A Lightweight YOLOv7 Algorithm for Steel Surface Defect Detection

There are many irresistible factors in steel production that will lead to the presence of slightly different defects on the steel surface. Unlike conventional object detection methods, the detection accuracy, prediction speed and lightweight deployment of the model are important issues to achieve the defect detection task. Traditional object detection methods are usually unable to achieve a balanced result in all aspects. Based on the above, an improved YOLOv7 model is proposed, we integrate Partial Convolution (PConv) into the YOLOv7 network to ensure detection accuracy while reducing the model parameters. After this, we add the Simplified SPPF (SimSPPF) module to replace the SPPCSPC in YOLOv7 to improve the computational rate and accurate detection capability, and finally we replace the original loss function with the Normalized Wasserstein Distance (NWD) to achieve small-scale feature extraction and fusion, reducing the number of model parameters and balancing model detection accuracy and detection speed. The experimental results show that the average accuracy(mAP) of our model is 78.6%, and the number of parameters is reduced by 10.06 M compared with the original YOLOv7. Therefore, the proposed YOLOv7 model is able to accurately locate and categorize steel surface defects with good real-time performance.