Research On Few Shot Learning For Image Classification Based On Metric Learning

The model of deep learning is data-hungry.In many scenarios of practical application,such as medical imaging of rare diseases,there is existing problem of data sparsity for model training.Therefore,studying how to improve the model performance from few samples is very important for both theory and practice.Metric learning is one of the effective approaches for few shot learning.The quality of learned mapping function determines the performance of the method.This paper focus on the mapping function's quality of metric learning from the aspects of improving the loss function and network structure,which can enhance the generalization ability of the model.(1)In order to solve the problem of low quality of feature representation in metric learning algorithm,a metric learning algorithm based on contrast average distance(CMD-MLA)is proposed.The CMD-MLA algorithm introduces the idea of margin in support vector machine algorithm,and considers the similarity to the same category around the sample and the difference between different categories at the same time.so the few shot learning model considers more category information when updating parameters,and the feature embedding is more global.The CMD-MLA algorithm is a general framework which can be combined with different metric-based few shot learning models.This paper combines the CMD-MLA algorithm with prototype network(ProtNet)and conducts a series of comparative experiments on the image classification benchmark dataset Omniglot and MiniImageNet.The classification accuracy of the CDM-MLA algorithm in few shot learning tasks has been improved by an average of 0.3%and 0.7%respectively,CDM-MLA effectively promotes the generalization ability of few shot learning models.(2)In order to solve the problem of limited number of target domain samples in ProtNet for few shot learning,a model based on channel feature fusion of image feature map named SE-ProtNet is proposed.SE-ProtNet uses the attention mechanism to add the channel features,which is ignored by ProtNet,to category feature representation of the model.It can learn the importance of different channel features.SE-ProtNet extracts the useful features and suppresses the little effect features for the current task according to this important degree This paper conducts a series of comparative experiments on the image classification benchmark dataset Omniglot and MinilmageNet.Compared with ProtNet,the classification accuracy of SE-ProtNet few shot learning tasks has been improved by an average of 0.5%and 1.1%respectively,which effectively improves the quality of the model's feature embedding and enhances the robustness of the model.