WARM: A wavelet adaptive restoration module for surface anomaly detection

Abstract

As a key component of lithium-ion batteries, surface defect detection of pole-piece is vital for quality control. Detection methods based on supervised have been used in pole-piece. However, the long cycle for data collection cannot meet the needs for fast construction and application of production line, making self-supervised a promising alternative. Although existing restoration-based methods have been widely studied, it’s still difficult to balance the problem of restoring normal regions and distinguishing abnormalities. We propose a Wavelet Adaptive Reconstruction Module by exploiting the obvious spatial and frequency differences between intra-class normal and abnormal features. Through training, the wavelet coefficients of the inputs are adaptively refined, potently obstructing the identity map from input image to restored image, thus improving the generalization of restoration network. The structure only consists of 2 simple Unets and WARM, it solves the anomaly detection task for pole-piece, and also outperforms other state-of-the-arts on MvTec AD.