Improved YOLOv10 Lightweight Bearing Surface Defect Detection Algorithm

Aiming at the problems of insufficient lightweight degree and incompatibility between accuracy and model lightweight of existing rolling bearing surface defect detection algorithms, based on the study of previous related algorithms, the YOLOv10 model algorithm was improved and the optimized model algorithm YOLO-ARE was obtained. The undersampling module ADown is innovatively introduced in Backbone and Head, which enables the model to capture the features of bearing defect images at a higher level, while reducing the amount of computation and optimizing the accuracy of target detection. In the backbone network, the fusion C2f-RVB-EMA module is inserted to optimize the model to extract the image defect features of the rolling bearing, which further improves the lightweight level of the model. The experiment was carried out on the self-built rolling bearing defect data set. Compared with the original YOLOv10n model, GFLOPs of the improved model decreased by 9.2%, Parameters decreased by 13.6%, model size decreased by 8.6%, accuracy P and mAP50 increased by 0.6% and 0.3% respectively. The detection speed is 158.1f/s, and the model can be arranged on the edge end to meet the requirements of industrial batch detection of bearing appearance defects.