Research On Cross Shaft Surface Defect Detection Algorithm Based On Deep Learning

The cross shaft is a key component of the universal transmission device in the transmission system of the automobile chassis,and the surface of the cross shaft will produce various defects in the production process,so the detection of the surface defect of the cross shaft is of vital significance.Compared with traditional detection algorithms,target detection algorithms based on deep learning have higher detection accuracy and stronger adaptability.However,in the actual complex industrial environment,the surface defect detection based on deep learning is more difficult,and the detection efficiency becomes low due to the huge amount of parameters of the neural network.In view of the above problems,the main work of this paper is as follows:(1)For the target detection algorithm based on deep learning,the production of the data set is very important.Since there is currently no public dataset of cross-axis surface defects,the defect images in the production of cross-axis visual inspection are collected and classified to create a cross shaft surface defect dataset.YOLOv5 is selected as the baseline algorithm of this paper,and its network structure and sub-modules in the network,the experimental environment and evaluation indicators of the algorithm model are introduced in detail.The cross shaft surface defect dataset is trained on the YOLOv5 model,the m AP reaches 79.2%,and the inference time is 38.6ms.(2)In view of the small size of defect targets in the cross shaft surface defect data set,the K-Means algorithm is used to cluster new anchors,which accelerates the convergence of the loss function and increases the matching degree between the predicted bboxes and the real bboxes.The m AP of the model is increased from 79.2%to 81.4%;by adding the SENet channel attention mechanism to the YOLOv5 backbone network,the importance of the target features is improved while the useless features are suppressed,and the m AP of the model is increased by about 1%.(3)The structural re-parameterization in the convolutional neural network is introduced.Taking the Rep VGG module as an example,the calculation process of the structural re-parameterization is described in detail.The C3 module in the YOLOv5 backbone network and the neck network is replaced with the DBBm reparameterization module.In the inference stage,the parameter amount of the network is reduced by re-parameterization.The inference time is 63% of the original model,and the model volume is only the original volume 42%,reducing inference time while improving model accuracy.For industrial application scenarios,the improved model has higher detection accuracy,faster inference speed,and lower missed detection rate,which is more practical.