Research On Deep Network Optimization Algorithm For Image Classification

Image classification can be divided into supervised classification,unsupervised classification and semi-supervised classification according to different sample data.This thesis first selects convolutional neural network as the research object from the perspective of supervised classification,and uses fireworks algorithm to optimize the model.Because the convolutional neural network uses gradient descent algorithm to optimize the parameters,which is easy to fall into the local optimum,and the firework algorithm has the characteristics of fast convergence and not easy to fall into the local optimum,using the firework algorithm to optimize the convolutional neural network has a certain improvement in the performance of the model.Supervised classification is done under the premise that the sample data is labeled,and most of the data in real life is unlabeled.Therefore,the classification accuracy of semi-supervised classification and unsupervised classification models is also particularly important,so this thesis improves the generated adversarial network from the perspective of optimizing the semi-supervised classification model.Since the generator of the original generated adversarial network uses random noise input,the randomness of the data is strong,and the quality of the generated pictures is low,the accuracy of the supervised classification is not high.The principal component analysis method can reduce the data dimension while retaining some of the characteristics of the original data,so this thesis uses the principal component analysis method to improve the input of the generator,and then use the improved network for semi-supervised classification.The main work done in this thesis is as follows:i.Research and implementation of improved convolutional neural network based on firework algorithm.First,compare the optimization effects of the firework algorithm,particle swarm optimization algorithm and genetic algorithm on the standard test function,and then use the firework algorithm to optimize the parameters of the convolutional neural network,and use the characteristics of the firework algorithm global search to avoid the parameter falling into the local optimal.The improved network is applied to image classification,and MNIST handwritten characters are used as sample data to test the classification effect of the model.The experimental results show that the improved network improves the classification accuracy of the model.ii.Research and implementation of a generative adversarial network improved based on principal component analysis.Combining principal component analysis with generative adversarial networks to improve the generative model of generative adversarial networks,through experiments on handwritten characters and face data sets,the model's generative effects were verified,then the second classifier of the discriminant model is changed to multi-classification,which is applied to semi-supervised image classification.Experiments are carried out on three different data sets.The classification accuracy and model stability of the improved network and the original network are compared.The experimental results verify the effectiveness of the principal component analysis method to improve the generated adversarial network.From the experimental results,this paper has achieved good results in the optimization of both networks.The convolutional neural network optimized by the fireworks algorithm has improved the classification accuracy of images,and the improved adversarial network generated by the improved principal component analysis method can improve The quality of the generated images can also improve the accuracy of semisupervised classification.