Adaptive Texture Transfer for Reference-Based Super-Resolution: A Novel RDA Module

This paper introduces a novel Residual Dense Aggregation (RDA) module for reference-based super-resolution (RefSR). The RDA module leverages a warping function (ω), concatenation operation ([;]), convolutional layers (Conv), and activation functions (Sigmoid and Tanh) to adaptively transfer textures from a reference image. The maximum magnitude (r) is set to 10 by default, and F l represents the features of the upsampled low-resolution (LR) images at the l-th scale.

The proposed method utilizes a mask to selectively transfer textures, effectively addressing correspondence mismatching issues. Even if the LR and reference images exhibit significant differences, the model can adaptively transfer relevant textures. When the reference image contains irrelevant textures or no information, our model can determine whether to transfer textures from the reference image, further mitigating correspondence mismatching.

This paper focuses on comparing the RDA module with existing RefSR methods that employ a single reference image. For a fair comparison, the model transfers one relevant texture from the reference image. Through our architecture, the RDA module enhances RefSR performance by transferring textures at each scale during both downscaling and upscaling, differentiating it from C 2 -Matching [12].