CSCP-YOLO: A Lightweight and Efficient Algorithm for Real-Time Steel Surface Defect Detection

Steel surface defect detection is critical for ensuring product quality, yet remains challenging due to multi-scale defects, small target prevalence, and complex background interference. To address these issues, a novel model named CSCP-YOLO is proposed. Specifically, the CGSF (C2f-GSConv-FasterNet) module is integrated into the feature extraction network to enhance the identification of small target defects on steel surfaces. To accelerate defect detection, the SPPF-DW module efficiently fuses multi-scale defect features through optimized computational pathways. The CCE (C2f-CIB-EMA) module strengthens feature extraction capabilities by optimizing weight sparsity distributions during model training. This architecture significantly improves feature map expressiveness and cross-channel information interaction, enhancing multi-scale and small-target perception. Finally, the Powerful-IoU loss function is employed to address optimization limitations in the original objective formulation. Extensive experiments on the NEU-DET dataset demonstrate that CSCP-YOLO achieves 79.1% mAP and 71.8% recall, outperforming YOLOv8s by 2.2% and 2.2%, respectively. Notably, through architectural simplification and quantization, the model size is reduced by 36.4% (from 22.5 MB to 14.3 MB) while achieving 186.2 FPS on an NVIDIA GeForce RTX3090, exceeding the 30 FPS threshold for real-time industrial inspection. This work provides a robust solution for high-speed, high-precision steel defect inspection, with potential applications in smart manufacturing.