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 5 在MVTec数据集上不同训练目标组合的性能表现

Tab.5 Performance of different training objective combinations on MVTec dataset

目标训练组合 AUROCI/% AUROCP/% AUPRO/% 可控异常生成策略 99.3 98.3 95.0 CutPaste 异常合成策略 98.9 98.0 94.6 基线 98.4 97.7 93.8 基线$ +{L}_{\text{S}} $ 99.0 98.1 94.5 基线$ +{L}_{\text{S}}+{L}_{\text{C}} $ 99.1 98.1 94.6 基线$ +{L}_{\text{S}}+{L}_{\text{C}}+{L}_{\text{R}} $ 99.3 98.2 94.8 基线$ +{L}_{\text{S}}+{L}_{\text{C}}+{L}_{\text{R}}+\text{AFRM} $ 99.3 98.3 95.0