基于多任务对比约束与可控缺陷生成的异常检测算法
|
|
刘宇轩,苏恩祈,井佩光,孙宏博,苏育挺
|
Anomaly detection algorithm based on multi-task contrastive constraints and controllable anomaly synthesis
|
|
Yuxuan LIU,Enqi SU,Peiguang JING,Hongbo SUN,Yuting SU
|
|
| 表 1 不同算法在MVTec数据集中的图像级异常检测结果 |
| Tab.1 Image-level anomaly detection results of different algorithms on MVTec dataset |
|
| 类别 | AUROCI/% | | PSVDD | RIAD | PaDiM | RD4AD | ATSN | NIGSF | 本研究 | | 地毯 | 92.9 | 84.2 | 99.8 | 98.9 | 99.8 | 96.5 | 100 | | 铁丝网 | 94.6 | 99.6 | 96.7 | 100 | 99.9 | 100 | 100 | | 皮革 | 90.9 | 100 | 100 | 100 | 100 | 100 | 100 | | 砖瓦 | 97.8 | 98.7 | 98.1 | 99.3 | 98.8 | 100 | 99.9 | | 木头 | 96.5 | 93.0 | 99.2 | 99.2 | 99.6 | 99.7 | 99.5 | | 瓶子 | 98.6 | 99.9 | 99.9 | 100 | 100 | 96.0 | 100 | | 电缆 | 90.3 | 81.9 | 92.7 | 95.0 | 100 | 85.6 | 99.2 | | 胶囊 | 76.7 | 88.4 | 91.3 | 96.3 | 96.8 | 93.9 | 97.5 | | 榛子 | 92.0 | 83.3 | 92.0 | 99.9 | 100 | 99.8 | 100 | | 金属 | 94.0 | 88.5 | 98.7 | 100 | 98.7 | 99.7 | 100 | | 药丸 | 86.1 | 83.8 | 93.3 | 96.6 | 96.7 | 97.2 | 98.2 | | 螺丝钉 | 81.3 | 84.5 | 85.8 | 97.0 | 96.8 | 86.6 | 98.6 | | 牙刷 | 100 | 100 | 96.1 | 99.5 | 93.1 | 100 | 100 | | 晶体管 | 91.5 | 90.9 | 97.4 | 96.7 | 98.3 | 95.3 | 98.8 | | 拉链 | 97.9 | 98.1 | 90.3 | 98.5 | 98.2 | 99.6 | 98.4 | | 平均值 | 92.1 | 91.7 | 95.5 | 98.5 | 98.4 | 96.7 | 99.3 |
|
|
|