DN-DR: Discriminative Network with Dual Reconstruction for Image Anomaly Detection

Abstract

One mainstream of image anomaly detection is based on reconstruction. Such methods still struggle with diverse anomalies, such as near-in-distribution or deformed types. To address the challenge, we propose a Discriminative Network with Dual Reconstruction (DN-DR), consisting of a Memory Reconstructor, a Corrector, and a Discriminator. DN-DR aims to better restore the defective image to its normal state through dual reconstruction, thereby obtaining superior Discriminator performance. Specifically, (1) the Memory Reconstructor is based on training multi-scale codebooks from normal images to rebuild unknown regions in the test images, also named preliminary reconstruction; (2) the Corrector, as a subsequent reconstruction module, addresses false anomalies caused by the patch-level replacement strategy in the Memory Reconstructor, achieving a final refined reconstruction; (3) a U-Net Discriminator follows. Experiments on the challenging MVTec AD dataset demonstrate excellent reconstruction performance and anomaly inspection, including defects of near-in-distribution or deformed types.