Abstract
Vision transformers (ViTs) excel at global context modeling with self-attention. However, standard self-attention leads to quadratic computational complexity, which restricts its practical use in high-resolution or latency-sensitive tasks. Existing methods achieve linear complexity via local window constraints or additive approximations. However, they often compromise long-range dependency modeling. To address this issue, we propose the channel-guided additive attention vision transformer (CGA-ViT), which achieves synergistic optimization of multi-scale feature extraction and efficient global context modeling. First, we propose multi-scale dilated feature embedding (MDFE). By designing multi-scale sampling and spatial feature embedding, we can expand the receptive field and capture fine-grained features simply by adjusting the dilation rate in the early stages; second, we design channel-guided additive attention (CGA), dynamically modulating key vectors using query-derived descriptors, enabling long-range semantic interactions while maintaining linear complexity growth. We adopt a hierarchical structure, and in the shallow layers, we use CGA to carry out local-global interactions and use efficient additive attention in deep layers for global integration. Evaluations on ImageNet-1K show that CGA-ViT achieves 84.0% Top-1 accuracy with 4.7 GFLOPs, outperforming Swin-T (81.3%) and ConvNeXt-T (82.1%) by 2.7 and 1.9 percentage points under comparable computational costs. Ablation experiments verify MDFE and CGA, which together contribute to 65.0% of performance gains, with the rest from token-level supervision. Overall, CGA-ViT effectively balances the intrinsic tradeoff between efficiency and global modeling capability, significantly boosts visual recognition performance without extra computational overhead, and provides an efficient solution for lightweight ViT design.