Research On SAR ATR Method With Limited Data Based On Causal Inference

With the rapid development of artificial intelligence technology,Synthetic Aperture Radar Automatic Target Recognition(SAR ATR)can quickly and accurately recognize target types with sufficient samples.It can be widely applied in areas such as agricultural and forestry environment monitoring,natural disaster assessment,battlefield situation awareness,and reconnaissance intelligence interpretation.However,due to factors such as the scarcity of targets of interest,difficulties in SAR image acquisition,and challenges in precise annotation,it’s hard to obtain enough SAR target training samples.This has made existing methods ill-suited for limited-sample SAR target recognition problems.This dissertation addresses the limited-sample SAR target recognition problem,focusing on the study of SAR target recognition causal modeling,causal feature extraction,causal approximate recognition,and causal interventional recognition.The innovations in model construction,mechanism,and methodology can be summarized as:1.This dissertation established a causal model for limited-sample SAR target recognition.By introducing causal inference theory into SAR target recognition,this dissertation derived the causal relationship between SAR image description variables and recognition performance.This reveals that spurious causal effects are the theoretical root of poor performance in limited-sample SAR target recognition,providing a theoretical foundation for improving its performance.2.This dissertation proposed a causal feature extraction method for limited-sample SAR targets.Using the counterfactual causal interventional processing framework and with stable inter-class distance loss,this dissertation suppressed the influence of SAR image areas with low discriminability on feature discriminability.Further,by optimizing the effective area search criteria,this dissertation not only effectively eliminated the spurious causal effects of low discriminability areas but also accurately extracted features from high discriminability areas of SAR images under limited-sample conditions.3.This dissertation proposed a causal approximate target recognition method for limited-sample SAR.By utilizing the maximum inter-class distance criterion to obtain inter-class differential recognition information and then employing informationaugmentation feature distribution optimization and hierarchical-feature adaptive weighting technology,this dissertation identified and eliminated the impact of outlier samples,achieving accurate recognition of limited-sample SAR targets.4.This dissertation introduced a causal interventional target recognition method for limited-sample SAR.Using the backdoor adjustment causal interventional processing framework,this dissertation selected features with high discriminability correlation.This effectively suppressed the spurious causal effects of confounding factors.On this basis,through the proposed intra-class feature invariance criterion and confounding invariance optimization criterion,this dissertation reduced the data requirement for deconfounding,achieving accurate recognition of limited-sample SAR targets.The effectiveness of the aforementioned models and methods has been verified through multiple real-measured datasets.Experimental results demonstrate that the models and methods proposed in this dissertation effectively address the key problems faced in limited-sample SAR target recognition,laying a theoretical and technical foundation for robust,high-performance SAR target recognition.