Study On Few-shot Classification Method Based On Meta-learning And Graph Neural Network

Few-shot image classification aims to categorize samples accurately based on a limited amount of data when available data is scarce.The main idea of the few-shot classification based on meta-learning is to utilize similar enough learning tasks to be adaptive in a new task.The new task only includes few labeled data.The few-shot meta-learning methods based on graph neural network further improve the model performance by processing information in non-Euclidean space.Recently,the few-shot learning based on meta-learning method gradually becomes the hot issue.Based on mentioned above,our contribution can be summarized as following:(1)In order to alleviate overfitting in traditional deep neural networks in a certain extent when the amount of data is scarce,based on the meta-learning method combined with the attention mechanism,we propose a few-shot classification method called union attention dense-in-residual meta-transfer learning(UADR-MTL)in this thesis.UADR-MTL uses the attention mechanism to explicitly model the interdependencies intra-& inter-channels of the convolutional features to improve the feature extraction ability.By using the transfer learning method,UADR-MTL can avoid the classic meta-learning method can only use the shallow feature extraction network problem.Experiments verify that this method can increase the classification accuracy and the generalization of the model in a certain extent.(2)For further improving the classification performance and highlighting the specificity and correlation between samples,we use graph neural network to map samples to non-Euclidean space and obtain feature information,and iteratively update the feature information through meta-learning tasks.We propose a few-shot classification method named graph-derived graph neural network(GDGN)in this thesis.GDGN constructs a sample graph network(SG)and a derivatization graph network(DG).The derivatization graph network is generated by the sample graph network and used to update the sample graph.Experiments on multiple few-shot data sets show that this method can achieve high-performance classification results in image classification scenarios with few-shot data.