Digital Semantic Communication with Neural Image Compression

Although analog semantic communication systems have attracted significant attention recently, there has been relatively less focus on digital semantic communication systems. In this work, we introduce a neural image compression-enabled semantic communication system to enhance the efficiency of digital image transmission, named NCSC. By employing an accurate and adaptable entropy model, NCSC obtains the efficiently compressed bitstreams, which are compatible with digital communication systems. Incorporating with the well-established digital components, our system trained on the MS-SSIM metric can achieve a significant bandwidth compression ratio of 0.002 at low SNR, reducing remarkably transmission overhead. Extensive simulations show that our approach outperforms traditional digital communication systems in terms of perceptual quality and bandwidth efficiency under challenging channel conditions.