Research On Image Classification Algorithm Based On Improved ResNet

Image classification uses a complex network structure to extract features from the information in the image and ultimately classify it into different categories.Since the deep residual network Res Net was proposed and won the championship in the 2015 ILSVRC(Image Net Large Scale Visual Recognition Challenge)competition,many researchers have focused on the improvement of Res Net,resulting in a large number of excellent Residualstructured convolutional neural network.So far,the improvement of the deep residual network Res Net is mainly reflected in four aspects:(1)perform a split-transform-merge operation on the data of the input module;(2)introduce an attention mechanism into the residual structure;(3)Optimize the module topology;(4)optimize the overall residual network.The main research work of this paper is as follows:(1)The bottleneck structure of Res Net network is analyzed in detail,especially the source of redundant information in the network and the method of reducing redundant information,which prepares the theoretical basis for the improvement of Res Net network structure later.(2)For the basic feature extraction module of Res Net,starting from the redundancy of the Res Net bottleneck structure,this paper designs a new dimension-raising structure RC(Residual Concatenate)using residual structure and connection operation to improve the bottleneck structure.This structure can reduce the resource consumption of the bottleneck structure and enhance the gradient transfer,which can effectively improve the network efficiency and accuracy for classification tasks with small categories.(3)In view of the overall topology of Res Net,this paper designs a long-connection feature fusion model SCFF(Selective Cross-layer Feature Fusion)from the perspective of multi-scale feature fusion,referring to the pyramid model and the feature fusion algorithm based on the attention mechanism,the model further enhances the gradient transfer and allows the receptive field features of different levels of Res Net to directly participate in the classification decision with trainable weights,effectively improving the network classification accuracy with only a small additional computational overhead.This paper combines two improved structures with various residual networks,and conducts image classification experiments on multiple datasets.The experimental results verify the effectiveness of the improvement.