Research On Lightweight Of Deep Learning And Its Application In Image Recognition

As a common information carrier in today's society,images are increasingly important.With the increasing use of images,image recognition algorithms have become a very popular research direction in the field of artificial intelligence.In recent years,relying on the excellent performance of convolutional neural networks in the field of image recognition,more and more convolutional neural network optimization methods and network model architectures have been proposed by researchers.However,as network performance improves,a dramatic increase in network complexity followed.Most advanced networks having millions of parameters and billions of computations.However,as the application of convolutional neural networks increases,more and more network models need to run efficiently on devices which have limited memory and computing capability.Therefore,the lightweight convolutional neural network has gradually attracted many researchers.This paper points at the lightweight convolutional neural network,designs a new network building module and trains a new lightweight convolutional neural network from scratch.Main tasks as follows:First of all,this paper sums up the research's background and significance of the lightweight convolutional neural network,indicating that the lightweight convolutional neural network is a hot and fast development direction.The research status at home and abroad has been detailed and the parts worthy of research and improvement are found out.Secondly,up to the current time point,this paper summarizes the basic theory and lightweight method of convolutional neural networks.After that,the classical lightweight convolutional neural network model proposed in recent years is introduced,and the advantages and disadvantages of the existing convolutional neural network lightweightingmethods are summarized.While reviewing the basic theory and the latest methods of lightweight convolutional neural networks,it provides the necessary theoretical basis for the reader to understand this paper.Then the paper combines and improves the group convolution and multi-branch structure,and proposes a new lightweight convolutional neural network building module—Slice Block,which can be used to build SFNet.This paper experiments on SFNet's tasks such as image classification and object detection.The experimentalizes results show that SFNet has better performance than existing lightweight convolutional neural networks on classification tasks while reducing network complexity,and can be applied to other computer vision tasks such as object detection.Afterwards,the paper combines the bottleneck structure and the group convolution to propose an embedded bottleneck structure.The structure effectively reduces the parameters and computations of the traditional bottleneck structure while maintaining the depth of the network.So that the network is free from the influence of vanishing gradient problem and easier to train.This paper combines the embedded bottleneck structure with the slice block to further reduce the complexity of the network.We builds the latest version of the embedded slice block and the latest version of the lightweight convolutional neural network SFNetV2 by using the embedded slice block.This paper has carried out detailed experiments on the proposed new structure on public datasets such as Cifar,SVHN and Pascal VOC.The experimental results indicate that the two new structures proposed in this paper are more efficient in terms of parameters and computations,and both have excellent performances.