Research On Few Shot Object Recognition Of Remote Sensing Image Based On Res-MSRN

Remote sensing image recognition technology has important roles in both military and civil fields.While,traditional gradient-base remote sensing image recognition networks often through extensive labeled examples and iterative training,which brings inconvenience to practical applications.Therefore,the ability of the algorithm to quickly identify new classes,merely using a small number or even one labeled example,has great significance in the field of remote sensing image analysis.In this article,the few-shot machine learning is introduced into the field of remote sensing image,and the study would focus on the multiresolution remote sensing image few-shot object recognition algorithm based on Relation Network(RN).The simplistic feature merge method of RN leads to unreasonable weight allocation of class features and affects recognition accuracy.In order to address this issue,the RN model based on self-learning feature merge(MRN)is proposed.A self-learning feature merge module is designed to extract class feature while reducing the weight of invalid features to ensure the effectiveness of class features.In the stage of feature comparison,to further improve the recognition accuracy,the MRN model based on split feature measurement(MSRN)is introduced.It applies piece-wise constraint to different class features,increases the spacing of different classes of samples in the feature space.In practical applications,remote sensing image data come from various sources and have different data distribution domains.Thus,the domain adaption ability of the model directly affects the recognition accuracy.This article introduces the Mixup data enhancement module into MSRN to increase the diversity of training samples.And the loss function is modified to regularize the model,to strengthen the generalization ability of the model.The resolution of the remote sensing image in a real-world situation is diverse,to which the MSRN turned out to be sensitive.In order to address the issue,the MSRN based on Residual equivariant mapping(Res-MSRN)is put forward.Residual Equivariant Mapping is introduced to modify the extract sample features,and a resolution coefficient is used to extend the effective feature information of the low-resolution samples.The module is proved to be able to enhance the robustness of the algorithm on different resolution while maintaining the recognition accuracy.Finally,the contrast experiments are conducted made on public remote sensing image data sets.By comparing the models proposed in this article with other state-of-the-art fewshot recognition algorithm,the experimental results show that the proposed Res-MSRN outperforms other models in recognition accuracy,and has better generalization ability and resolution adaptability as well.