A high-precision segmentation network for industrial surface defect detection

Accurate surface defect detection is essential for improving product quality and reducing manufacturing costs, particularly in high-precision industries. However, existing deep learning methods struggle with multi-scale feature fusion and spatial information preservation. To address these challenges, this paper proposes the Spatial-Positional and Cross-Scale Fusion Network (SPCS-Net), an advanced segmentation network incorporating the Spatial-Positional Attention Module (SPAM) and the Cross-Scale Attention Fusion Module (CSAFM). SPAM integrates multi-scale convolutions, spatial attention mechanisms, and positional encoding to enhance the perception of defects with varying shapes. This design helps mitigate spatial information loss commonly observed in traditional U-Net models. CSAFM optimizes multi-scale feature fusion in the decoding stage by employing asymmetric convolutions and an adaptive feature weighting mechanism, effectively bridging the gap between high-level semantic information and low-level spatial details. Experimental results demonstrate that SPCS-Net outperforms state-of-the-art models on the NEU-Seg, MBP-Seg, and USB-Seg datasets, achieving F1 scores of 0.8854, 0.6543, and 0.7793, respectively, with Mean Intersection over Union scores of 0.7964, 0.5654, and 0.6674. Furthermore, SPCS-Net attains an inference speed of 104 f/s on MBP-Seg, striking a balance between accuracy and efficiency. These results highlight SPCS-Net as a promising solution for surface defect segmentation, with potential applications in automated quality inspection and material science.