Research Of Few-Shot Image Recognition Based On Prototypical Networks

Few-shot image classification is a difficult problem in the field of computer vision and artificial intelligence.Few-shot learning is of great significance in practical application and a hot research topic in the field of machine learning.The deep learning model has achieved superior performance in image classification.But in practical application,data acquisition is difficult and the cost of data annotation is very large,which leads to the small amount of data with tags,and the effect of depth model classification in Few-shot sample scenes is not good.Many researchers have made some achievements in data enhancement,meta-learning,metric learning and so on,which promote the development of small sample learning.However,the current methods generally have complex network structure,do not consider the differences between samples,and can not make full use of the information provided by samples.This thesis starts with the problem of Few-shot image recognition based on Prototypical Networks,and studies on network structure and loss function.The main work are as follows:(1)A Prototypical Networks model based on attention mechanism for Few-shot image classification is proposed.Prototypical Networks has some problems in sample feature extraction: 1)because of the local limitation of convolution operation,it is difficult to transfer information between features which are relatively far away;2)it does not use the relationship between feature channels.In order to solve these problems,we use the attention mechanism of deep network to improve the structure of the prototype network,introduce the attention mechanism to improve the feature extraction ability of the model,and introduce the Shortcut to alleviate the gradient divergence problem of the model in training.The 5-way 1-shot and5-way 5-shot classification accuracy of the hybrid attention prototype network on the miniImagenet color image dataset are improved by 5.52% and 2.85% respectively.The N-way Meta-testing experiment also proves that the hybrid attention prototype network has better generalization in multi classification task.(2)A global dynamic scaling loss function for Few-shot image classification is proposed.The loss function of the existing Prototypical Networks only focuses on the distance within the class,so that the distance within the class is the minimum,but not the distance between the classes.To solve this problem,this thesis proposes a global dynamic scaling loss function,which takes into account the distance within the class and the difference between classes.The loss function is based on the loss function of Prototypical Networks,adding a weight term with global full sample view,which has variable parameters and gives attention to the global samples,so as to reduce the confusion between classes and further improve the accuracy of the model in Few-shot scene.In this thesis,the global dynamic scaling loss function is applied to the Prototypical Networks and hybrid attention Prototypical Networks model,and the experimental verification is carried out on Omniglot data set and miniImagenet data set.Experimental data show that the application of global dynamic scaling loss function improves the accuracy of model classification task.The results show that the application of global dynamic scaling loss function can accelerate model optimization and improve training efficiency.The N-way Meta-testing experiment also proves that the global dynamic scaling loss function has better generalization on multi classification tasks.This thesis has 27 figures,9 tables and 63 references.