Research And Application Of Lightweight Network In Image Classification

In the field of image classification,deep neural networks have achieved excellent performance,but the training process may require graphics processing units with basic computing power as support,and the training classification models often occupy a large amount of memory.With the rapid development of Internet of Things devices,the deep learning model demands to be deployed on small intelligent devices with limited computing resources.Therefore,the deployment that is high-performance and lightweight network models on embedded devices or mobile phone terminals has become the research focus.For the current deep learning solutions in image classification,most of them can not take into account the lightweight and accuracy of the model at the same time.A new convolutional neural network architecture is proposed.Firstly,a deep separable asymmetric convolution feature extraction module is designed,which is mainly used to reduce the computational complexity of the model.Then,using the matrix decomposition idea of the Kronecker product,a separable fully connected layer classifier module is designed,which is mainly used to reduce the number of model parameters.A network architecture that can be used for end-to-end training is finally obtained by weighing the depth and width of the network,called XSNet.To verify the effectiveness of the network architecture,MNIST,CIFAR-10,CIFAR-100 and SVHN datasets are used.On the CIFAR-10 dataset,the model parameters of XSNet11 are reduced by 96.71% compared to VGG19,and the model accuracy is only reduced by 0.23%.Compared with Res Net34,the designed 16-layer XSNet16 reduces the amount of model parameters by 64.37%,and the model accuracy increases by 3.24%.The experimental structure demonstrates that XSNet has a better performance than the current mainstream methods based on maintaining the performance of the model.The trained model has the advantages of lightweight and high classification performance,and XSNet makes it possible to deploy the model on small equipment with limited resources.The paper has 29 pictures,14 tables,and 55 references.