Leveraging Foreground Collaboration and Augmentation for Industrial Anomaly Detection

Reconstruction-based methods, as one of the mainstream and advanced methods for anomaly detection, have attracted significant attention in the academic community. Although these methods may achieve good performance on some ideal industrial datasets, background factors have considerable influence on detecting anomalies due to a complex and ever-changing environment, resulting in overkill and false detection. In this work, we extend our previous implicit foreground-guided network (IFgNet) with a comprehensive consideration of the interference from complex backgrounds, and an incorporation of foreground constraints throughout the entire process. Thus, we propose a foreground collaboration and augmentation (ForeCA) network for anomaly detection, consisting of foreground homology augmentation (FHA) and foreground collaboration reconstruction (FCR). To be specific, FHA adopts a shuffled homology augmentation (SHA) strategy to synthesize pseudo-anomalous samples, as inputs of FCR, disrupting the original spatial structure of normal samples while preserving some structural relevance. Furthermore, FCR flexibly injects two sets of task-specific attention blocks into each convolutional block as task attention, integrating foreground detection with image reconstruction. We discriminate anomalies by the difference between the reconstructed images and the inputs and utilize the obtained foreground predictions to refine the coarse anomaly map. Extensive experiments on two challenging, widely used industrial anomaly detection datasets, including visual anomaly (VisA) and metal parts defect detection (MPDD), demonstrate our proposed method can achieve competitive results in both anomaly detection and localization. Our code is available at https://github.com/gloriacxl/ForeCA.