Cross-domain Few-shot Image Classification Based On Domain Adaptation

Few-shot image classification specifically addresses the problem of insufficient training sample data to achieve fast learning and generalization of models from a small number of samples,which has important value and practical significance.In traditional few-shot image classification,the training set and test set samples have the same data distribution,however,there is a domain distribution offset between the training set and test set for cross-domain few-shot image classification,which seriously affects the classification accuracy.Domain adaptation can extract more robust domain invariant features to solve the domain distribution offset problem.To address the problems of domain distribution offset,finite labeled samples in the target domain and class nonoverlap between domains in cross-domain few-shot image classification,two crossdomain few-shot image classification methods based on domain adaptation are proposed in this paper.The main research content includes:1.Aiming at solving the problem of domain distribution deviation between source domain and target domain in cross-domain few-shot image classification,a crossdomain few-shot image classification method based on contrastive domain adaptation network is proposed by combining self-supervised contrastive learning and deep adversarial domain adaptation.Firstly,the feature extractor based on affine transformation is used to fully extract the features of the samples in the source domain to ensure that the few-shot image classification task in the source domain is fully matched.Then,according to the idea of adversarial domain adaptation,the parameters of feature extractor are updated by the adversarial relationship between feature extractor and domain discriminator to narrow the distribution of the two domains.Finally,data augmentation is carried out for the target domain samples and self-supervised learning auxiliary classification task is constructed.The contrastive loss function is constructed by using the unlabeled samples of the target domain to realize the clustering and interclass classification of the target domain samples,so as to improve the classification effect of the target domain.2.Aiming at the problem that there are few labeled samples in the target domain and unlabeled samples are not fully utilized,this paper combined the unsupervised clustering method to generate pseudo labels for the target domain samples,combined the domain adaptation with the target domain sample clustering method,and proposed a cross-domain few-shot image classification method based on the clustering domain adaptation network.In order to better extract the target domain sample information,firstly,the feature extractor is trained through the source domain prototype network,and the trained feature extractor parameters are obtained.Secondly,the trained feature extractor is used to generate pseudo-tags for the target domain samples,and the divergence between the feature distribution and the original distribution after the pseudo-tag clustering is calculated.Simultaneous adaptation of two antagonistic domains;Finally,the feature distribution of the target domain is learned by the selfsupervised method,the feature extractor and classifier are trained jointly,and the target domain samples are input into the trained classifier to get the prediction results.Experiments are carried out on Mini Image Net,CUB,Cars,Places,Plantae and other datasets.The experimental results show that compared with the classical few-shot image classification methods,the two few-shot image classification methods based on domain adaptation proposed in this paper can improve the performance of cross-domain few-shot image classification.There are 24 figures,18 tables and 110 references in this thesis.