Research On Improved Image Classification Methods Based On Convolution Neural Networks

Deep learning was described firstly in 2006.Its power has demonstrated in many areas in recent years.Convolution neural network is a typical deep learning model.The performance of convolutional neural networkhas surpassed traditional machine learning algorithms in computer vision problem,such as object detection,scene classification,object recognition,face recognition and so on.As the convolutional neural networks are widely used in more and more areas,how to improve the network performance in practical application becomes more and more important.In this paper,the basic convolution neural network and the general steps of training it are summarized.This paper introduce the the existing methods to imporve the performance of networks for every step during training,and the factors that affect the final performance of networks.Finally,this paper propose several methods to impove the perfomace based on original improvement methods.Different from simply modifying the network structure,the methods proposed in this are more reasonable in image classification.The main improvements in this paper can be summarized as follows:1.In this paper,we propose a new convolution kernel,named as complex convolutionkernel.Complex convolution kernel will be transformed correspondingly in theimage when the features of the object are transformed.Compared with the originalconvolution kernel,the complex convolution kernel can extract the correspondingfeatures with different deformations effectively.2.Taking effect on the pre-training phase,the denoise constraint based on restrictedboltzmann machine can make the distributions of the learned features more regular.3.The framework of ensemble feature learning based on the pretrained base networkscan make use of the advantages of ensemble learning to improve the performanceof the network without modifying the network's structure.The methods proposed in this paper are more general and can be more specific to guide how to improve the performance of the networks.The improvements of the proposed methods are proved by the comparison experiment and the suggestions of how to imporve performance in practical applications can be found in this paper.