Research Of Metric-based Few-shot Image Classification

In recent years,with the development of deep learning,various advanced models and algorithms have enabled artificial intelligence to surpass humans in many fields.Despite this,the success of artificial intelligence largely depends on learning from large-scale data sets,and it is difficult to gain an advantage in situations where there are only a few data.Therefore,few-shot learning has become a research hotspot in recent years,aiming to cope with the situation of insufficient training data.This article focuses on the problem of few-shot image classification,and conducts research on metric-based few-shot image classification methods.The main contents are as follows:This article found that in the current commonly used few-shot image classification methods based on label propagation or feature propagation,the initial distribution of feature vectors has a great impact on the subsequent propagation results.Therefore,this paper proposes a subgraph-based feature propagation mechanism that distributes the input samples on several subgraphs,effectively avoiding excessive assimilation of labels and features,and reducing the bias between predicted labels and true labels.And based on this,a subgraph propagation network(SPN)is designed,which achieves better or similar classification results than popular few-shot image classification models on four commonly used few-shot image classification datasets.The results of the ablation experiment show that using subgraph propagation in several different settings on four mainstream datasets can improve the accuracy rate by 2%to 6%compared to not using subgraph propagation.Previous studies mostly mainly explored the feature similarity between samples,while the sample relationship gap between feature space and label space was often ignored.This paper proposes an information fusion mechanism based on attention mechanism for this situation,which fuses sample relationships in feature space and label space,thereby reducing sample inconsistency in feature space and label space Relationship design an information fusion network based on attention mechanism(RFPN),which also achieved good classification results.The results of the ablation experiments show that using information fusion under several different settings can improve the accuracy rate by 3%to%compared to not using information fusion on four mainstream datasets.