Surface Defect Recognition Of Varistor Based On Deep Convolutional Neural Networks

The appearance defect classification technology is one of the key technologies for varistor quality inspection,which can greatly improve the efficiency and accuracy of product quality inspection.In order to identify the appearance defects of varistor body and pins more accurately,this thesis proposes a classification method of appearance defects based on deep convolutional neural networks(Convolutional Neural Networks,CNN).The main research content of this thesis is the classification of the appearance defects of varistor based on convolutional neural network.Around this subject,a detection method of appearance defects of varistor based on convolutional neural network is proposed.The method mainly includes the following work: First,based on the appearance characteristics of the varistor,based on Alex Net improvement,a convolutional neural network model(named CNN4VDR)that is sensitive to the minute defects of the varistor appearance is obtained.Based on the AlexNet network,this model combines the Global Average Pooling(GAP)layer and replaces the original 3 fully connected layers with a new convolutional layer and a global average pooling layer.A CNN4 VDR model with 6 convolutional layers,3 pooling layers and1 global average pooling layer is developed.At the same time,in order to prevent overfitting and improve the efficiency of training,the entire network is regularized in structure.Then,pre-process and segment the collected varistor image samples to make six classification data sets of appearance defects of varistor;input these image data sets with category labels into the proposed CNN4 VDR network for training,in the process Continue to fine-tune various parameters of the network until a high accuracy varistor appearance defect classification model is obtained;finally,1300 test samples in the dataset are used in the testing stage to test the trained classification model,and at the same time on Alex Net The same parameters and data sets were used for classification experiments under the network,and the experimental results of the two were compared to verify the effectiveness of the model proposed in this thesis.In this thesis,Mean Average Precision(mAP)is used to evaluate the performance of varistor appearance defect classification for the proposed model.Under the CNN4 VDR model,the mAP for the six classifications of the appearance defects of the varistor can reach 97.58%.Compared with the AlexNet model,it has achievedbetter results.The average detection time of one sample is about 17 ms.Experiments show that the CNN4 VDR model based on the proposed model can efficiently and accurately identify various appearance defects of the main body and pins of the varistor.