Aiming at the problem of low accuracy and many parameters of the existing bearing defect detection models

a lightweight bearing defect detection method based on YOLOv5s is proposed in this paper. Firstly

the improved EfficientViT-B0 is used to reconstruct the YOLOv5s feature extraction network

which can reduce the computational complexity of the model and extract deeper semantic feature information. Secondly

in order to solve the problem of difficult and easy sample imbalance

an F-CIoU is designed as a loss function to improve the positioning regression accuracy and robustness of the detection frame. Finally

a Dynamic Head (DyHead) based on multiple attention mechanisms is used to strengthen the feature semantic information and further optimize the classification and location of bearing surface damage. The experimental results show that map@0.5 of the improved YOLOv5s is 93.8%

the parameter amount and calculation amount are reduced by 42.3 % and 51.9 % respectively. The algorithm meets the lightweight requirements of industrial inspection deployment while maintaining high accuracy.