Research On Cross Domain Fine-Grained Few-Shot Classification Based On Local Descriptor And Feature Level Transformation

Cross domain few-shot learning is that after learning the model with generalization ability on the source domain with large samples,it can quickly learn and identify new categories in the new target domain with only a small number of samples.However,the current work is aimed at the cross domain few-shot classification of coarse-grained images.When used for cross domain few-shot learning of fine-grained images,it can’t accurately describe the rich local information of finegrained images,and the effect is not good.Because fine-grained images have large intra class changes and small inter class changes,how to accurately capture differentiated local information and how to improve the generalization ability of cross domain few-shot learning model in fine-grained image classification become very important.For the local region detail features in fine-grained images,this paper first studies a cross domain fine-grained few-shot classification method based on local descriptor,uses residual network and local descriptor to represent richer and more distinctive features in fine-grained images,and realizes few-shot classification in the target domain.The experimental results show that the feature representation of local descriptor has a good effect in cross domain fine-grained few-shot classification.Due to the difference of feature distribution between cross domain fine-grained images,this paper studies a cross domain fine-grained few-shot classification method based on feature transformation of local descriptor.Specifically,the features obtained from each residual block of the residual network are further integrated into a local descriptor representation,and then an improved feature transformation module is added after the local descriptor to simulate various distributions of image features in different domains and improve the generalization ability of the model in cross domain fine-grained image few-shot classification.In addition,this paper also studies a cross domain fine-grained few-shot classification of feature transformation layer fusion based on local descriptors,that is,multiple transformation layer fusion based on local descriptors.Specifically,the features obtained from each residual block of the residual network are divided into four branches,the features of each branch are integrated into local descriptor features,and then an improved feature transformation module is added after each local descriptor feature to obtain the transformed local descriptor features.Finally,the local descriptor features obtained from each branch are added and fused together to further improve the generalization ability in the new domain.