Mobile Image Communications Using Deep Neural Networks for Heterogeneous User Devices in MIMO Channels

Abstract

This article considers a mobile multimedia communication scenario in which two types of heterogeneous users with different numbers of antennas and different display resolutions are served within the same service area. We propose an efficient image transmission scheme that makes use of deep neural network-based super-resolution techniques, spatial diversity, and diversity-multiplexing tradeoff (DMT) achieving codes. The proposed scheme transmits a low-resolution (LR) image to two types of users, along with residual pixel-error map that contains high-frequency details of high-resolution (HR) images and is produced by a convex optimization technique. Then, a user retaining an HR screen employs super-resolution to reconstruct an HR image from the received LR image, and also exploits the residual map to further enhance the image quality. Our scheme properly trains deep neural network models for super-resolution, based on the source coding rates of the images to be transmitted over bandwidth-limited multiple-input multiple-output (MIMO) channels. That is, various quantization errors resulting from image compression are considered when training the deep neural network-based super-resolution models. For MIMO systems, considering and utilizing the relationship between the number of antennas and the screen resolution, which is based on hardware space of user devices, the proposed scheme encodes an LR image with spatial diversity techniques, and encodes a residual pixel-error map with DMT-achieving codes. Numerical evaluation shows that our scheme significantly outperforms the strategy that broadcasts either HR or LR images to two types of heterogeneous users.