Multi-feature reconstruction network using crossed-mask restoration for unsupervised industrial anomaly detection

Unsupervised anomaly detection using only normal samples is crucial for industrial quality inspection. Although existing reconstruction-based methods have shown promising results, they suffer from two critical limitations: (1) poor distinguishable information in image reconstruction and (2) well-regenerated anomalies caused by model under-regularization. To overcome the above issues, we propose a novel Multi-feature Reconstruction Network leveraging Crossed-mask Restoration in this paper. Specifically, a multi-scale feature aggregator is developed to generate rich hierarchical representations from a frozen pre-trained model, providing discriminative features beyond raw pixels. Subsequently, a crossed-mask generator is adopted to randomly partitions the feature map into complementary regions covered by disjoint masks, followed by a restoration network based on the Transformer structure for high-quality repair of the missing regions, thus suppressing the model over-generalization. Finally, a hybrid loss is equipped to ensure comprehensive feature restoration and robust anomaly scoring. Extensive experiments show that our method is highly competitive with or significantly outperforms other state-of-the-arts on three public available datasets and one self-made dataset.