Study On Few-shot Learning Based On Deep Learning For Image Classification

The rapid development of Deep Learning technology has opened an unknown era of machine intelligence and is gradually and profoundly influencing and changing our life.However,Deep learning techniques rely on a large number of labeled sample data.In practical application,it is costly to obtain a large amount of labeled data,however,there is no huge data in some fields to feed to the deep learning model for training.When the small training data is available,it will lead to model overfitting.Therefore,few-shot learning technology is particularly important to solve the problem of insufficient training data.Researchers also pay attention to few-shot Image Classification as one of the basic tasks in the few-shot learning.In this thesis,two classification algorithms for few-shot image data is proposed under the existing framework of few-shot image classification algorithm.This thesis main contributions are as follows.1.Classify the existing few-shot image classification algorithms.According to different network modeling methods,few-shot image classification algorithms are divided into two categories: convolution neural network(CNN)model and graph neural network(GNN)model.Convolution neural network model is mainly based on CNN for modeling image data.Graph neural network model further applies CNN to graph neural network,the image data is modeled by the nodes and edges in the graph structure.In recent years,GNN algorithm has attracted more and more attention.2.A few-shot graph neural network classification algorithm based on cross attention is proposed.GNN is composed of nodes and edges,which can represent complex sample relationships and build a more powerful network model.The initial features of nodes are input by external modules,which has an important impact on the classification performance of graph networks.In order to improve the representativeness of initialization graph node features,cross attention is introduced to enhance the representativeness of target domain features.In order to make full use of the sample information of the query set in the process of building model,the similarity association of the support set sample features and the query set sample features is carried out.Through the information transfer on the graph,the correlation of the samples of intra-classes and the differences between the samples of inter-classes are enhanced,and the classification performance of GNN is improved.3.A few-shot image classification algorithm based on subspace is proposed.The samples are classified by distance measurement.In the past,the distance between paired samples or between samples and the class was measured.In this algorithm,cross attention mechanism is used to associate the support set sample features with the query set sample features,and then they are projected into the subspace respectively.Different subspaces represent different class.By measuring the distance between the sample to be predicted and the subspace,and increasing the distance between different subspaces to distinguish different class,the classification accuracy of the model is improved.To sum up,this thesis summarizes the current research of few-shot image classification algorithm,and proposes two classification algorithms suitable for few-shot image data.