A Few-Shot Image Classification Algorithm Combining Graph Neural Network and Attention Mechanism

Jie Zhou, Jinfeng Qian, Qinghua Guo, Yong Zhao

Abstract: Deep learning has achieved success in various applications, depending on the abundance of training data. However, in practical applications, it is challenging to gather a substantial number of training samples. We employ few-shot learning algorithm to tackle this issue. Graph neural network can effectively capture the intra-class similarities and inter-class differences, which is highly beneficial for few-shot learning. This paper introduces an enhanced architecture based on graph neural network. Initially, we utilize a pre-trained residual network to extract features, which serve as the initial node features. And the similarity information between node features is defined as edge features. Subsequently, the graph neural network iteratively updates node features and edge features. Finally, we incorporate an attention mechanism into the metric network to compute the similarity of node features, which is used for classifying test samples. The attention mechanism can mitigate the interference from irrelevant information of neighbouring nodes, thereby improving the robustness and performance of the algorithm. Experiments demonstrate that our algorithm exhibits excellent performance in few-shot image classification.