Research And Optimization Of MobileNet Compression Model

In recent years,the deep neural network has developed rapidly,especially in computer vision,constantly refreshing the performance of traditional models.Although the depth model has strong performance,there are numerous parameters.The huge number of parameters can improve performance and express the complexity of the model.At the same time,it also brings disadvantages such as huge network volume,slow operation,high computational cost,and large storage capacity.It is difficult to deploy on a limited hardware platform such as a mobile device.This is not conducive to embedding deep learning models into devices and applications with limited computing resources.Deep network compression is a key technology for solving such problems.Deep network compression is a key technology for solving such problems.This article belongs to the research direction of deep network compression and its specific work is as follows:Firstly,the paper introduces the basic structure of convolutional neural network,and then studies several classic convolutional neural network compression models,expounds their basic ideas,analyzes the compression methods they use in detail,and summarizes their respective characteristics.Secondly,this paper focuses on the specific structure of the lightweight neural network compression model MobileNet.According to its structural characteristics and parameter distribution structure,the regularization method and Dropout technology are adopted to improve the network structure of the MobileNet model,and an optimization scheme based on the MobileNet compression model is proposed.Finally,the article chooses to use the classic dataset CIFAR10 for experimentation and testing.After dividing the CIFAR10 data set into a training set and a test set,we trained and tested the improved MobileNet model on the TensorFlow deep learning framework.In this way,the optimization scheme proposed in this paper is achieved and the experimental results are compared and analyzed.The final experimental results show that the improved method proposed in this paper can effectively constrain the network parameters,reduce the coupling between neuronsand improve the accuracy of the model for image recognition,and finally have a good optimization effect on the original model.