Unified Model Based on Reinforced Feature Reconstruction for Metro Track Anomaly Detection

Metro track anomaly detection can prevent accidents, thus avoiding severe life safety and property losses. Unsupervised methods that rely on one model per category or scene are unsuitable for complex and diverse track environments and unified detection, exhibiting poor stability. For most feature-based methods, the multistage features extracted by the pretrained model contain the redundant information and noise, which interferes the feature reconstruction and anomaly detection. Additionally, the presence of abnormal information in the reconstructed feature further degrades the performance of anomaly detection. To address the aforementioned issues, a unified anomaly detection model based on feature reconstruction, named reinforced feature reconstruction-based anomaly detection network (RFReconAD), is proposed. The proposed efficient channel feature reinforcement (ECFR) module cooperated with the designed loss function weakens the interference of redundant information and noise on feature reconstruction task. The layer-wise learnable queries embedded in the decoder alleviate the problem of anomaly reconstruction. Moreover, the proposed detection scheme achieves more accurate anomaly detection. Under unified training and inference, our method achieves 99.8% and 98.2% image-level AUROC, as well as 99.2% and 97.2% pixel-level AUROC, on the track foreign object detection (TFOD) dataset and MVTec-AD dataset, respectively; and its inference speed reaches 37 frames/s, outperforming the state-of-the-art methods.