Research On Few-shot Synthetic Aperture Radar Target Recognition

Synthetic aperture radar(SAR)is an all-weather,day and night observation microwave sensor,which plays an important role in various fields.Among them,automatic target recognition(ATR)is one of the important applications of SAR image interpretation.Data-driven target recognition methods have become the mainstream in current SAR ATR field,and these methods can achieve excellent performance under the condition of sufficient training samples.However,in practical scenarios,it is often difficult to acquire and label a large number of SAR images.Therefore,the few-shot problem caused by the phenomenon of sample scarcity has become an important challenge in SAR target recognition at present stage.In response to the challenge of few-shot SAR target recognition,this thesis takes meta-learning as a guiding principle within the framework of deep learning,and refines several urgent issues in feature extraction and classifier design.It focuses on theoretical exploration and methodological research in three aspects: knowledge transfer,reasoning principle,and optimization function design,with the aim of achieving robust SAR target recognition under the condition of limited samples.The main contents and innovations of this thesis are as follows:(1)A few-shot SAR target recognition method named bi-similarity with capsule embedding network is proposed.This method establishes an equivariant feature extraction network based on capsule routing units to improve the distinguishability of features.In order to improve the robustness of target inference under the condition of few training samples,a classification criterion based on bi-similarity measurement is proposed.Furthermore,to enhance the discriminative ability in the feature space,a hybrid contrastive loss function is proposed,which can promote intra-class compactness and inter-class separability.(2)A few-shot SAR target recognition method called transductive prototypical attention reasoning network is proposed.In the feature extraction stage,a region awareness module is firstly proposed,which aims to locate target area of interest and to weaken redundant background information.On this basis,a spatial mutual feature attention is proposed to strengthen the correlation between support and query samples.In the inference stage,a transductive prototype inference method is proposed,which implements prototype enhancement by adaptively weighting pseudo-labeled high-confidence samples.At the same time,a marginal adaptive hybrid loss function is proposed to optimize the classification boundary while enhancing the discriminability of feature embedding space.(3)A few-shot SAR target recognition method named deep hyperbolic hybrid graph inference network is proposed.This method proposes a global-local feature extraction model,which aims to extract discriminative and rich knowledge from limited SAR images.Besides,abandoning traditional Euclidean embedding,a target reasoning space based on deep hyperbolic embedding is established to learn the hierarchical relationship between samples.To address the issue of degraded recognition capability caused by abnormal samples and class-overlapping samples,a classification method based on instance-prototype label hybrid propagation mechanism is proposed,which combines instance graph label propagation and prototype graph label propagation to improve the recognition performance of the model.