Limited-Data SAR ATR Based on Causal Intervention

This chapter introduces a causal intervention processing architecture based on backdoor adjustment to address challenges in limited-data SAR target recognition. Initially, a multi-center causal target recognition method for limited-data SAR feature selection is proposed. By identifying shared intra-class similar features and incorporating a multi-feature center classifier, this method reduces the reliance on single features in recognition and achieves approximate estimation and elimination of confounding factors. Furthermore, a dual-invariance causal target recognition method for limited-data SAR is developed. By utilizing invariant feature proxies within classes and integrating confounding-invariance loss, this method reduces the data requirements for the optimization function and accomplishes the estimation and elimination of confounding factors. These approaches enable true causal effects in limited-data SAR target recognition and provide innovative solutions for enhancing recognition performance in data-scarce scenarios.