Small Sample Image Classification Based On Deep Bayesian Network

Small sample learning is the main cognitive way for humans to understand the world.In the field of machine learning,small sample learning can effectively solve the problem of a sharp drop in network model performance due to insufficient training data for deep learning algorithms.Therefore,small sample learning is also one of the important research directions of artificial intelligence technology.In recent years,small sample learning has been widely used in medical diagnosis,military strategy,business decision-making and other fields,avoiding the consumption of a large number of resources on data sets,so that more energy can be devoted to the improvement of deep learning models.Small sample learning uses a small number of samples to effectively analyze new things based on the existing knowledge structure,so as to quickly acquire relevant knowledge of new things,and has good generalization ability.Therefore,learning with small samples has important application value and theoretical significance.This paper researches and improves the small sample image classification learning method based on deep learning,mainly using deep learning technology to effectively combine deep network model and Bayesian learning.Through the Bayesian learning idea,this paper uses the prior probability knowledge obtained from a small sample to infer the overall probability distribution.At the same time,in order to obtain effective prior probability knowledge,a deep network model based on transfer learning is used to obtain a better feature extraction effect.The main research contents of this paper are as follows:Firstly,this article gives an overview of image classification based on small sample learning,explains the importance of small sample image classification,and deeply studies the current mainstream small sample learning methods,especially transfer learning and Bayesian learning,which lays the theoretical basis for the work of this article.Secondly,based on the Res Net model from transfer learning,this paper effectively combines Bayesian classification and Soft Max classification to obtain a deep Bayesian network,the purpose of this method is to enable the model to effectively recognize known classes images in the training set and achieve good closed set classification.The above two classification methods will analyze the sample feature vectors separately to obtain different classification results.The network model uses feature thresholds to reasonably select the results to improve the learning effect of small samples,and finally experiments is designed for comparison and verification.Finally,this paper designs an open set classification method based on the Bayesian learning idea,the goal of this method is to make the model not only can effectively identify the known classes images in the training set,but also reject the unknown classes images that are not included in the training set.The method first performs feature distribution statistics on the samples in the training set based on Bayes’ theorem,and then the network model uses the feature distribution and the sample information to make causal derivation using Bayesian learning ideas,inferring whether the sample is an unknown category,and Finally,design experiments for comparison and verification.The experimental results show that the deep Bayesian network model performs more comprehensively on small sample data sets than other small sample learning methods,and has better classification effects.In addition,its classification performance of open set classification is also better than other existing methods,therefore,the open set classification algorithm also has certain advantages.