Bi-Grid Reconstruction for Image Anomaly Detection

In the domain of image anomaly detection, significant progress has been made in unsupervised and self-supervised methods with datasets containing only normal samples. Although these methods perform well in general industrial anomaly detection scenarios, they often struggle with over- or under-detection when faced with fine-grained anomalies in products. In this paper, we propose GRAD: Bi-Grid Reconstruction for Image Anomaly Detection, which utilizes two continuous grids to detect anomalies from both normal and abnormal perspectives. In this work: 1) Grids serve as feature repositories to assist in the reconstruction task, achieving stronger generalization compared to discrete storage, while also helping to avoid the Identical Shortcut (IS) problem common in general reconstruction methods. 2) An additional grid storing abnormal features is introduced alongside the normal grid storing normal features, which refines the boundaries of normal features, thereby enhancing GRAD's detection performance for fine-grained defects. 3) The Feature Block Pasting (FBP) module is designed to synthesize a variety of anomalies at the feature level, enabling the rapid deployment of the abnormal grid. Additionally, benefiting from the powerful representation capabilities of grids, GRAD is suitable for a unified task setting, requiring only a single model to be trained for multiple classes. GRAD has been comprehensively tested on classic industrial datasets including MVTecAD, VisA, and the newest GoodsAD dataset, showing significant improvement over current state-of-the-art methods.