Reconstruction-Based Anomaly Detection with Noise-Induced Normal Sample Augmentation
Yong‐Wan Kwon, Dong‐Joong Kang · 2023
Anomaly detection has been an actively researched field, particularly in the context of industrial automation. In recent years, significant efforts have been made towards reconstruction -based approaches for anomaly detection. This method involves using a model trained on normal data patterns to generate reconstructions of given data, and then evaluating the difference between this input and the model's reconstruction to detect anomalies. However, in practice, models often struggle to control the generalization boundary or encounter ambiguity, resulting in inadequate separation between normal and abnormal instances or performance degradation due to overfitting. To address these challenges, this paper proposes a process that applies noise to normal images and reconstructs them into normal images. The proposed method learns patterns of normal images through a reconstruction -based network, allowing for the detection of anomalies by comparing input images with their reconstructed counterparts. Additionally, various normalization techniques are explored to improve the performance of the network. The effectiveness of our approach is demonstrated through superior performance on the MVTec AD benchmark, providing opportunities for early detection and intervention in potential anomaly situations in real-world applications.