Single image super-resolution with channel attention and diffusion

Single image super-resolution (SISR) is a fundamental vision task that facilitates a range of applications. In SISR, human perception, objective quality index, and complexity are three main concerns. In this paper, we propose a new SISR framework known as single image super-resolution with channel attention and diffusion (SRCAD). The proposed SRCAD combines the stochastic iterative mechanism of the denoising diffusion probabilistic model (DDPM) with advanced feature encoding techniques. With the guidance of encoded features, the diffusion model better predicts SR images with more details, and the model size is also reduced as well. To be specific, in feature encoding, SRCAD introduces a pre-trained encoder combined with dimensional interleaved product channel attention (DIP-CA), which extracts key features at a low computational cost. The extracted deep features guide iterative denoising and are combined with distribution-aware feature subset fusion (DFSF) to reduce the data dimension of the features. Experimental results demonstrate that SRCAD performs well on four datasets and two SR tasks. It also outperforms other state-of-the-art models in terms of objective quality metrics, including PSNR, SSIM, and LR-PSNR. Besides, it also reduces the number of parameters, thus delivering high performance with low complexity.