Task-Oriented Semantic Communication with Semantic Reconstruction: An Extended Rate-Distortion Theory Based Scheme

Compressed semantic representation of source is essentially important to accomplish various artificial intelligent (AI) tasks in task-oriented semantic communication (TOSC). In this paper, by extending the rate-distortion theory to multiple tasks, we propose a TOSC scheme with semantic reconstruction (SR), named as TOSC-SR, in the joint source and channel coding (JSCC) framework. Besides extracting and compressing task semantics, our basic idea here is to reconstruct images with task semantics rather than traditionally in the pixels or features. The main purpose is to share the semantic-reconstructed images among multiple tasks with enhanced generalization ability under certain rate. We formulate the TOSC-SR scheme as a rate-distortion optimization problem, where a novel semantic distortion measurement is defined by mutual information of source, the semantic-reconstructed images, and task labels, pairwise. We derive an analytic solution for the formulated problem, where the self-consistent equations are obtained to determine the optimal mapping of source and the semantic-reconstructed images by taking task labels into account. In the TOSC-SR scheme which is feasible in practice, a relaxed version of loss function is derived based on variational approximation of mutual information. Then we adopt the classification task to train TOSC-SR, and the object detection task to validate the generalization ability. Experimental results show that the proposed TOSC-SR scheme outperforms conventional JPEG, JPEG2000 based communication schemes and deep learning based TOSC with general reconstruction schemes in terms of reconstruction quality, classification and object detection performance at the same source compression ratio and signal-to-noise (SNR) regime.