A Two-Stage Autoencoder For Visual Anomaly Detection

Deep convolutional autoencoder (DCAE) is usually optimized to minimize the difference between the input and the reconstruction, and the reconstruction error has been widely used as an indicator for visual anomaly detection. However, DCAE sometimes can reconstruct anomalies very well and thus may yield misdetections. To tackle this issue, we propose a novel non-symmetrical DCAE, which is trained in a two-stage manner. Specifically, a single RotNet is first trained to serve as encoder. Then, discriminative representations generated by the frozen encoder are used to train two parallel decoders for image reconstruction. Finally, the reconstruction errors obtained by the two decoders are combined as the anomaly score. Massive experiments on three public datasets and one practical industrial dataset demonstrate the superiority of the proposed method among existing reconstruction based methods.