Research On Image Classification Method Of Convolutional Neural Networks Based On Multi-Scale Pooling

The convolutional neural network adopts local receptive fields,shared weights and spatial downsampling technique.These make the network structure simple and has few training parameters.These also bring a certain degree invariance against distortion,translation and rotation.Compared with traditional image classification methods,convolutional neural networks do not need to extract specific artificial features for specific image datasets,Rather it is a hierarchically extracted feature that is autonomously end-to-end.Thus these networks have a higher recognition rate and wider utility.The emergence of convolutional neural networks have greatly stimulated the updating of computer vision and pattern recognition algorithm models,and also spurred researchers to study image classification and recognition algorithms.Traditional convolution neural network feature extraction is random.The same structure of convolutional neural network will extract different features which input the same data set.The difference is complementary which is helpful for mining the deep features of the image。Image categorization can be divided into coarse-grained image categorization and fine-grained image categorization;just as the name implies,the differences between classes are even smaller.Only subtle local differences can be used to distinguish different categories.Many unknown factors such as attitude,illumination,occlusion,background interference and so on have a greater interference with fine-grained image classification;The low rank recovery model ignores the local spatial correlation of the image which causes the sparse matrices are difficult to separate from similar low rank matrices.Based on the above problems,in this paper,on the basis of optimizing the pooling combination,introducing adaptive pooling into another sub-network to enrich the characteristics of differences,achieve multi-scale pooling and improve the level of feature expression.Measuring the complementarity of feature differences between sub-networks according to the complementary measurement function,so as to optimize the loss function which can fine tune the model weights by back propagation to improve the accuracy of image classification.In order to solve the problem of traditional convolutional neural network classification accuracy on fine-grained image dataset and background interference and background interference and relying on manually annotated information in fine-grained image classification,this paper introduces an improved saliency analysis model based on multi-scale pooling(the Tree structure induction norm and the Laplacian regularization term are introduced into the low rank recovery model,which is used to extract the structural information of the image,suppress the background interference).The experimental results on the MNIST and CIFAR-10 image sets show that the classification ability of multi-scale pooled convolutional neural networks is better than the existing deep convolutional neural networks.And after introducing the improved saliency analysis model,the algorithm achieved 86.83% and 82.71% classification accuracy on the fine-grained image datasets Cars and Stanford Dogs,which is superior to other comparison algorithms.The above experimental results show that this paper proposes a convolutional neural network classification algorithm based on multi-scale pooling and saliency analysis,which is better than traditional fine-grained image algorithm for classification accuracy without manual labeling information,and the proposed algorithm has strong robustness.