Research On Construction And Application Of Lightweight Neural Network

With the development of information technology,digital images have penetrated into every aspect of our lives.Deep learning is a research topic that has attracted attention in the field of artificial intelligence in recent years.As a classic model of deep learning,Convolutional Neural Networks(CNN)has made a series of important breakthroughs in image classification,object detection and natural language processing.The convolutional neural network model has the characteristics of high computational complexity and large number of parameters.It is this complex multilayer network structure that provides the model with powerful feature representation capabilities.The popularity of smart devices has increased the need to transplant convolutional neural networks to embedded devices,but their huge parameter redundancy and computational costs limit their deployment in embedded devices,especially mobile devices.The demand for lightweight convolutional neural networks is increasing,so neural network compression algorithms have emerged.In this paper,by studying the existing convolutional neural network compression algorithm,focusing on the binarization method in the parameter quantization method,two kinds of binary convolutions that can effectively reduce the storage space of the model parameters while maintaining the model effect are proposed Neural network structure.The main contents of this article are as follows:(1)We propose a new type of compact portable deep learning network called Modified Binary Clique Net(MBClique Net),which aims to improve the portability of convolutional neural networks based on binary filters,while achieving Equivalent performance to high-precision convolutional neural networks such as Res Net.In MBClique Net,it is based on the modulation convolution module proposed in this paper.This module contains a special modulation filter and a binary convolution filter.This paper designs a special modulation operation.This operation uses modulation filtering.The generator and the binary convolution filter generate a reconstruction convolution kernel for performing the convolution operation,so as to compensate for the accuracy loss caused by the binary convolution filter.Compared with the full-precision model,MBClique Net can reduce the storage space required by the convolution filter by at least 32 times,and has better performance than other latest binarization models.More importantly,our model is even better than the high-precision models(such as Res Net)on the data set used.(2)The quaternion convolutional network has more advantages in processing color images than the ordinary convolutional network.To solve the problem that the quaternion convolutional network has twice as many parameters as the ordinary convolutional network,in this article we propose A quaternion local binary convolutional neural network(LBQCNN)based on a local binary neural network is presented.This paper proposes to replace the quaternion convolution operation with two quaternion convolution layers.The first layer uses a binary filter that does not need to update the parameters through learning,and the second layer uses a quad with a convolution kernel size of 1×1The convolutional layer of arity,proved the effectiveness of the network through image classification experiments,at the same time reduced the amount of parameters to be learned in the network by at least 6 times,and reduced the model storage space by at least 4 times.In addition,this paper uses the proposed quaternion local binary convolutional neural network to perform experiments on face recognition tasks,and achieves 97.93% accuracy on LFW data,which exceeds the original local binary neural network.