Wavelet-Based Texture Reformation Network for Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Reference-based image super-resolution (RefSR) typically uses raw features from pretrained encoders.
• Directly using raw features can lead to suboptimal texture matching due to irrelevant information and loss of high-frequency details.

Purpose of the Study:

• To propose a novel Wavelet-based Texture Reformation Network (WTRN) for enhanced RefSR.
• To address limitations in current RefSR methods regarding texture transfer and matching accuracy.

Main Methods:

• Decomposing texture features into low-frequency and high-frequency sub-bands using wavelet transformation.
• Performing feature matching on low-frequency components and transferring wavelet-domain features.
• Introducing a wavelet-based texture adversarial loss for visually plausible texture generation.

Main Results:

• The WTRN method demonstrates superior performance compared to existing RefSR methods.
• Quantitative and qualitative experiments on benchmark datasets validate the effectiveness of the proposed approach.
• Improved texture matching and transfer capabilities were observed.

Conclusions:

• Wavelet transformation offers significant advantages for RefSR by capturing multi-scale contextual and textural information.
• The proposed WTRN effectively overcomes the limitations of raw feature utilization in RefSR.
• The method achieves state-of-the-art results in reference-based image super-resolution.