Research On Feature Recalibration Densely Connected Convolutional Networks And Clothing Image Classification

With the rapid development of e-commerce,more and more consumers tend to buy goods online.At present,e-commerce platforms mainly retrieve clothing through keywords matching.These clothing are photographed by the seller and classified manually.As an important link in the clothing retrieval process,clothing classification directly affects the sales of goods and the consumer experience of users.Due to the large number of clothing images,clothing classification is a tedious and arduous task.The manual classification method is inefficient and has a high error rate.Therefore,clothing image classification based on computer vision has become an important research direction in the neighborhood of industrial clothing sorting and e-commerce.Aiming at the problems of existing convolutional neural networks in describing different types of information in images,weak generalization capabilities,and limited representation capabilities,this paper proposes a new feature recalibration classification algorithm based on convolutional neural networks and visual attention mechanisms.The main work is as follows:(1)In order to improve the ability of convolutional neural networks to describe different types of information in images,inspired by the visual attention mechanism,this paper designs two single-dimensional feature recalibration modules based on the channel attention mechanism and the spatial attention mechanism.The recalibration modules are channel feature recalibration module(CFRM)and spatial feature recalibration module(SFRM).Through the feature recalibration operation,CFRM can capture important information in the image in the channel dimension of the feature maps,and SFRM can enhance the relevant information in the image in the spatial dimension.In order to improve the image classification accuracy of the DenseNet,this paper designs two new Dense Block structures,and proposes channel feature recalibration DenseNet(CFR-DenseNet)and spatial feature recalibration DenseNet(SFR-DenseNet)two single-dimensional feature recalibration networks.Experiments on the CIFAR dataset show that the image classification accuracy of the CFR-DenseNet and SFR-DenseNet proposed in this paper is slightly higher than that of DenseNet,which proves that the single-dimensional feature recalibration module can improve the image classification accuracy of the DenseNet.(2)In order to further improve the image classification accuracy of the DenseNet,in view of the limited ability of the single-dimensional feature recalibration modules to improve the image classification accuracy of the DenseNet,this paper proposes two two-dimensional feature recalibration modules that can perform feature recalibration operations in the feature map channel and space at the same time,namely parallel channel and spatial feature recalibration module(PCSFRM)and tandem channel and spatial feature recalibration module(TCSFRM).By designing two new Dense Block structures,this paper proposes two two-dimensional feature recalibration networks of parallel channel-wise and spatial feature recalibration DenseNet(PCSFR-DenseNet)and tandem channel-wise and spatial feature recalibration DenseNet(TCSFR-DenseNet).The experimental results on the CIFAR dataset show that the image classification accuracy of the PCSFR-DenseNet and TCSFR-DenseNet is significantly higher than that of the DenseNet,CFR-DenseNet and SFR-DenseNet,which proves that the two-dimensional feature recalibration module can more significantly improve the image classification accuracy of DenseNet.(3)In order to verify the classification performance of the proposed models for clothing images and the generalization ability of the models,this paper conducted a lot of experiments on the clothing image dataset Fashion＿MNIST.The image classification accuracy of CFR-DenseNet,SFR-DenseNet,PCSFR-DenseNet and TCSFR-DenseNet are 95.52%,95.57%,95.63% and95.60%,respectively.Experimental results prove that the models proposed in this paper can accurately classify clothing images and has a strong generalization ability.