OTB-AAE: Semi-Supervised Anomaly Detection on Industrial Images Based on Adversarial Autoencoder With Output-Turn-Back Structure

Abstract

Due to the unbalanced proportion of positive (non-anomalous) and negative (anomalous) samples obtained from industrial data collection, the development prospect of supervised algorithms in industrial field is limited. Recently, Adversarial Autoencoder (AAE) has been used in the field of anomaly detection, and the complexity and unknown of negative samples increase the difficulty of this task. Here we propose an anomaly detection framework based on AAE to capture the normality distribution of high-dimensional images and identify abnormalities in industry. The Output-Turn-Back structure (OTB) is proposed to be added to the AAE structure to improve the discriminant capability of the discriminator. In particular, the L1 loss function and the Structural Similarity (SSIM) loss function are combined to assist the OTB. Then the Squeeze-and-Excitation (SE) model is embedded into AAE to improve the ability to capture the normality distribution of positive samples. The characteristic distribution of the negative and positive samples is different in testing, which leads to the significant reconstruction error of the latent space vector encoding, indicating anomaly. We test the model and compare it with the other models on different datasets, using the existing anomaly detection evaluation indexes to evaluate their performance. Experiments prove the superiority of the proposed method and its unique application prospect.