Exploiting Connection-Switching U-Net for Enhancing Surface Anomaly Detection

Surface defect detection in manufacturing is essen-tial but challenging due to data imbalance from the rarity of abnormal data. To address this, unsupervised anomaly detection (UAD) has been widely adopted, which trains reconstruction Neural Networks (NNs) with normal samples only. Recent studies have focused on reconstruction-by-inpainting approaches, evolving from multi-random masking to single deterministic masking strategies in order to maximize UAD performance without changing the NN structure. Single deterministic masking methods use 1) an U-Net structure to reduce false positives through precise normal pattern reconstruction and 2) an identity shortcut (IS) to avoid increasing false negatives by partially masking the input image. We investigate into further improvement of U-Net based UAD models without training or modifying their NN structure. To achieve performance improvement, we propose connection-Switching U-Net (CS-U-Net) that is trained with fully activated skip connections and then selectively deactivates some of the connections during inference to mitigate IS issues. This simple scheme effectively contains the generalization ability of deployed UAD models to reconstruct abnormal patterns into normal patterns. Using the KolektorSDD2 dataset, we have verified that our method achieves better throughput than state-of-the-art methods, suggesting its potential to be used as a widely adopted inference strategy for UAD where contained generalization ability is necessary.