Abstract: When using neural network models such as deep residuals to classify images,some important features lost during feature extraction will affect the classification performance of the model.The black box problem brought about by the “end-to-end” learning mode of neural network can also limit its application and development in many fields.In addition,neural network models often require longer training time than traditional methods.In order to improve the classification effect and training efficiency of the deep residual networks,this paper introduces the model transfer method and soft thresholding method,proposes the deep residual soft thresholding network(DRSTN) network,and fine-tunes the network structure to generate different versions DRSTN network.The performance of the DRSTN networks benefit from the organic integration of three aspects:1)Visualize the feature extraction of the network through the gradients-weighted class activation mapping(Grad-CAM) method,and select further optimized ones based on the visualization results.2)Based on model transfer,researchers do not need to build a model from scratch,and can directly optimize the existing models,which can save a lot of training time.3) Soft thresholding,as a nonlinear transformation layer,is embedded into the deep residual network architecture to eliminate irrelevant features in samples.Experimental results show that under the same training conditions,the classification accuracy of the DRSTN_KS(3*3)_RB(2:2:2) network on the CIFAR-10 dataset is 15.5%,8.8% and 10.9% higher than that of SKNet-18,ResNet18 and ConvNeXt_tiny networks,respectively.The network also has a certain degree of generalization.It can achieve rapid transfer on MNIST and Fashion MNIST datasets,and the classification accuracy reaches 99.06% and 93.15% respectively.

Key words: Transfer learning, Residual network, Gradient weighted class activation mapping, Soft thresholding method, Image classification