Sparse recovery networks for target detection in OFDM-based passive radar

With advances in deep learning for signal recovery based on sparse representation, this paper proposes two efficient sparse recovery (SR) networks to solve the target detection problem and generate range-Doppler maps in orthogonal frequency division multiplexing-based passive radar systems. First, by unfolding the alternating direction method of multipliers (ADMM) into a deep network with learnable parameters, a deep unfolding SR-ADMM-Net is developed, which avoids manual parameter configuration and achieves superior performance to ADMM. However, SR-ADMM-Net integrates sparsity prior information into the sparse recovery task via regularization but access only partial data domain knowledge. Therefore, this work proposes another method leveraging generative priors, which is derived from the least squares generative adversarial network (LSGAN) framework and termed SR-LSGAN. Generative priors can capture complex structural features of clutter and target signals, but their effectiveness depends on the solution space of the pre-trained generator. SR-LSGAN leverages prior distributions from the generative model and extends the solution space by dynamically fine-tuning parameters through gradient descent and internal learning, thereby improving recovery accuracy and target detection performance. Experiments on simulated and measured data demonstrate the generalizability and effectiveness of the proposed networks.