Prior embedding multi-degradations super resolution network

Nowadays, multi-degradations super-resolution (MDSR) has gradually received wide attention for its ability to deal with multiple degradations including blur, noise in a low-resolution (LR) image. Deep neural networks based MDSR methods have gained rapid development by utilizing the degradation prior. Though achieving promising results, most existing MDSR models fail to fully utilize the degradation prior and efficiently reconstruct high-resolution (HR) image by simple concatenation and iterative operations. This paper proposes an end-to-end MDSR approach to effectively extract features from the degradation prior and embed the degradation features into image features via the degradation feature extraction module and the prior embedding module. In addition, the axis attention mechanism and the pixel attention mechanism are proposed to augment the representation power of image features. Extensive experiments demonstrate that our approach can simultaneously solve different degradations and produce state-of-the-art results on standard benchmark datasets, with advantages in terms of quantitative metrics, visual quality, and reconstruction efficiency.