Research On Improving The Prediction Ability Of Deep Learning Classification Network Model Based On Bayesian Theor

With the continuous improvement of computer hardware,deep learning has made a breakthrough in various research fields.As one of the important branches,image classification task has become the hot topic in recent years.Image classification technology has changed from the artificial feature extraction methods to deep learning,which extracts features through a large number of samples.The classification of deep learning is based on the end-to-end method,which has achieved good results in recognition speed and accuracy.Generally,the performance of deep learning models is highly related to the captured features.To improve the accuracy of the classification model,we need to label a large number of training samples.However,in many practical application scenarios,it is difficult to obtain large-scale high-quality training data.Therefore,in these applications,how to improve the prediction ability of image classification model is still a problem,which should be solved.In fact,human being does not only classify objects according to their features,but also based on some other information.For example,when we know that the probability of a class of object in a certain environment is higher than that of other classes of objects,even if we can’t see the object,we can accurately classify the label of the object.In this paper,we call the distribution of labels in this environment as "local environment probability distribution",which is related to local environment.According to the above,this paper proposed a new framework based on Bayesian theory,which is named as "deep learning classification framework based on local environment label distribution".This framework can improve the classification accuracy and robustness.The details are as follows:(1)Firstly,the deep learning model is selected and trained on the training set.(2)Secondly,the posterior probability of each label is computed on the validation set,which is also called local environment label distribution in this paper.(3)Finally,on the test set,we run the model on each sample and output the probability of labels.Then,our framework combines it with the local environment label distribution to increase the accuracy.This paper selects three popular deep learning frameworks(Vov Net57、VGG16and Res Nest50)on three real data sets(CIFAR-100、CIFAR-10 and Mini-Image Net),the experimental results show that the proposed method achieves higher accuracy than the existing methods.The accuracy of Vov Net57,VGG16 and Res Nest50 is improved by 2.56%-13.06%,2.57%-18.51% and 2.26%-13.94% respectively,which proves that the proposed framework can be a choice to improve the performance of deep learning model.At the same time,it also compared the fusion methods with ours,where our methods achieved higher accuracy.