Abstract
Transformer-based methods have shown strong potential for single image super-resolution (SISR) due to their ability to model long-range dependencies. However, most existing approaches primarily focus on global representation learning while underexploring high-frequency priors that are crucial for recovering fine textures and sharp edges. In this work, we present a wavelet-transform-based analysis showing that transformers are effective at modeling low-frequency structures but comparatively less effective than convolutional neural networks (CNNs) in recovering high-frequency details. Motivated by this observation, we propose WaveletMixer, a novel SISR framework that integrates wavelet decomposition into feature learning to exploit the complementary strengths of transformers and CNNs in the frequency domain. WaveletMixer adopts a dual-branch architecture for low- and high-frequency wavelet feature extraction, introduces Wavelet Channel Attention (WCA) to adaptively emphasize informative wavelet subbands, and employs Deformable Cross Attention (DCA) for cross-wavelet subbands interaction and fusion. In addition, an Omni Channel Attention (Omni-CA) module is designed to enhance spatial-channel feature aggregation through multi-axis interaction. Extensive experiments on benchmark datasets demonstrate that WaveletMixer consistently outperforms several state-of-the-art methods and achieves a favorable trade-off between reconstruction performance and model complexity.
Funding Information
  • National Natural Science Foundation of China (62071174)
  • Basic and Applied Basic Research Foundation of Guangdong Province (2025A1515010293)

This publication has 59 references indexed in Scilit: