Low Redundant Attention Network for Efficient Image Super-Resolution

Transformer-based models have demonstrated impressive performance in image super-resolution (SR), but they come with a high computational overhead. In this paper, we present a low redundant attention network (LRAN) for efficient image SR. We observe that there is significant similarity in attention maps across heads and blocks, leading to computational redundancy. First, to mitigate the redundancy in attention maps among heads, we introduce a multi-element mechanism in the self-attention computation. This mechanism allows for the incorporation of various types of self-attention, thus increasing inter-head diversity. Second, to address this redundancy in attention maps among blocks, we propose the hamburger architecture, which introduces enhanced local perception units to capture local information. Moreover, this architecture incorporates a single self-attention layer between several efficient MLP layers. Extensive experiments demonstrate that LRAN outperforms the latest models in lightweight SR, achieving a better trade-off between SR quality and latency. For instance, LRAN surpasses SwinIR-light by 0.25dB PSNR in ×4 SR on Urban100, while running ×5 faster.