Research And Application Of Convolutional Network Model Compression Method

As a representative technology in the field of deep learning,convolutional neural network has outstanding performance in computer vision tasks such as target detection and image classification and recognition.As the network becomes deeper and deeper,the model structure becomes more and more complex,and the number of parameters and the amount of computation in the network becomes more and more large,which makes the model difficult to be applied to the devices with poor computing and storage resources.Therefore,the research on the model compression technology of convolutional neural network becomes more and more important.Based on the idea of parametric pruning,the convolutional network model compression method is studied in this paper.The main research contents are as follows:A model compression algorithm based on mixed parameter pruning is proposed.The algorithm prunes the convolutional layer and the fully connected layer in the convolutional neural network according to different strategies.In the convolution layer,the mean value of the weight parameter of the convolution kernel of each trainable parameter is combined with the standard deviation,and the convolution kernel connection significance evaluation index is established,and the convolution kernel is pruned according to the index.In the fully connected layer,the weight amplitude of the corresponding node is selected according to the weight parameter of each node and the input feature fully connected,the weight amplitude is used as the evaluation index of the importance of the fully connected layer node,and the fully connected layer node pruning is performed according to the index.In this way,the convolutional neural network model can be compressed by mixing parameter pruning.Using this algorithm to experiment with LeNet-5 and AlexNet-like models on the MNIST data set,the results show that the algorithm can achieve 2.64 times the parameter compression effect on the LeNet-5 model without loss of accuracy,and reach 10.96 on the AlexNet-like model.This method effectively reduces the amount of parameters and calculations in the model,and achieves a better model compression effect.A traffic sign classification and recognition application based on the compressed convolutional neural network model is designed and implemented,the functional requirements of the application are analyzed,the application development is completed by using Qt Designer and PyQt5 toolkit,and the application testing is carried out.The application can select the data set for preprocessing,then train and save the model,and use the compressed network model saved after the training to classify and recognize the image,which reflects the engineering application value of the model compression algorithm studied in this paper.