Image Semantic Communications: An Extended Rate-Distortion Theory Based Scheme

Abstract

Semantic communication is a promising paradigm for future artificial intelligence-empowered communications. In this paper, we propose a task-oriented semantic communications with semantic reconstruction (TOSC-SR) scheme for image semantic communications. Different from existing studies focusing on minimizing the pixel-level or feature-level distortion of images, TOSC-SR aims to preserve the semantic information of the image so as to ensure the performance of downstream artificial intelligence (AI) tasks. To this end, a new form of rate-distortion is first derived, which takes the semantic information related to AI tasks into consideration. We formulate TOSC-SR as an extended rate-distortion optimization problem. Then we propose a joint source and channel coding (JSCC) based semantic communication system, which are followed by multiple AI tasks. Experimental results show that the proposed approach outperforms the traditional JPEG, JPEG2000-based communication system and deep learning based multi-task communication system in terms of image reconstruction quality, AI task performance, and multi-task generalization ability.