MAFT: An Image Super-Resolution Method Based on Mixed Attention and Feature Transfer

Abstract

Reference-based image super-resolution methods, which enhance the restoration of a low-resolution (LR) images by introducing an additional high-resolution (HR) reference image, have made rapid and remarkable progress in the field of image super-resolution in recent years. Most of the existing methods use an implicit correspondence matching approach to transfer HR features from the reference image (Ref) to the LR image. However, these methods lack the further judgment and processing of the HR features from Ref, which limits them in challenging cases. In this paper, We propose an image super-resolution method based on mixed attention and feature transfer (MAFT). First, we obtain the deep features of the LR and Ref images through the encoder network, then extract the transferred features from Ref through the attention network, and perform adaptive optimization processing on the features, and finally fuse the transferred features with LR features to achieve a high-quality image reconstruction. The quantitative and qualitative experiments on these benchmarks, i.e., CUFED5, Urban100 and Manga109, show that MAFT outperforms the state-of-the-art baselines with significant improvements.