Few-shot Image Classification Based On Critical Region Perception Graph Neural Network

The goal of few-shot learning is to reduce the dependence of the model on labeled data by training a model that can be rapidly generalized to new categories.In order to enhance the inference ability of the model,some few-shot learning methods construct a graph from query samples and support samples,use graph neural network to update the graph representation and propagate label information,then infer the category of the query samples according to the final graph representation.In this thesis,we propose our few-shot learning method based on graph neural network by analyzing the existing problems and make improvements.The main work of this thesis is as follows:(1)To solve the problem of category semantic ambiguity caused by image level labels,this thesis proposes a few-shot learning method based on semantic perceptual graph neural network.This method firstly refines the edge representation from a scalar which represent the global similarity of the adjacent nodes to a vector which represent the similarity of each pixel of the adjacent nodes.This method further introduces semantic alignment module to update the edge features,it makes the edge explicitly expressed semantic similarity between nodes by calculating the relationship matrix and transfer them into the edge value between pixels,and then make it tend to aggregate the semantic area in the process of node updates,the multilayer graph neural networks then propagate semantic information to improve the final classification result.Comparison and ablation experiments conducted on mini Image Net and tiered Image Net datasets prove the superiority of the proposed method,and visualization experiments demonstrate that the proposed method realize the perception of semantic region.(2)In order to reduce the ambiguity caused by background,this thesis proposes a few-shot learning method based on edge transformer graph neural network.Based on the edge refinement in(1),this method introduces transformer model to update the edge features in the graph.The updated edge features are obtained by dividing the difference feature maps between nodes into feature patch sequences and inputting them into transformer module.The idea of this method is to use the self-attention mechanism of transformer to help the model automatically focus on the key areas used to measure the similarity between nodes,so as to suppress the background information.Comparison and ablation experiments conducted on mini Image Net and tiered Image Net datasets demonstrate the effectiveness of the proposed method,and some classification results are visualized to show that the proposed method is robust to background interference.(3)In order to further improve the few-shot learning ability of the model,a self-supervised enhanced few-shot learning method is proposed in this thesis.This method designs a selfsupervised learning task based on location prediction,that is,to randomly cut out an image patch of the sample,and then predict the location of the image patch to be cut out.In this method,we train the self-supervised task and the few-shot classification task jointly to enhance the model's ability to extract effective features,and thereby improve the accuracy of classification task.The comparison and ablation experiments conducted on mini Image Net and tiered Image Net demonstrate the effectiveness of the proposed self-supervised auxiliary task.