Key points are not available for this paper at this time.
Abstract In industrial production, the steel surface may incur different defects owing to the influence of external factors, thereby affecting the performance of steel. With the increasing requirements for steel quality, achieving efficient detection of steel surface defects is a difficult problem that urgently needs to be solved. Traditional steel surface defect detection methods are limited by poor detection performance and slow detection speed. Therefore, a model named LMS-YOLO, based on YOLOv8, is proposed in this paper for achieving efficient steel surface defect detection. Firstly, in backbone, the light weight multi-scale mixed convolution (LMSMC) module is designed to fuse with C2f to obtain C2fLMSMC, so as to extract the features of different scales for fusion and achieve the light weight of the network. Meanwhile, the proposed efficient global attention mechanism was added to backbone to enhance cross dimensional information interaction and feature extraction capabilities, and to achieve a more efficient attention mechanism. In neck, using channel tuning to achieve better cross scale fusion in BiFPN. Finally, the model uses three independent decoupled heads for regression and classification, and replaces CIoU with NWD as the regression loss to enhance the effect of detecting small scale defects. The experimental results showed that LMS-YOLO achieved 81. 1 mAP and 61. 3 FPS on NEU-DET, 80. 5 mAP and 61. 3 FPS on GC10-DET, respectively. The mAP increased by 2. 8 and 4. 7 compared to YOLOv8, and decreased by 17. 4% in floating point operations (GFLOPs) and 34. 2% in parameters (Params), which indicates that the model proposed in this paper has a better comprehensive performance compared with other methods in steel surface defect detection.
Xie et al. (Wed,) studied this question.