改进YOLOv8的轻量化轴承缺陷检测算法OA北大核心CSTPCD

In view of the low accuracy and the large number of model parameters of the surface defect detection algorithms for industrial bearings,a lightweight object detection algorithm(MFA-YOLOv8)combined with improved YOLOv8 is proposed.A lightweight multi-scale feature convolution module EMFC(efficient multi-scale feature convolution)is designed,based on which the Bottleneck structure in the partial C2f of the trunk and neck is reconstructed to maintain lightweight while also capturing the detailed features of the information with different scales effectively.The focal modulation(FM)module is introduced to enhance the model´s ability to characterize the defective objects and the receptive field of the model.The attentional scale sequence fusion(ASF)module is introduced to further improve the model´s detection accuracy of bearing defects and reduce the quantity of the parameters.The experimental results show that the mean average precision mAP@0.5 of MFA-YOLOv8 is as high as 91.5%on the GGS dataset,which is 2.4%higher than that of the YOLOv8,and its number of parameters decreases by 21.9%,so the MFA-YOLOv8 can satisfy the requirements of bearing appearance defect detection in industrial sites.