Improved YOLOv8-Based Defect Detection Model for Hot Rolled Strip Steel

Production defects in hot-rolled strip steel, arising from process design issues or variations in material properties, adversely affect both economic returns and production safety. Existing deep learning–based surface defect detection methods are often too slow and computationally heavy for real-time industrial applications. This paper proposes an optimized YOLOv8s algorithm for defect detection on hot-rolled steel strips. By integrating Cloformer and CBAM attention mechanisms, the model enhances its ability to capture spatial relationships, while the combination of partial convolution and depthwise separable convolution substantially reduces computational load. Furthermore, the introduction of SimSPPF accelerates inference. Experimental results show that the optimized YOLOv8s achieves a mean average precision (mAP) of 80.4% and an inference speed of 182.4 FPS, with a significant reduction in model size compared to conventional methods.