Discriminative Feature Learning Framework With Gradient Preference for Anomaly Detection

Abstract

Unsupervised anomaly detection methods can detect product defects in industrial images by leveraging only normal samples during model training. Currently, the representation-based method, as a popular unsupervised anomaly detection method, has achieved impressive performance. Extracting valuable feature vectors that can remarkably improve the performance of anomaly detection is essential in unsupervised representation learning. To this end, we propose a novel discriminative feature learning framework with gradient preference for anomaly detection. Specifically, a gradient preference-based selector is designed to construct a representative feature repository, which alleviates the interference of the redundant feature vectors. To further make the feature vectors of normal samples adapt to the target distribution, a discriminative feature learning framework with center constraint is presented. By introducing our proposed framework, the distribution of normal samples becomes more compact. In the inference stage, the anomaly score calculated by the distance between feature vectors of test data and the normal cases is used to detect anomalies. Moreover, our method can be easily extended to anomaly localization. Extensive experiments on three popular industrial anomaly detection datasets demonstrate our proposed framework can achieve competitive results in both anomaly detection and localization. Additionally, the results on two medical anomaly detection datasets validate the generalization performance of our proposed method.