Research On Few-shot Learning Method Based On Deep Feature Mining

With the rise of artificial intelligence,deep learning is developing rapidly.However,traditional deep learning requires a large number of labeled samples for deep model training,which is costly.Therefore,how to make the deep models reduce the dependence on the number of training samples,namely few-shot learning,has become the key of the research.The few-shot learning methods based on metric learning are simple and efficient.They are mainly realized by mapping the samples into the embedding space and measuring the distances in it.In this paper,feature information extraction ability of the base metric method is improved by using the multi-scale feature extractor and the graph neural network,and then the class representation calculation of the traditional method is improved,so as to optimize the metric learning.The main research contents are as follows:1.Aiming at the problem that the traditional metric-based few-shot learning methods use single-scale feature extractors,which causes the feature information to be flat,a multiscale class feature network is proposed.Firstly,a multi-scale feature extractor based on convolutional blocks and residual blocks is designed to obtain multi-scale fusion features with rich abstract and detailed information.Then,the feature contribution degree is proposed.The sample weights are given to calculate the real representations,so as to optimize the representaions in the embedding space and improve the classification effect.2.In order to solve the problem of insufficient information mining caused by traditional metic learning methods only use the labels of samples themselves and ignore the label association between the samples,a deep mask GNN feature network is proposed to further mine the feature information of samples by using graph neural network after feature extraction.The deep mask GNN feature network adaptively generates the edge mask through the meta-learner and guides the graph to be updated selectively.In addition,the mutual exclusion loss is proposed in the classification stage.The sample clusters of different classes are driven away from each other by the optimized network,so as to further optimize the representations calculation and improve the classification performance of the method.