Wavelet-SVDD: Anomaly Detection and Segmentation with Frequency Domain Attention

Abstract

Anomaly detection is a formidable challenge that entails the formulation of a model capable of detecting anomalous patterns in datasets, even when anomalous data points are absent. Traditional algorithms focused on learning knowledge regarding the typical features that arise in images, such as texture, shape, and color, to distinguish between normal and anomalous examples. However, there is untapped potential in frequency domain features for differentiating anomalous patterns, and current methodologies have not exhaustively exploited this avenue. In this work, we present an extension of the deep learning version of support vector data description (SVDD), a prevalent algorithm used for anomaly detection, through the introduction of Wavelet transformation and frequency domain attentions in the feature learning network. This extension allows for the consideration of frequency domain patterns in defect detection, and improves detection performance significantly. We performed extensive experiments on the MVTecAD dataset, and the results revealed that our approach attained advanced performance in both anomaly detection and segmentation localization, thereby confirming the efficacy of our proposed innovative designs.