A New Causal Inference Framework for SAR Target Recognition

In synthetic aperture radar (SAR) automatic target recognition (ATR) tasks, deep learning-based methods usually work with the assumption that the training and test target samples are independent and identically distributed. However, the performance of the deep model degrades dramatically when there exists a large distribution variation between training and test data. The collected target samples include not only the target entity but also the target's complicated surrounding environment. So it is difficult to accurately identify targets when they appear in a new background. In this article, we propose a causal inference framework for SAR ATR by removing the background-related bias. This framework can handle more challenging recognition scenarios, SAR background out of distribution (o.o.d) recognition task. First, the SAR ATR task is modeled as a causal graph from a causal inference perspective, and this graph clearly explains the sources of background-related bias in traditional deep models. Then, this graph is intervened to calculate the causal effect caused by background on prediction in accordance with the frontdoor adjustment. This postintervened graph cuts the spurious correlations between background and prediction. Finally, the total effect is used as the final unbiased prediction by removing background-related bias from the original prediction. Our framework does not impose constraints on the specific implementation of the model and does not add any new training parameters. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) and synthetic and measured paired labeled experiment (SAMPLE) benchmarks demonstrate the effectiveness of our proposed framework in the background out-of-distribution case, and it scientifically improve the recognition capabilities of several baseline models.