Deep Learned-CoSaMP Network for Signal Denoising

Pagoti Reshma, Pradyumna Pradhan, Hrithik Singh, Srinivas Tenneti, Pradip Sasmal, Ramunaidu Randhi

Session: General

Keywords: Sparse signal recovery, OMP, CoSaMP, Algorithm unrolling, Deep neural networks

Abstract:

Deep Neural Networks (DNNs) have garnered considerable attention in the field of sparse signal recovery due to their powerful learning capabilities. However, they face challenges such as a lack of interpretability and a strong dependence on large training datasets.
To address these issues, algorithm unrolling has emerged as a promising approach that systematically transforms iterative algorithms into neural network architectures. Recently, the unrolling of orthogonal matching pursuit (OMP), termed learned OMP (L-OMP), has demonstrated improved performance over existing unrolled methods. Nonetheless, L-OMP exhibits limitations in high-noise scenarios due to its slower dictionary learning process. To overcome this limitation, we propose the unrolling of compressive sampling matching pursuit (CoSaMP), leveraging its batch-wise support selection and noise-pruning capabilities. This method, termed learned CoSaMP (L-CoSaMP), effectively addresses noise-dominated components and accelerates dictionary learning. Experimental results indicate that L-CoSaMP consistently outperforms state-of-the-art unrolled networks, such as the learned iterative soft thresholding algorithm (LISTA) and L-OMP, particularly in high-noise environments. These findings highlight the robustness and efficiency of L-CoSaMP in signal-denoising tasks.

Submission Number: 74