Deep Learning-Based Optimal Transmission of Embedded Images Over Interference Channels

In this paper, we consider the transmission of embedded (or progressive) images over interference channels for the purposes such as surveillance, and special missions of public safety or emergencies. When transmitting embedded images, the joint optimization of source and channel coding for a series of numerous packets has been a challenging problem. Further, the problem is more complicated if the interference is present and the relevant transmit signal power control is also involved with the optimization. This is because the number of ways of jointly optimizing transmit power and a set of channel codes to be assigned to progressive packets is much larger than that of solely optimizing a set of channel codes. For a single communication link, there have been many studies about the optimal allocation of channel codes to progressive packets, but those studies do not provide a solution for the aforementioned optimization problem. We propose a neural network based optimization method to address such issues in interference channels. Specifically, the proposed scheme incorporates a fully-connected layer based neural network, where the input layer takes the distortion-rate characteristics of images to be transmitted, and the estimate of channel propagation effects such as path loss and shadow fading. Then, the network predicts the optimal transmit power and a set of optimal channel codes to be assigned to progressive packets. We demonstrate that a neural network can analyze and learn the highly nonlinear distortion-rate characteristics of images, which are associated with wireless channel fading effects, and can com up with the relevant optimization strategy. Compared to the baseline method that separately optimizes transmit power and a set of channel codes, our scheme offers significantly improved peak-signal-to-noise ratio performance.