ASRMF: Adaptive image super-resolution based on dynamic-parameter DNN with multi-feature prior

Abstract

In recent years, single-image super-resolution has made great progress due to the vigorous development of deep learning, but still has challenges in texture recovery for images with complex scenes. To improve the texture recovery performance, we propose an adaptive image super-resolution method with multi-feature prior to model the diverse mapping relations from low resolution images to their high resolution counterparts. Experimental results show that the proposed method recovers more faithful and vivid textures than static methods and other adaptive methods based on single feature prior. The proposed dynamic module can be flexibly introduced to any static model and further improve its performance. Our code is available at: https://github.com/zzsmg/ASRMF.