Research On Automatic Detection Algorithm For Lightweight Wheel Hub Defects Based On Convolutional Neural Network

As the core of a fully automated wheel hub inspection system,wheel hub defect detection algorithms play a key role in the efficiency and accuracy of leave factory wheel inspection.In order to improve the efficiency and accuracy of wheel hub defect detection in enterprises,this paper discusses from two aspects of model lightweight and detection accuracy,which is based on the mainstream recognition network YOLOv4 and segmentation network U-Net.The following are the primary research contents:To improve the inspection efficiency and achieve real-time online inspection of wheel hub,this paper builds a defect recognition algorithm based on a lightweight YOLOv4 network.The algorithm uses the lightweight convolutional neural network Mobile Net V3 to replace the YOLOv4 backbone feature extraction network,and introduces the channel attention mechanism SE(Squeeze and Excitation)module and depth-separable convolution in the PANet(Path Aggregation Network)module,which greatly reduces the number of parameters in the original YOLOv4 network while ensuring the recognition accuracy.In order to compensate for the loss of recognition accuracy of the lightweight algorithm,this paper uses K-means++ to cluster the edges of the self-built defect dataset in order to further improve the recognition accuracy of internal wheel hub defects by combining the target characteristics and distribution features of wheel hub defects.Through the test of the self-built dataset,the class average accuracy of the improved algorithm m AP(mean Average Precision)is 87.46% and the test time is 0.075 s,which can realize real-time online detection of wheel hub defects.To extract the exact contour of the defect and calculate the defect area,this paper builds a defect segmentation algorithm based on a lightweight U-Net network.The algorithm transposes the bneck module from Mobile Net V3 to the encoding part of the U-Net network,deepening the network depth while reducing the network parameters,so as to improve the network fitting ability.To improve the accuracy of the network segmentation,a gated attention mechanism is introduced to the path where the encoder and decoder supplement the target feature information.Weights are also given to the feature maps during layer-by-layer sampling.Through the test of the self-built dataset,the average pixel accuracy of the improved algorithm m PA(mean Pixel Accuracy)is 91.57%,the test time is 0.067 s,and the segmented defect area error is able to meet the defect grading requirements.