Channel Estimation for Underwater Acoustic OFDM Communications: An Image Super-Resolution Approach
In this paper, by exploiting the powerful ability of deep learning, we devote to designing a well-performing and pilot-saving neural network for the channel estimation in underwater acoustic (UWA) orthogonal frequency division multiplexing (OFDM) communications. By considering the channel estimation problem as a matrix completion problem, we interestingly find it mathematically equivalent to the image super-resolution problem arising in the field of image processing. Hence, we attempt to make use of the very deep super-resolution neural network (VDSR), one of the most typical neural networks to solve the image super-resolution problem, to handle our problem. However, there still exist significant differences between these two problems, we thus elegantly modify the basic framework of the VDSR to design our channel estimation neural network, referred to as the channel super-resolution neural network (CSRNet). Moreover, instead of training an individual network for each considered signal-to-noise ratio (SNR), we obtain an unified network that works well for all SNRs with the help of transfer learning, thus substantially increasing the practicality of the CSRNet. Simulation results validate the superiority of the CSRNet against the existing least square (LS) and deep neural network (DNN) based algorithms in terms of the mean square error (MSE) and the bit error rate (BER). Specifically, compared with the LS algorithm, the CSRNet can reduce the BER by 44.74% even using 50% fewer pilots.