Physical-Abstract Bidirectional-Guided Learning for High-Resolution Radar Target Recognition

Target recognition based on high-resolution radar has garnered increasing attention, with target-aspect sensitivity being one of the primary challenges. For that, this paper proposes a Physical-Abstract Bidirectional-Guided Learning Network (PABGL-Net) that leverages scattering center based physical characteristics to guide deep models training, thereby enhancing the robustness and interpretability of deep features. The core innovation lies in modeling this process as a bidirectional integration, enabling simultaneous parameter estimations of scattering center based physical models and mapping abstract representation to local scattering structures of targets. Furthermore, to improve adaptability and reduce computational complexity, several simple yet effective training strategies are introduced within the proposed framework. First, an adaptive method for determining the number of scattering centers and neural network architecture is presented. Second, a soft-threshold based target region extraction algorithm is developed, significantly reducing the parameter search space. The performance of the proposed algorithm is validated using 1D Carrier-free ultra-wideband (UWB) radar echoes and synthetic aperture radar (SAR) imagery. Experimental results show that the proposed method is capable of handling challenging conditions where there are significant differences in target aspect between the training and testing datasets. Moreover, integrating the bidirectional -guided learning strategy with a lightweight network yields comparable recognition performance with lower computation complexity, requiring only 0.64 million parameters and 0.018 GFLOPs per layer for 2D SAR images.