Heterogeneous Attention for Efficient JSCC-Based Image Transmission

Abstract

Semantic communication, which lies at the intersection of sixth-generation wireless networks and artificial intelligence, has attracted widespread attention. Although existing Transformer-based architectures can achieve high image-reconstruction quality, their substantial computational overhead remains a critical challenge. To address this issue, a communication-oriented JSCC model is proposed, which employs Outlook attention to capture local details and Transformer self-attention to model global semantic dependencies. Meanwhile, semantic-aware adaptive modulation and rate-adaptive feature selection are introduced to adjust the transmitted representation according to the signal-to-noise ratio (SNR) and channel bandwidth ratio (CBR), respectively, with the selected-channel mask transmitted as compressed side information to resolve channel-position ambiguity. Experimental results demonstrate that the proposed model achieves a peak SNR above 26dB and outperforms Transformer-based baselines under low SNRs. Moreover, compared with the Transformer baseline, it reduces the parameter count and inference time by 37% and 33%, respectively. These results indicate that allocating local attention to early stages and global attention to later stages can reduce complexity while maintaining reconstruction quality.