Research On Image Classification Based On Ensemble Convolutional Neural Networks

Large-scale image classification has always been the focus of research in the field of computer vision,and is the key to achieving universal intelligence.In the face of massive image data,the traditional feature-based classification method seems to be powerless.With the rise of deep learning,especially convolutional neural networks,image classification performance has been greatly improved.At present,convolutional neural networks generally face problems such as high computational cost,easy gradient vanishing,and low classification efficiency,which makes the model unable to be further applied.In view of the above problems,this thesis proposes three improved convolutional neural network models from the perspective of ensemble,as follows:Firstly,for most models,the lack of robustness to the feature scale of the input image leads to the problem that the network needs more layers and parameters.To this end,a weight-sharing multi-stage multi-scale ensemble convolutional neural network(WSMSMSE-CNN)is proposed.First,the training samples which are pooled several times are sent into the weight-sharing multi-stage network to extract the features of the multi-scale input;Then,within each stage,multi-scale features of the same input are extracted by multi-layer multi-scale convolution;Finally,the two kinds of features are integrated,and network parameters are iteratively optimized.Therefore,the efficient feature extraction and learning of the input image by the network is achieved.Secondly,the gradient vanishing causes the deep network to be insufficiently trained.Besides,simply deepening the network has limited performance improvement.Aiming at these problems,a multipath ensemble convolutional neural network(ME-CNN)is proposed.First,the shallow and deep output features of the network are directly concatenated into multipath ensemble structure through shortcut connection.The multipath ensemble structure integrates multiple layers of features while avoiding the vanishing of gradients,thus enabling the construction of deep networks;Then,based on the multipath ensemble mode,the number of feature channels of the network is multiplied,and the improved channel-wise attention module is introduced to make the feature extraction more complete and efficient;Finally,iteratively optimizes network parameters and completes image classification tasks.Thirdly,for most models,the large amount of redundancy in the network structure and convolution operation leads to the problem of large calculation amount and parameter quantity.To this end,a multi-scale multipath ensemble convolutional neural network(MSME-CNN)is proposed.First,on the basis of multipath ensemble,the highrank convolution is replaced by multi-scale low-rank convolution,thus the network complexity is reduced while extracting various features;Then,the convolution kernel sparse connection mechanism is introduced to force the convolution output correlation to be reduced,thus the feature redundancy is removed and the network is compressed;Further,the input channels is linearly weighted by the linear sparse bottleneck structure.The linear sparse bottleneck structure not only reduces the number of channels,but also integrates different scale features separately and preserves input information to the greatest extent;Finally,optimizing network parameters by iteration and achieving light weight of the network.Experimental results on image datasets such as CIFAR and tiny ImageNet show that the proposed convolutional neural network models achieve high classification accuracy and network performance.