A surface defect recognition network model YOLOv5s-CDOD was proposed to address the issues of high small target miss rate,slow detection speed,insufficient model accuracy and feature extraction generalization ability in multi-object situations during bearing surface defect detection.The B-ConvNeXt network was employed before the convolution operation to balance the accuracy and complexity of model,preserving the features of small target defects on the bearing surface while improving the generalization ability of model.The traditional convolution modules in the YOLOv5s network were replaced with convolution modules (DOD-Conv) with two deep over-parameterized convolutions,to enhance the recognition accuracy and speed of the model without increasing the model parameters.Finally,VariFocal Loss function was used in the feature processing stage to increase the learning of model for positive sample targets,further improving the detection accuracy of small target defects on the bearing.The experimental results show that compared to the original YOLOv5s network,the optimized network reduces the parameter count by 10%,significantly improving the detection speed;moreover,the average detection accuracy of the proposed model reaches 94%,enhancing the recognition rate of small target defects on the bearing surface.