Image Classification Based On Convolutional Neural Network

Image classification is one of the basic research topics in the field of computer visual recognition.The research goal is to predict the category label of input image by classification method to identify a given image and classification label set.In the whole classification process,feature extraction and selection are generally called feature expression,and good feature expression plays an important role in improving image classification accuracy.Therefore,for image classification tasks,deep learning has attracted wide attention and application with its powerful feature extraction ability.Convolutional neural network is one of the important models of deep learning,its weight sharing network architecture greatly reduces the network complexity and the number of parameters,thus avoiding the complex feature extraction and data reconstruction process in traditional recognition algorithms.However,there are still some problems to be solved,such as:(1)Gradient dispersion.The existing deep neural network usually adopts gradient descent for parameter training.However,as the number of network layer increases,the back propagation gradient gradually disappears during the training process,accuracy gets saturated and then degrades rapidly,resulting in degradation of network performance.(2)Network redundancy.An ultra-deep network can be constructed by simply stacking more network layers.However,as the structure scale increases,a large number of parameters are generated in the network,which increases the redundancy of the network,and then leads to network performance degradation.Aiming at the above problems,a more efficient deep convolutional network model is desgned to improve image classification accuracy.The main research contents are as follows:1.During the learning process of deep convolutional network,the initial values of convolution kernels are usually randomly assigned,which will cause the learning process to fall into local optimum.In addition,as the depth increases,the learning rule of network parameters based on gradient descent usually results in gradient vanishing phenomenon.Based on the above problems,a novel learning method for deep convolutional network based on deconvolution feature extraction is proposed.Firstly,an unsupervised two-layer stacked deconvolution network is used to learn feature mapping matrixes from the original images.Then,the learned feature mapping matrixes are used as the convolution kernels to convolve and pool with images in a layer-wise manner.Finally,the deep convolutional network is fine-tuned by the mini-batch stochastic gradient descent method with momentum coefficient,which can avoid the gradient vanishing problem.2.Deep convolutional network demonstrates that the classification accuracy can be remarkably improved by increasing the number of network layers.However,with the network depth increasing,the gradient will gradually disappear during the training process,and accuracy gets saturated and then degrades rapidly,which leads to network performance degradation.Based on the above problems,a novel learning method for deep convolutional network based on the pyramid structure is proposed.Firstly,as the number of layers increased,the feature map dimensions are gradually increased at each layer to distribute the burden concentrated at locations of structural units affected by down-sampling,such that all units are equally distributed.Then,by exploring the sequence between the stacked elements inside the structural unit,we design a pyramidal building block.Finally,gradient dispersion is further avoided by using the mini-batch stochastic gradient descent method with momentum coefficient for parameter training.3.Deep fused network can improve the training process of deep network,due to its capability of learning multi-scale representations and of optimizing information flow.However,the depth in a deep fused network does not contribute to the overall performance significantly.With the network depth increasing,accuracy gets saturated and then degrades rapidly.Based on the above problems,a deep convolutional network consisting of deep fused network and branch channel is proposed.Firstly,two base networks are combined in a concatenation and fusion manner to generate a deep fused network architecture.Then,an ensemble block with embedded learning mechanisms is formed to improve the feature representation power of model.Finally,the computational efficiency is improved by introducing group convolution without loss of performance.4.It is known that the classification accuracy of deep convolutional network can be remarkably improved by increasing the depth and width.However,as the network scale increases,the number of network parameters will increase significantly,which results in network redundancy and performance degradation.Based on the above problems,a highly modularized and light-weight deep interleaved fusion group convolution network is proposed.Firstly,a template block is constructed using the same network topology and the split-transform-concatenate strategy,and the deep network is formed by stacking template modules.Then,a small convolution kernel and group convolution are introduced into deep network to construct more efficient convolution kernel.Finally,the structured sparse convolution kernel and deep network are combined to form a highly modular and light-weight network architecture.Experiments on the image classification standard recognition tasks have shown that the proposed deep convolutional networks achieve better classification performance and has superior generalization ability.