A deep learning-inspired compressed sensing framework for sparse channel estimation in optical wireless communications

Abstract

Optical wireless communication is an important candidate for next-generation high-speed communication systems due to its large bandwidth, small power cost and freedom from EM noise. Hard turns the task of tracing sparse channels that signals meet in rough air, where heavy turbulence twists their paths. Tight grows the gap before old sensing tools, and the rule is that better guesses need slower runs. Fresh comes the path offered by this paper, while old deep models still ask for long training and big marked sets. Under fixed weights do OMP and CoSaMP fuse, which builds a fused base; a sparse polish step then sharpens this base. Unlike conventional deep learning methods, DL-CS does not require neural network training or large labeled datasets, making it lightweight, interpretable, and easy to implement. Simulations show that the proposed method improves NMSE and BER over OMP and CoSaMP, with runtime close to that of OMP. Under the current simulation setting, SP achieves the lowest NMSE among the tested baselines, while DL-CS still demonstrates a favorable balance between accuracy and complexity. Runtime analysis further shows that DL-CS maintains moderate computational cost, and statistical evaluation using confidence intervals and Welch’s t-test is provided to support the efficiency comparison. Overall, the proposed DL-CS method offers a practical training-free enhancement of classical greedy algorithms for sparse channel estimation in optical wireless communications.