Reinforcement Learning for SAR Target Orientation Inference with the Differentiable SAR Renderer

This paper attempts to infer the orientation angle of the target in synthetic aperture radar (SAR) images using reinforcement learning. It is intended to address the challenges like limited interpretability, the scarcity of SAR data and complex imaging mechanisms restrict the broader application of learning-based approaches. We propose an interactive deep reinforcement learning (DRL) framework, where an electromagnetic simulator named differentiable SAR renderer (DSR) is embedded to facilitate the interaction between the agent and the environment. Specifically, DSR generates SAR images at arbitrary orientation angles in real time, helping to simulate a human-like process of angle estimation. The differences in sequential and semantic aspects between images of different orientation angles are leveraged to construct the state space in DRL, which effectively suppress the complex background interference, and enhance the sensitivity to temporal variations. Moreover, to maintain the stability and convergence of our approach, reward mechanisms such as memory difference, smoothing and boundary penalty are incorporated to contribute to the formulation of the comprehensive reward function. Extensive experiments performed on both simulated and real datasets demonstrate the effectiveness and robustness of our proposed method. In addition, when utilized in the cross-domain area, the proposed method mitigates inconsistency between simulated and real domains, outperforming reference methods significantly.