Prototype-complemented few-shot image classification with intra-class and cross-class information
The metric-based few-shot learning models struggle to fully exploit the relationship between the intra-class samples and the cross-class ones,and treat a single sample feature as an independent item during training,which results in inaccurate prototypes and low-quality representation.To handle the issue,a prototype-complemented few-shot image classification with intra-class and cross-class information is proposed.Firstly,the features of the support set are fed to the intra-class information extraction branch to exploit the intra-class feature,which is further processed to obtain the information of the category description to complement the initial prototypes.Then,the query samples from different classes are fused to generate the new samples through the cross-class information extraction branch.The labels of the query samples are constructed as soft labels of the new samples.Finally,the complemented prototypes are employed to classify the query and new samples,and the model is optimized by classification loss.In this paper,the comparison experiments are performed on four public few-shot learning datasets.The accuracy is improved by 2.03％to 5.48％,2.25％to 8.55％,2.61％to 10.03％,and 5.10％to 8.82％on the MiniImageNet dataset,TieredImageNet dataset,CUB dataset and CIFAR-FS dataset,respectively.Experimental results show that the proposed model achieves superior generalization and classification performance than other methods.
few-shot learningmetric learningmeta-learningintra-class and cross-class informationprototype-complemented
万方数据
维普
