ESC: An Efficient Semantic Communication Architecture with Feature Selection and Adaptive Inference

Semantic communication is a novel communication paradigm that demonstrates great potential in information transmission applications, particularly in challenging scenarios characterized by low signal-to-noise ratio (SNR) conditions. It extracts task-relevant semantic information and performs end-to-end optimization of source and channel coding. However, current mainstream architectures do not account for the importance of features in the subsequent reconstruction process. This shortcoming leads to the transmission of all features, resulting in inefficient bandwidth utilization. Furthermore, existing methods reconstruct all images equally, regardless of their differences in complexity, which is inefficient and wastes computational resources. To address these issues, we develop an efficient semantic communication architecture, termed ESC. Specifically, we design feature selection and reconstruction modules to filter out unimportant information, addressing the problem of transmission feature redundancy and improving bandwidth utilization efficiency. In addition, we develop a multi-branch decoder architecture and an adaptive inference strategy to accommodate the varying complexities of images, allowing samples with satisfactory reconstruction results to exit the decoder network early, thus reducing inference costs. By introducing feature selection, our architecture's reconstruction quality surpasses that of 5G systems and mainstream semantic communication under the same bandwidth on the Kodak and CLIC2021 datasets. Our adaptive inference strategy achieves speed-ups of approximately 1.37 × and 1.43 × respectively, with only minimal degradation in image reconstruction quality.