Deep Metric Learning Method For Few-shot Image Classification

Traditional deep neural network(DNN)has excellent performance in image classification with a large number of samples.If DNN is directly applied to the field of few-shot image classification such as pedestrian re-identification,it is easy to produce overfitting and result in a significant decline in the performance of classification.Therefore,the study of few-shot image classification has important significance and value to effectively reduce the dependence of these fields on training samples.Metric learning based few-shot image classification has recently received much attention for simplicity and efficiency.Its performance depends highly on the feature extractor and classifier.In order to improve the performance of deep metric learning furtherly,according to the design of feature extractor and classifier,the work of this paper has the following two aspects:(1)Propose Res-SVDNet for few-shot image classificationAiming at the problems of insufficient feature representation and overfitting in current metric learning,this paper proposes Res-SVDNet for few-shot image classification,which can obtain good performance on relatively complex images classification tasks.Taking the advantages of good representation and anti-overfitting,the proposed Res-SVDNet utilizes ResNet architecture as the backbone to extract image features.As for the classifier,Euclidean distance is employed to meet the few-shot classification preference for simple inductive bias.Considering the requirement of Euclidean distance for orthogonality,singular value decomposition(SVD)is further involved between feature extraction and classification to improve the classification performance of Res-SVDNet.Experimental results on Mini-ImageNet and Tired-ImageNet demonstrate that the proposed Res-SVDNet outperforms the state-of-the-art methods such as prototypical networks by 2%?6%.(2)Propose Bi-TSNet for few-shot image classificationThis paper proposes Bi-TSNet based on the content(1),which make full use of the contextual information between few examples to improve the performance of classification.It has good performance on different few-shot image classification tasks.Bi-TSNet uses Task Encoder to obtain the contextual information of the support set to generate Feature-wise Linear Modulation(FiLM)parameters.The activation of the feature extractor ResNet18 intermediate map is linearly adjusted by FiLM to achieve the effect of suppressing overfitting by changing the feature distribution,thereby improving the performance of classification.Bi-TSNet uses Euclidean distance as classifier.Experimental results on Mini-ImageNet,Tired-ImageNet and CUB200₂011 demonstrate that the proposed Res-SVDNet outperforms the state-of-the-art methods such as prototypical networks by 12%?20%.