High Resolution Range Profiles (HRRPs) play a critical role in radar based Automatic Target Recognition (ATR) by revealing detailed structural information along the range dimension. In legacy radar systems transmitting unmodulated narrow pulses, low inherent range resolution presents significant challenges for HRRP reconstruction, especially under low signal to noise ratio (SNR) conditions. Spectral Inverse Filtering (SIF) is a recently introduced method that enhances resolution through frequency domain deconvolution, but it remains highly sensitive to noise. This paper proposes a physics-informed Convolution Neural Network (CNN) trained using SIF generated clean HRRPs as supervision. The network is trained on a synthetic dataset generated from randomized target profiles under various SNR levels ranging from –5 dB to 40 dB. Quantitative evaluations using Root Mean Square Error (RMSE), Peak Signal to Noise Ratio (PSNR), and Mean Structural Similarity Index (MSSIM) show that the CNN consistently outperforms SIF, especially under low-to-moderate SNR conditions. Visual comparisons confirm the CNN’s ability to suppress noise while preserving key structural features such as peak positions and sidelobes. The results demonstrate that data-driven learning can effectively complement physics-based methods, offering robust, high-fidelity HRRP reconstruction without need to modify the radar hardware or transmitted waveform.
NGUYEN et al. (Thu,) studied this question.