Deep Denoising and Unfolding for IRS Cascaded Channel Estimation With Element-Grouping

Abstract

In intelligent reflecting surface (IRS) assisted communication, reliable channel estimation with reduced training overhead is crucial for the timely reconfiguration of wireless environments through IRS reflection control. We consider IRS cascaded channel estimation with element-grouping, which reduces the training overhead by sharing a reflection coefficient among a subgroup of elements at the cost of interference. Based on the maximum a posteriori estimation, we design a deep denoising network that supports various noise levels for noise-dominated regions and a deep unfolding network that optimizes extrapolation and denoising alternately for interference-dominated regions. The proposed networks provide a more reliable estimation through balanced noise and interference mitigation than the conventional networks proposed recently for element-grouping and super-resolution.