Hierarchical Multi-Granularity Joint Source-Channel Coding for Image Semantic Transmission

Deep Joint Source-Channel Coding (DEEPJSCC) is an effective method for realizing semantic communication. However, the existing DEEPJSCC methods cannot fully exploit the potential of deep semantics, posing serious challenges to the quality and reliability of communication. To address this issue, we propose a multi-granular DEEPJSCC (MDEEPJSCC) framework for image semantic transmission. The framework employs a hierarchical multi-scale encoder, where each layer corresponds to a different scale representation of the image, and the distinct attention mechanisms (AMs) are leveraged to capture diverse semantic information at each scale level. Strengthening the final semantic representation by integrating the semantic features extracted from each layer. On the other hand, a novel adaptive masking technique is also incorporated into the framework to optimize the allocation of code lengths, improve the quality of code rate control, and reduce the communication overhead based on the compression requirements of transmission and the dynamic assessment of the semantic importance of images by the two networks. Simulation results show that the MDEEPJSCC framework effectively improves the performance of wireless image transmission and provides strong support for high-quality semantic communication.