PatchCore 异常检测模型

基于 PatchCore (CVPR 2022) 的工业异常检测预训练模型，支持 MVTec AD 和 VisA 数据集。

模型信息

支持的类别

MVTec AD (15 类)

平均 AUROC: 0.942

VisA (12 类)

平均 AUROC: 0.886

使用方法

1. 安装依赖

pip install torch torchvision safetensors numpy scikit-learn tqdm Pillow

2. 下载模型

from modelscope import snapshot_download

model_dir = snapshot_download('Brilliantccc/PatchCore-Anomaly-Detection')

3. 推理示例

import torch
import json
from PIL import Image
import torchvision.transforms as transforms
from models.patchcore import PatchCore

# 加载模型
model = PatchCore.load('mvtec/bottle/patchcore_model.pth', device='cuda')

# 加载阈值
with open('mvtec/bottle/config.json') as f:
    threshold = json.load(f)['threshold']

# 预处理
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

image = transform(Image.open('test.jpg').convert('RGB')).unsqueeze(0).cuda()

# 推理
with torch.no_grad():
    scores, _ = model.compute_anomaly_scores(image)

print(f'异常分数: {scores.item():.4f}')
print(f'是否异常: {scores.item() > threshold}')

模型文件说明

mvtec/bottle/
├── config.json                # 配置（含阈值）
├── patchcore_model.pth        # 模型参数
└── memory_bank.safetensors    # 记忆库

引用

@inproceedings{roth2022towards,
  title={Towards total recall in industrial anomaly detection},
  author={Roth, Karsten and Pemula, Latha and Zepeda, Joaquin and Sch{\"o}lkopf, Bernhard and Brox, Thomas and Gehler, Peter},
  booktitle={CVPR},
  year={2022}
}