Research On Improvement Of Lightweight Convolutional Neural Network And Application Of Image Classification

With the development of intelligence in social life,Machine Learning has been applied to modern intelligent production.In Machine Learning,Deep Learning is an important part of it.Convolutional Neural Network(CNN)is widely used in the research of computer vision,and the performance of CNN determines the result of related tasks.As the accuracy requirements for CNN increase,the model becomes more and more complex,and the required hardware resources become more and more expensive.In recent years,the research on CNN has gradually shifted from building high-accuracy large-scale CNN to how to build more practical and efficient lightweight model architectures.The purpose of research on lightweight CNN is to greatly reduce the amount of calculations and parameters on the basis of ensuring that the accuracy of large-scale CNN is equivalent,so as to reduce the hardware resources required for model training and deployment.To solve the problems of large-scale CNN as large amount of parameters and computation,expensive hardware resources,limited feature extraction ability of existing lightweight CNN,a lightweight convolutional unit is designed to replace traditional convolution operations in this thesis.The convolution unit is designed based on the depthwise separable convolution,and on this basis,the strategy of feature reuse,channel compression and activation function are designed to improve it.Combined with the general principle of model design,the lightweight convolution unit is adjusted.The LCUNet is designed based on the network structure of large-scale CNN such as VGG and Res Net network structure,and carries out an image classification experiment on the CIFAR data set.The results show the LCUNet constructed in the thesis is better than the Mobile Net series and Shuffle Net series in classification accuracy,the required computational cost is less,and compared with the more advanced Ghost Net,the performance is equivalent in accuracy.In ablation experiments,the effectiveness of the feature reuse and channel compression strategy and activation function designed in this thesis is proved.