Lightweight Augmentation of SimpleNet for Industrial Anomaly Detection via Residual Spatial Attention

Abstract

Industrial anomaly detection requires both image-level recognition and pixel-level localization when anomaly samples are scarce. SimpleNet has become a representative lightweight baseline because of its single-stream inference pipeline, compact structure and low deployment cost. However, its multi-layer features lack an explicit spatial selection mechanism before patch extraction and fusion, and local anomaly evidence may be weakened in the front-end aggregation stage. To solve this problem, this paper introduces the residual spatial attention module in the front-end of layer2 and layer3 features without changing the SimpleNet discrimination target and the overall reasoning link, and pre-refines the local spatial response. Experiments on MVTec AD shows that this modification does not bring a unified gain on the overall average, but shows a clear category-specific performance pattern in terms of positioning indicators: P_AUROC has increased in 10 of the 15 categories, AU_PRO has increased in 9 categories, and the improvement in categories such as carpet, capsule, zipper and bottle is more obvious. The results show that the spatial refinement before fusion can enhance the local anomaly characterization ability of some categories at a lower structural cost and provide a reusable empirical basis for the front-end enhancement of lightweight discriminant industrial anomaly detection framework.