Normalizing Flow-Based Industrial Complex Background Anomaly Detection

Table 2

Anomaly detection performance of AUROC (%) on the BTAD dataset.

Categories	AE MSE	AE MSE + SSIM	VT-ADL [20]	CSA-Flow (ours)
0	49	53	99	99.61 (0.62%)
1	92	96	94	87.20 (−9.17%)
2	95	89	77	99.93 (5.19%)
Average	78.67	79.33	90.00	95.58 (6.20%)

Note: We compared CSA-Flow with convolutional autoencoders trained with MSE-loss and MSE + SSIM loss. Bold values denote the best result in the category.