Patch-Based Prototypical Cross-Scale Attention Network for Anomaly Detection

Tung-Lin Wang, Jun-Wei Hsieh, Yi-Kuan Hsieh

Abstract: Anomaly detection and localization play crucial roles in industrial manufacturing to help maintain product quality and minimize defects. However, anomalies are rare and challenging to collect, leading to imbalance data that cause a biased model to be trained and sensitive to noisy or irrelevant features. In addition, anomalies are often subtle, diverse, and change over time, making them difficult to differentiate, further complicating the detection and localization tasks. To address these challenges, we propose a new Patch-based Protopical Cross-Scale Attention Network (PPCA-Net) to effectively identify anomaly regions by learning residual features across different scales and sizes, distinguishing abnormal from normal patterns. It consists of two key components: the Scale-Aware Channel Attention Module (SACAM) and the Patch-based Cross-Scale Attention Module (PCSAM). These modules facilitate interactive feature inferences across multiple scales, significantly enhancing the ability to capture abnormal features of various sizes in various environments. Furthermore, we incorporate diverse anomaly generation strategies, including multi-scale prototypes to better represent feature disparities between abnormal and normal patterns, thereby enhancing overall effectiveness. Through extensive experimentation on the challenging MVTec AD [1] benchmark, PPCA-Net demonstrates superior performance in both unsupervised and supervised methods, highlighting its effectiveness in anomaly identification.