Deep Causal Metric Learning

ABSTRACT Deep learning, with abundant training data, has the potential to greatly improve the performance of Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR). However, due to the scarcity of SAR data and the high cost of collection, there is a challenge in training deep learning-based SAR-ATR models. Therefore, it is essential to investigate few-shot recognition models to address the challenge of data scarcity in SAR-ATR tasks. This paper proposes a recognition model based on deep causal metric learning, which integrates causal inference with Brownian distance covariance metric (BDCM). The recognition model utilizes causal inference to unveil causal relationships among critical factors in SAR-ATR. The model also designs an effective implementation of causal intervention using feature-level data augmentation to mitigate the adverse impact of confounders on model predictions. BDCM effectively captures the nonlinear relationships between image features, enhancing the model’s ability in feature learning and representation. Experimental results demonstrate the proposed model’s superior few-shot recognition performance. In the 1-shot and 5-shot tasks, our model achieves accuracies of 92.8% and 97.9% on the MSTAR dataset, with performance improvements of 1.0% (1-shot) and 1.8% (5-shot) over state-of-the-art methods. On the OpenSARShip dataset, the model achieves competitive results of 64.9% (1-shot) and 79.2% (5-shot), further verifying its effectiveness in diverse scenarios.