Research On Lightweight Image Classification Technology Based On Convolutional Neural Network

In recent years,convolutional neural networks have shown excellent performance in computer vision fields such as image recognition,target detection and semantic segmentation.During the development of deep convolutional networks,researchers have designed complex model structures in terms of network depth,width,and multi-paths to improve network performance,but also caused a series of problems such as excessive number of network parameters and computation,low efficiency,and increased hardware requirements,making their deployment on end devices with limited computation or low storage a great challenge.Therefore,it is important to design a lightweight network with fewer parameters and more suitable for small mobile devices while maintaining network performance.The main research content of this topic is as follows:(1)To address the problems that ShuffleNetV2,a lightweight convolutional neural network,still has redundant parameters and computational effort,slower response time and less accurate model classification,this paper proposes a more efficient and lightweight network ShuffleNeXt on the basis of ShuffleNetV2.In order to alleviate the intrinsic feature redundancy while reducing the number of network parameters,two modules,D-Shuffle and B-Shuffle,are designed to generate more expressive spatial features to further improve the network performance.The network is validated on four classical image classification datasets,CIFAR10,CIFAR100,Fashion MNIST and Flower102,respectively.The experimental results show that the ShuffleNeXt network can effectively reduce the model computation while achieving a better classification accuracy.(2)To address the problem that the ShuffleNeXt proposed above cannot effectively perform global feature characterization,a new combinatorial network,the ShuffleViT network,is proposed in this paper based on the ShuffleNeXt model.In order to realize the modeling between local and global information,this paper designs a new T-Shuffle module based on Transformer model,and introduces this module with D-Shuffle and B-Shuffle in a tandem combination into the ShuffleNeXt network,so that this network model can extract and fuse image features more fully.To validate the effectiveness of the proposed ShuffleViT network model,it is applied to a variety of classification datasets.The experimental results show that the ShuffleViT network exhibits superior classification performance compared to the traditional lightweight network model.