基于异常特征对抗学习的工业图像异常检测方法

王天飞,周文俊,项圣,贺宇航,彭博

Industrial image anomaly detection method based on adversarial learning of abnormal features

Tianfei WANG,Wenjun ZHOU,Sheng XIANG,Yuhang HE,Bo PENG

表 3 不同方法像素级AU-PR结果对比

Tab.3 Pixel-level AU-PR comparison results for different models

类别 US[9] AE-SSIM RIAD PaDim CutPaste CLGAN MB-PFM ATSNM 本研究算法 bottle 74.2 — 76.4 73.0 77.9 79.6 76.7 78.7 86.6 capsule 25.9 — 38.2 33.4 32.3 69.6 46.2 52.7 72.8 grid 10.1 — 36.4 58.0 42.6 64.9 45.3 45.1 68.5 leather 40.9 — 49.1 45.2 54.6 75.4 46.8 57.4 76.3 pill 62.0 — 51.6 60.2 51.8 76.3 78.6 66.4 72.5 tile 65.3 — 52.6 51.7 67.2 87.6 80.3 89.1 95.6 transistor 27.1 — 39.2 71.3 70.8 67.7 56.8 70.3 77.5 zipper 36.1 — 63.4 16.6 68.5 68.7 55.6 72.6 86.3 cable 48.2 — 24.4 34.3 55.6 73.8 67.7 69.4 75.3 carpet 52.2 — 61.4 49.7 57.3 87.6 58.3 80.2 90.7 hazelnut 57.8 — 33.8 37.4 53.7 68.4 60.7 76.3 96.3 metalnut 83.5 — 64.3 39.4 62.5 67.7 78.1 77.6 75.6 screw 17.8 — 43.9 51.7 58.6 69.3 52.6 69.4 74.2 toothbrush 37.7 — 50.6 40.6 46.8 59.6 53.4 54.9 67.4 wood 53.3 — 38.2 42.3 79.3 77.3 46.7 78.7 76.8 平均 46.14 — 48.2 47.0 58.6 67.5 61.1 68.92 81.1