Untrained Deep Learning Techniques for Channel Estimation With Limited Pilots in RIS-Enhanced SIMO-OFDM Systems

The estimation of channel state information (CSI) is a well-known challenge in reconfigurable intelligent surface (RIS)-assisted communication systems due to the limited signal processing capabilities of the RIS node. Conventional methods often resort to employing a substantial number of pilot sequences to obtain the necessary degrees of freedom, leading to a significant reduction in the data transmission rate. In this study, we present deep learning-based channel estimator designed for scenarios with limited pilots in an RIS-assisted single-input-multiple-output (SIMO) orthogonal frequency division multiplexing (OFDM) system. Specifically, we introduce an untrained deep neural network (DNN) based on the deep image prior (DIP) network. This DIP-based DNN is employed to denoise the estimated channel using orthogonal matching pursuit (OMP), ultimately achieving a more accurate channel estimation. Numerical results from our experiments demonstrate that the proposed untrained DIP-based DNN estimator offers significant improvements in accuracy when compared to both conventional OMP-based approaches and training-based DNN estimators.