Abstract:The ability to learn and generalize from a small number of samples is a primary distinction between artificial intelligence and humans.In the field of few shot learning,most graph neural networks focus on propagating labeled sample information to unlabeled query samples,while overlooking the crucial role of semantic features in the classification process.To address this,we propose a semantic feature propagation graph neural network,which embeds semantic features into the graph neural network to resolve the issue of low classification accuracy caused by fine grained image feature similarity.We then merge attention mechanisms with backbone networks to strengthen foreground and enhance feature extraction quality.By calculating class similarity using Mahalanobis distance,we achieve superior classification performance.Finally,we utilize the Funnel ReLU function as the activation function to further enhance classification accuracy.Experimental results on benchmark datasets demonstrate that our proposed algorithm improves accuracy by 9.03%,4.56%,and 4.15%,respectively,compared to the baseline algorithm on 5 class 1/2/5 sample tasks.