LGFDR: local and global feature denoising reconstruction for unsupervised anomaly detection

Unsupervised anomaly detection is a challenging task in many visual inspection scenarios and has attracted significant attention. Anomalies are typically related to local low-level features or require global semantic information to be detected. However, most of the existing methods fail to strike a balance between local and global features and thus lack versatility and practicality. To address this issue, we propose local and global feature denoising reconstruction (LGFDR). The proposed method can implicitly learn the latent distribution of local and global features for normal images via a dual-tower reconstruction network. Next, a selective reconstruction head (SRH) is designed to adaptively fuse the information from local and global reconstructions. Moreover, adding noise to the features proves a simple and general operation that can further enhance the generalization of reconstruction networks. On the MVTec AD benchmark, LGFDR achieves 98.8% and 65.3% of pixel-level AUROC and AP for anomaly localization and 99.3% of image-level AUROC for anomaly detection, respectively. In addition, a real-world metal plate surface defect detection project is adopted to validate LGFDR. Both the public and the practical experimental results show the effectiveness of our proposed approach. The code will be available at https://github.com/Karma1628/work-1.