YOLO-KD: An Improved YOLOv8 Algorithm for Strip Steel Surface Defect Detection

To address the issues of low accuracy, high missed detection rates, and insufficient robustness in strip steel surface defect detection under complex industrial environments, this paper proposes an enhanced YOLOv8-based detection algorithm named YOLO-KD. Building upon the YOLOv8n baseline model, the algorithm improves model performance through three key modifications: feature extraction, feature fusion, and loss function optimization. Specifically, for feature extraction, the original C2f module is replaced with a C2f_KAN structure to enhance feature representation capability, significantly improving the model’s ability to extract defect features. In feature fusion, the conventional UpSample operator is innovatively substituted with the Dynamic Upsampling Operator (DySample), enabling more refined multi-scale feature fusion and effectively enhancing the model’s detection performance for minute defects in complex backgrounds. Finally, regarding optimization objectives, the CIoU loss function is replaced with InnerPIoU to further improve detection accuracy for defect locations on strip steel surfaces. Experimental results on the NEU-DET dataset demonstrate that the YOLO-KD algorithm outperforms the baseline model significantly, with the recall rate reaching 75.2% (an increase of 4.3 percentage points), mAP reaching 77.5% (a rise of 2.1 percentage points), and the F1 score rising to 75% (an improvement of 2.25 percentage points). Additionally, the FPS of the YOLO-KD algorithm proposed in this paper reaches 121.7. Compared with the industrial production line requirement of 30-100 FPS, it demonstrates significant performance advantages, meets the real-time requirements for strip steel surface defect detection, and highlights its important engineering application value.