Attention-enabled Channel Estimation for STAR-RIS-aided Indoor and Outdoor MIMO

The innovative idea of simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) allows for a full-space manipulation of signal propagation. In STAR-RIS-aided communication systems, which integrate both transmission and reflection functionalities, the distribution of channels and the complexity of channel estimation are notably augmented due to the dual-mode operation. Therefore, the practical application of analytical channel estimation techniques combined with integration computation is not feasible. To improve the channel estimation accuracy, we model the channel estimation as a denoising task retrieving the channel coefficients from noisy pilot-based observations. We design a novel channel attention-based deep residual neural network, named CANet, to handle intricate channel estimation challenges in such systems. The CANet can adaptively focus on useful features of the input signal, effectively denoising and improving the accuracy of the channel estimation. Our comprehensive simulations demonstrate that CANet can robustly and satisfactorily perform channel estimation with inputs of different distributions. The performance and convergence speed of the proposed CANet are significantly faster than some existing deep residual neural networks without channel attention.