WDLReconNet: Compressive Sensing Reconstruction With Deep Learning Over Wireless Fading Channels

Hancheng Lu, Lei Bo

Abstract: Deep learning has been exploited in compressive sensing to reduce the computational complexity of reconstruction algorithms. However, existing deep-learning-based reconstruction algorithms might result in poor performance, when applied in wireless transmission environments. First, the impact of channel noise and fading on reconstruction has not been considered. Second, with the fully connected layers, most of these algorithms have been designed for a given sampling ratio, and thus cannot handle bandwidth variations. To tackle these problems, we propose a wireless deep learning reconstruction network (WDLReconNet). The most distinctive aspect of the WDLReconNet is that a feature enhancement layer (FEL) is designed and combined with the convolutional neural network (CNN). To combat the adverse impact from wireless transmission environments (i.e., channel noise and fading and bandwidth variations), FEL enhances features remained in compressive measurements by roughly recovering the signal based on dictionary learning. Then, this rough signal is processed by CNN to finalize reconstruction. In this way, the advantages of CNN in signal feature learning can be fully exploited for reconstruction. Furthermore, we propose fast-WDLReconNet to accelerate the training process of WDLReconNet. The experimental results demonstrate that the proposed algorithms outperform existing traditional and deep learning-based reconstruction algorithms under various scenarios.