Self-supervised anomaly detection based on foreground enhancement and autoencoder reconstruction

Anomaly detection algorithms typically utilize generative models to reconstruct anomaly regions. Post-processing is used to pinpoint the anomalies. However, the paucity of real-world anomaly samples and the complex image backgrounds pose significant challenges for anomaly detection. The work innovatively proposed a self-supervised anomaly detection method. An efficient channel attention mechanism in the autoencoder was introduced to improve the reconstruction performance. Besides, a foreground enhancement strategy was designed to distinguish the foreground from the background by maximizing the inter-class variance. The strategy reduced the effect of background noises and simulated various anomalies that were rare in real samples. The MVTecAD and BTAD datasets were used to experiment with anomaly detection and location. Experimental results demonstrated that our method achieved higher AUC and AP scores at both the image level and pixel level compared to other advanced methods. In particular, the average AP score increased by 12.5% at the pixel level.