FC2P: Feature Cross-Channel Projection for Unsupervised Anomaly Segmentation

Abstract

Unsupervised anomaly segmentation plays a critical role in real-world industrial product quality inspection. While feature reconstruction-based methods have shown promising performance by detecting anomalies through differences between pretrained features and their reconstructions, existing approaches often suffer from shortcut learning, and leading to reconstruction failures and inaccurate anomaly representation across multistage features. To address these limitations, we propose feature cross-channel projection (FC2P), a novel approach for anomaly segmentation. FC2P divides features into two subsets based on neighboring channels and employs two autoencoders for closed-loop prediction, effectively mitigating shortcut effects while capturing semantic relationships for efficient reconstruction. In addition, we introduce an anomaly exposure network (AExNet), which progressively amplifies anomalies across multistage feature residuals, generating precise anomaly score maps for accurate segmentation. Extensive experiments on MVTec AD and Visa benchmark datasets demonstrate that the proposed FC2P achieves state-of-the-art (SOTA) performance, with average precision (AP) scores of 79.8% and 44.8%, respectively. Moreover, visualization results on real industrial data further show the practicality of our proposed method. The code will be made publicly available at https://github.com/Karma1628/work-2 to ensure reproducibility and facilitate further research.