Research On SAR Image Target Recognition Method Under Data Imbalance

Synthetic Aperture Radar(SAR)plays an important role in both civil and military fields because it is an active microwave detection system that can operate around the clock in complex environments and has excellent penetration and camouflage recognition capabilities.Automatic Target Recognition(ATR)is an important technology in the field of radar signal processing interpretation and SAR image interpretation.However,in practical application scenarios,it is costly to obtain sufficient and balanced number of SAR image samples due to factors such as the working platform of sensors,target acquisition cost and non-cooperative characteristics of targets,which results in the data imbalance problem and makes it difficult for the existing SAR ATR methods to achieve the expected results.This thesis aims to address the problems in the existing unbalanced SAR image target recognition models,focusing on theoretical exploration and methodological research at the data level,feature extraction level and classifier level,and proposes three methods to achieve SAR target recognition under imbalanced samples conditions.The main research works in this thesis are as follows:(1)A two-stage adaptive sampling unbalanced SAR target recognition method is proposed to address the problem of poor model training performance due to insufficient amount of sample data at the data level for SAR image target recognition under data imbalance conditions.The use of a few classes of adaptive synthetic oversampling techniques to expand the sample set during the data pre-processing phase to initially mitigate data imbalances;Then,a simple sample selection strategy based on self-step learning is proposed to select high confidence samples with low loss of the current model from the majority class samples to further alleviate the imbalance between target data.Finally,the trained network model is used to complete the inference of class labels for new samples.(2)A data augmentation feature attention network is proposed to solve the problem of feature information scarcity due to insufficient minority class samples at the feature extraction level of SAR image target recognition under data imbalance conditions.First,virtual samples are synthesized based on the data neighborhood relationship to expand the number of minority class targets; then a feature attention network is built to extract rich and discriminative features for minority class targets; finally,an asymmetric loss function is used to further alleviate the problem of data imbalance,so as to robustly achieve SAR target recognition under samples imbalance conditions.(3)A cost-sensitive ensemble classification network is presented to solve the problem of skewed classifier models due to differences in the amount of sample data at the classifier level for SAR image target recognition under data imbalance conditions.The proposed method first establishes a multiscale feature extraction network to learn rich features from limited samples of minority class targets.Then a minority class penalty factor based on cost-sensitive matrix is proposed to solve the problem that the classifier prefers the majority class.Meanwhile,in order to prevent the model from overfitting on the minority class due to insufficient samples of the minority class,the propose method constructs multiple classifiers in different layers of the network to further improve the inference ability of the model.