Context-Aware Adaptive Weighted Attention Network for Real-Time Surface Defect Segmentation

Surface defect detection is an important step in ensuring product quality in various manufacturing industries. Existing methods have achieved significant results, but there are still challenges, such as the lack of real-time detection speed, low contrast between defects and background, and insufficient handling of defect details. To address these issues, we introduce a lightweight and efficient method called context-aware adaptive weighted attention network (CAWANet) for real-time surface defect segmentation. To handle the computational resource constraints when dealing with deep features, we introduce context-aware adaptive weighted convolution (CAWAConv), which aims to extract deep features while suppressing noise interference. This allows the model to remain lightweight without compromising its ability to recognize subtle defect characteristics. In addition, during the feature fusion stage, we propose the feature detail improvement (FDI) module to capture more complex defect details. The FDI module enhances the representation of defect-related information, further optimizing the segmentation results. We conducted experimental evaluations of CAWANet on three surface defect detection datasets: the magnetic tile, NEU-Seg, and MSD. The experimental results indicate that our proposed CAWANet strikes a favorable balance between accuracy and inference speed compared to other state-of-the-art methods. Our code will be available at https://github.com/ZGWzzu/CAWANet.