Attention-Based Adaptive Rate Allocation for Semantic Communication

Abstract

Semantic communication and deep joint sourcechannel coding (Deep JSCC) have emerged as promising AIenabled coding paradigms for next-generation wireless communications. For wireless image transmission, existing Deep JSCC methods can jointly optimize source representation and channel protection in an end-to-end manner. However, most Transformerbased semantic transmission schemes still adopt a fixed channel bandwidth ratio (CBR) and transmit encoded features uniformly, which limits the efficient utilization of channel resources under bandwidth-constrained wireless channels. To address this issue, this paper proposes an attention-guided adaptive rate allocation method based on the wireless image transmission Transformer (WITT). Specifically, a dual-branch importance modeling module is introduced after the Swin Transformer encoder to jointly evaluate the semantic importance of encoded features from both channel and spatial dimensions. Based on the estimated importance, an adaptive transmission mask is generated to allocate more channel resources to semantically important features while suppressing low-contribution redundant features. Experiments on the CIFAR-10 dataset over additive white Gaussian noise (AWGN) and Rayleigh fast-fading channels show that the proposed method maintains stable reconstruction performance under different signal-to-noise ratio (SNR) conditions. The proposed adaptive coding mechanism provides a semantic-aware resource allocation solution for 6G-oriented wireless image transmission.