Deep Conditional Generative Semantic Communication for Image Transmission

For semantic communication, the knowledge base is the most significant component. Based on it, both the sender and the receiver are able to perform semantic encoding and decoding processes. In this context, we consider that semantic communication is in “generative form.” Semantic encoding serves to create a carrier of meaning, which can be viewed as a trigger or catalyst. This results in the generation of corresponding content with high similarity to the sender's, thus enabling complete semantic communication. Just like when humans hear the word “hamburger,” their minds conjure up an image of two slices of bread enclosing a meat patty. In this cognitive process, prior knowledge aids the brain in creating mental images, with the word “hamburger” acting as the trigger for this visualization. Building upon this perspective, we propose a deep conditional generative semantic communication system for images. This approach will not only have better communication performance at a low channel SNR but will also address the prevalent issue of the “cliff effect” in conventional separate communication systems. The proposed method is expected to provide high-throughput image transmission under poor communication conditions, advancing the development of generative semantic communication.