Abstract:

Image itself information is naturally robust to image reconstruction, yet most current super-resolution methods do not fully utilize global feature information. This study proposes a new image super-resolution model mixing multiple and multi-scale attentions, including two new modules: Multi-scale hybrid non-local attention upsampling module and residual dense attention block. Different from previous nonlocal methods, multi-scale hybrid non-local attention upsampling module mixes pixel-based and patch-based nonlocal attention and establishes patch-level upsampling mapping relationships at multiple scales, which enables a wider global search space. The residual dense attention block establishes attention associations in channel and spatial dimensions, which enhances the transfer and fusion of front-to-back attention information through dense connections. In this study, quantitative and qualitative evaluations are conducted on several benchmark datasets, and the experimental results show that the model outperforms similar super-resolution models in terms of performance and reconstruction quality.

Key words: image super-resolution, multi-scale, attention mechanism, non-local, recurrent network, dense connection, upsampling, self-similarity