LASDN: A Lightweight Adaptive Segmentation-Decision Network for Industrial Surface Detection

Industrial surface defect detection continues to pose significant persistent challenges due to uneven illumination, complex textures, and multi-scale defect patterns. To address these issues, we propose Lightweight Adaptive Segmentation-Decision Network (LASDN), which is a lightweight two-stage framework that integrates a segmentation and a decision network. The architecture incorporates an Adaptive Attention Module (AAM) in both stages to enhance the model's focus on defect regions while suppressing background interference. A Lightweight Segmentation Enhancement Module (LSEM) is embedded at the input of the segmentation network to improve the representation of low-contrast, small-scale, and blurred-boundary defects. At the decision stage, an Optimized Decision Layer Module (ODLM) is designed to hierarchically fuse local and global features, thereby enhancing classification accuracy. Extensive experiments on four public industrial defect detection datasets demonstrate that the optimized model significantly outperforms baseline and state-of-the-art methods, achieving superior defect detection accuracy while maintaining computational efficiency.