In the underlay cognitive radio networks, the radio environment maps (REMs) estimation is the main challenge in sensing the idle wireless spectrum resources. Traditional deep learning-based algorithms estimate the REMs on the basis of the high-quality, large-scale complete training images. However, collecting the complete radio environment images is time-consuming and requires a numerous number of power spectrum sensing nodes. For this reason, we propose a generative adversarial networks-based pixel regression framework (PRF) for underlay cognitive radio networks. The PRF algorithm relaxes the requirement of the complete training images, and estimates the radio environment maps only on the basis of the incomplete REMs images, which are easier to be collected. First, we transform the radio environment maps estimation task into a pixel regression task through the color mapping progress. Then, to extract helpful information from the incomplete training data, we design a feature enhancing module for the PRF algorithm, which intelligently learns and emphasizes the important features from the training images. Finally, we use the trained pixel regression framework to reconstruct the radio environment maps in the target area. The proposed algorithm learns accurate radio environment characteristics from the incomplete training data rather than making direct biased or imprecise radio propagation assumptions as in the traditional methods. Thus, the PRF algorithm has a better REMs reconstruction performance than the traditional methods, as verified by simulations.
Keywords: cognitive radio networks; deep learning; generative adversarial networks; image reconstruction; wireless communication.