Cascaded Semantic-Guided Network for Surface Anomaly Localization in Steel-Cased Batteries

Localization of surface anomalies in steel-cased batteries remains challenging due to subtle defect textures, nonuniform illumination, and complex surface reflections. Existing methods often struggle to separate fine-grained anomalies from normal patterns, leading to incomplete or inaccurate localization. To address these challenges, this paper presents the cascaded semantic-guided network (CSG-Net) framework that performs coarse-to-fine anomaly localization through cross-stage semantic refinement. Within the cascaded architecture, each stage refines anomaly localization using semantic priors from the preceding stage. In the coarse stage, a reconstruction–discrimination subnetwork models normal appearance distributions to generate preliminary anomaly cues. In the fine stage, a semantic-guided enhancement module (SGEM) transfers semantic priors from the coarse stage to adaptively modulate encoder features, thereby strengthening semantic consistency and enhancing discriminative capability. Furthermore, a deep supervision mechanism with weighted multi-scale loss is introduced to improve sensitivity to small defects and stabilize convergence. A complete steel-cased battery surface anomaly detection system is also developed and deployed on an industrial production line to validate engineering feasibility. Experiments on the steel-cased battery surface anomaly dataset (SCBSA), together with evaluations on MVTec AD and VisA, show that CSG-Net delivers strong localization performance across industrial inspection settings. Our code and dataset are available at https://github.com/yikuizhai/CSG-Net.