Research On Bayesian Deep Learning Classification Framework For Multiple Dataset

Image classification is an important research direction in the field of computer vision.It is the basis of other computer vision tasks such as target detection and image segmentation.It has strong application value in daily life.Deep learning technology has strong feature expression ability in image classification,and has shown quite good classification performance.At present,many deep neural network models have been applied to image classification,and the performance of these neural network classification models is highly correlated with the characteristics of the captured training samples.When samples come from a variety of different fields,the diversity of features will increase the difficulty for the model to obtain correct features,thereby reducing the accuracy of model classification.If only the network structure of a single model is changed to improve the classification accuracy,the network structure of the model may become larger and deeper,and the problem of gradient explosion or disappearance will occur.Both feasibility and practicality are great challenges.Based on the above problems,this paper proposes a new framework based on Bayesian theory,which trains multiple models and integrates them with Bayesian methods to improve the accuracy of the final classification results.First,the framework selects some existing convolutional neural network models,trains these models on different data sets to obtain multiple training models,and realizes the use of multiple models instead of a single model to solve the classification task.In this process,in order to avoid the similarity between models,this paper proposes a model diversity training strategy.Finally,the Bayesian method is used recursively on the output results obtained by these models to obtain the final correct output.In order to prove the effectiveness of this framework,we have achieved good results in comparison with other methods on the CIFAR10 dataset and Euro STA dataset.This article also collected some pictures from real data sets to form a data set composed of multiple types of samples.The experimental results on this data set also show that the framework of this article can effectively improve the accuracy of classification and has stronger robustness.