageAnalyzer] 难点: {result.get('difficulty')}")
|
||
if result.get('gemini_prompt'):
|
||
logger.info(f"[ImageAnalyzer] Gemini提示词: {result.get('gemini_prompt')}")
|
||
return result
|
||
|
||
except asyncio.TimeoutError:
|
||
logger.warning(f"[ImageAnalyzer] 分析超时: {image_url[:50]}...")
|
||
return self._fallback(image_url, "分析超时")
|
||
except Exception as e:
|
||
logger.warning(f"[ImageAnalyzer] 分析失败: {e}")
|
||
return self._fallback(image_url, str(e))
|
||
|
||
def _parse_result(self, url: str, content: str) -> Dict[str, Any]:
|
||
"""解析视觉模型返回的文本"""
|
||
import re
|
||
from datetime import datetime
|
||
|
||
def extract(key: str, default: str = "") -> str:
|
||
pattern = rf"{key}:\s*(.+)"
|
||
match = re.search(pattern, content, re.IGNORECASE)
|
||
return match.group(1).strip() if match else default
|
||
|
||
complexity = extract("复杂度", "normal").lower()
|
||
if complexity not in ("simple", "normal", "complex", "hard"):
|
||
complexity = "normal"
|
||
|
||
price_min, price_max = self.PRICE_MAP.get(complexity, (15, 20))
|
||
price_suggest = round((price_min + price_max) / 2 / 5) * 5
|
||
|
||
# 文字加价
|
||
has_text = extract("含文字", "no").lower()
|
||
if has_text in ("yes", "partial") and complexity in ("simple", "normal"):
|
||
price_suggest += 5
|
||
|
||
# 人脸加价
|
||
has_face = extract("含人脸", "no").lower()
|
||
if has_face == "yes":
|
||
price_suggest += 5
|
||
|
||
# 精细度加价
|
||
detail_level = extract("细节层级", "中等")
|
||
if detail_level == "极精细":
|
||
price_suggest += 10
|
||
elif detail_level == "精细":
|
||
price_suggest += 5
|
||
|
||
return {
|
||
"url": url,
|
||
"complexity": complexity,
|
||
"reason": extract("原因"),
|
||
"subject": extract("主体"),
|
||
"proc_type": extract("类型"),
|
||
"quality": extract("质量"),
|
||
"resolution": extract("分辨率评估", "足够"),
|
||
"flatness": extract("平整度", "mild").lower(),
|
||
"has_text": has_text,
|
||
"text_amount": extract("文字量", "none"),
|
||
"has_face": has_face,
|
||
"face_clarity": extract("人脸清晰度", "无"),
|
||
"has_shadow": extract("阴影", "no").lower(),
|
||
"color": extract("色彩", "全彩"),
|
||
"main_color": extract("主色调", ""),
|
||
"detail_level": detail_level,
|
||
"edge_clarity": extract("边缘清晰度", "清晰"),
|
||
"background": extract("背景", "简单"),
|
||
"risk": extract("风险", "none").lower(),
|
||
"feasibility": extract("可做", "yes").lower(),
|
||
"sensitive": extract("敏感内容", "no").lower(),
|
||
"perspective": extract("透视", "no").lower(),
|
||
"aspect_ratio": extract("比例", "1:1"),
|
||
"est_time": extract("预估工时", "15分钟"),
|
||
"difficulty": extract("难点", ""),
|
||
"suggest_method": extract("建议方案", "AI处理"),
|
||
"gemini_prompt": extract("提示词"),
|
||
"note": extract("备注"),
|
||
"price_min": price_min,
|
||
"price_max": price_max,
|
||
"price_suggest": price_suggest,
|
||
"analyzed_at": datetime.now().isoformat(),
|
||
"success": True
|
||
}
|
||
|
||
def _fallback(self, url: str, reason: str) -> Dict[str, Any]:
|
||
"""分析失败时的默认结果"""
|
||
from datetime import datetime
|
||
return {
|
||
"url": url,
|
||
"complexity": "normal",
|
||
"reason": reason,
|
||
"subject": "",
|
||
"proc_type": "",
|
||
"quality": "",
|
||
"resolution": "",
|
||
"flatness": "",
|
||
"has_text": "no",
|
||
"text_amount": "none",
|
||
"has_face": "no",
|
||
"face_clarity": "无",
|
||
"has_shadow": "no",
|
||
"color": "",
|
||
"main_color": "",
|
||
"detail_level": "中等",
|
||
"edge_clarity": "",
|
||
"background": "",
|
||
"risk": "none",
|
||
"feasibility": "yes",
|
||
"sensitive": "no",
|
||
"perspective": "no",
|
||
"aspect_ratio": "1:1",
|
||
"est_time": "",
|
||
"difficulty": "",
|
||
"suggest_method": "",
|
||
"gemini_prompt": "",
|
||
"note": "",
|
||
"price_min": 15,
|
||
"price_max": 20,
|
||
"price_suggest": 20,
|
||
"width": 0,
|
||
"height": 0,
|
||
"analyzed_at": datetime.now().isoformat(),
|
||
"success": False
|
||
}
|
||
|
||
|
||
# 全局实例
|
||
image_analyzer_service = ImageAnalyzerService()
|