Lightweight Strip Steel Surface Defect Detection Algorithm Based on YOLOv8s

To address the issues of low accuracy, high number of parameters and computational cost in strip steel surface defect detection algorithms, a lightweight strip steel surface defect detection algorithm based on YOLOv8s is proposed. Initially, the Cross-stage Focused Linear Attention (C2F-FLA) module is introduced to enhance the feature extraction capability of the backbone network. Subsequently, to improve the capture of multi-scale defect contours and reduce the model’s parameters and computational complexity, a Wavelet Convolution-Bidirectional Weighted Feature Pyramid Network (WT-BiFPN) is employed in the model’s Neck section for adaptive fusion of features at different resolutions. Finally, Haar Wavelet Downsampling (HWD) is utilized to retain more defect-related information, further reducing the parameters and computational complexity. The proposed method was validated on the NEU-DET dataset, and the results demonstrate that compared to YOLOv8s, this approach achieves a 4.1% increase in mAP, a 44.4% reduction in parameters, and a 32.8% decrease in computational complexity. The method ensures lightweight implementation while improving detection accuracy, offering more possibilities for defect detection on edge terminal devices.