Steel surface Defect Detection Algorithm Based on Improved YOLOv5s

To address challenges in steel surface defect detection, such as low accuracy and slow processing speed, an enhanced algorithm is proposed. The C3 module is replaced with GSConv (multi-channel shuffle convolution) to improve parameter efficiency and detection accuracy. The nearest neighbor up-sampling is substituted with CARAFE (a lightweight operator) to enhance the receptive field and utilize semantic information. GhostNetV2 replaces the CBS convolutional module for efficient feature extraction through stacking operations. Additionally, the Hard Swish activation function is incorporated to boost accuracy. Experimental results on the NEU-DET dataset demonstrate that the average accuracy of the improved YOLOv5s algorithm is 81.2%, a 1.7% increase compared to the original algorithm, the amount of computation is 17.2GFLOPs and the number of model parameters is 8.65M. The model effectively enhances detection accuracy.