Reference feature enhancement with spatial-channel fusion for image super-resolution

Reference-based image super-resolution (RefSR) utilizes features from a reference image that are related to the low-resolution image to enhance its super-resolution. Most previous approaches use convolutional operations to fuse the features of the reference image with those of the low-resolution image, often overlooking global feature fusion, inadequately considering the channel information of the fused features, and lacking dynamic refinement of the reference image features. In this paper, we propose a novel reference-based image super-resolution method with dynamic refinement of reference features and spatial-channel fusion. First, designing Spatial Transformer, we achieve global-local feature fusion between the low-resolution image and the reference image at the spatial level. Next, Channel Transformer facilitates the interaction of features at the channel level. Finally, we design Window Mixformer, which dynamically refines the reference image features through a hybrid attention mechanism, enabling the extraction of more relevant feature information. Experimental results on public datasets demonstrate that our method outperforms several state-of-the-art reference-based image super-resolution methods, achieving better reconstruction results.