Intelligent Reflecting Surface-Aided Wireless Networks: Deep Learning-Based Channel Estimation Using ResNet+UNet

Abstract

Accurate channel estimation is essential for optimising intelligent reflecting surface-assisted multi-user communication systems, particularly in dynamic indoor environments. Conventional techniques such as least squares (LS), linear minimum mean square error (LMMSE), and orthogonal matching pursuit (OMP) suffer from noise sensitivity and fail to effectively capture spatial dependencies in high-dimensional intelligent reflecting surface (IRS)-assisted channels. To overcome these limitations, this work proposes a deep learning-driven ResNet+UNet framework that refines initial LS estimates using residual learning and multi-scale feature reconstruction. While UNet enhances channel estimation through hierarchical processing, efficiently decreasing noise and enhancing estimate accuracy, ResNet gathers spatial features. Simulation results show that the proposed method significantly outperforms existing methods across various performance metrics. In NMSE versus signal-to-noise ratio assessments, the proposed approach surpasses convolutional deep residual network (CDRN) by 59%, OMP by 81%, LMMSE by 114%, and LS by 115%. When IRS elements are modified, it overcomes CDRN by 60%, OMP by 78%, LS by 107%, and LMMSE by 110%. Along with this, recommended structure performs more effectively than CDRN by 39%, OMP by 44%, LS by 122%, and LMMSE by 129% across various antenna configurations. The proposed approach is particularly beneficial for augmented reality (AR) applications, where real-time, high-precision channel estimation ensures seamless data streaming and ultra-low latency, enhancing immersive experiences in AR-based communication and interactive environments. These results illustrate the proposed method's scalability and resilience, making it a suitable choice for next-generation IRS-assisted wireless communication networks.