AgentJSCC: Image Transmission based on Transformer Block with Agent Attention

Efficient deep joint source-channel coding (JSCC) is a key technology driving advances in semantic communications. However, the high computational cost of global attention and the limited global modeling capacity of local window attention have partially constrained previous wireless image transmission solutions. In this paper, we design a novel JSCC scheme named AgentJSCC, introducing an Agent Attention module into a Vision Transformer backbone for wireless image transmission, aiming to strike an effective balance between computational efficiency and representational capacity. The Agent Attention mechanism is defined as a quadruple ($Q, A, K, V$), where a set of agent tokens A is incorporated to enhance the traditional attention module. This mechanism operates in two stages: in the first, the agent tokens act as intermediaries for the query tokens Q, extracting and integrating information from the keys K and values V. In the second, the integrated global features are fed back into the original Q via reverse information flow, forming a more efficient attention mechanism. By replacing certain encoderstage attention modules in a Swin Transformer-based backbone with Agent Attention, we achieve a model offering a broader receptive field under the same or even reduced computational overhead. Experimental results show that AgentJSCC not only outperforms the SwinJSCC scheme (based on Swin Transformer) in terms of performance, but also features lower computational complexity and reduced inference latency. Code is available at https://github.com/cxp-ustc/AgentJSCC.