Aiming at the low detection accuracy and low efficiency of traditional defect detection methods of metal surface

an improved algorithm for defect detection of metal surface based on YOLOv5 is proposed. The improved algorithm firstly introduces Ghost convolution to replace the standard convolution in the original backbone network

generating more feature mappings by fewer operations

realizing the lightweight of the model

and thus improving the detection efficiency. Meanwhile

the attention mechanism module is added to strengthen the channel information and weaken the redundant information to enhance the model's ability to extract the target feature information

thus achieving the improvement of detection accuracy. Using the DIoU-NMS to replace traditional NMS algorithms to improve the recognition of duplicate and occluded targets. Experimental results show that the average accuracy of the improved model reaches 76.2%

which is 2.7% higher than the average accuracy of the original YOLOv5 model. Comparing with the original YOLOv5 algorithm

the detection accuracy and efficiency are significantly improved

and the detection and identification of defects of metal surface can be performed quickly and accurately.