Research On ResNet Image Classification Model Based On Tensor-synthetic Attention

There are insufficient feature extraction issues and ambiguous feature contributions of the residual network in image classification tasks.To address this problem,we propose a novel Res Net image classification model based on tensor-synthetic attention.Firstly,we utilize the Res Net-101 backbone to extract features and introduce the tensor-synthesis attention module after the residual convolution structure.The attention feature matrix is generated from three tensor product calculations on the acquired features;then,use the Softmax function to normalize the attention feature matrix and assign different weights to distinguish the feature contributions.In this way,the model focuses on the most promising features that are decisive factors for the classification results;After obtaining the weights,the weighted sum of the corresponding key values constructs the overall feature information of the image.The final image classification results from the average pooling and fully connected layer.The comparison experiments were carried out on the natural image datasets,CIFAR-10,CIFAR-100,and the street brand dataset,SVHN.The classification accuracy rates reached 96.12%,81.60%,and 96.67%,respectively,with the average test run-time of 0.0258 s,0.0260 s,0.0262 s.It is proved that the tensor synthesis attention module has good robustness capability when the image is disturbed by the Gaussian noise,random rotation,and center cropping.Also,compared with the original Res Net-18,Res Net-34,Res Net-50,and Res Net-101,adding the tensor-synthesis attention module to the residual network successfully increases the classification accuracy by 0.97%,2.47%,6.26%,and0.65%,respectively.The experimental results show that the proposed model obtains higher classification accuracy and shorter testing run-time than the state-of-the-art image classification methods,such as Improved GAN,CLS-GAN,NNCLR CCT-6/3x1,Res Net56 with re SGHMC,it can effectively enhance the feature learning ability of the network.Thesis has 36 pictures,14 tables,and 72 references...