FDSR: An Interpretable Frequency Division Stepwise Process Based Single-Image Super-Resolution Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning excels in single-image super-resolution (SISR) but often lacks interpretability, limiting its use in critical fields like medical imaging.
• The 'black box' nature of current SR networks hinders trust and validation in high-stakes applications.

Purpose of the Study:

• To develop an interpretable deep learning framework for SISR operating in the image frequency domain.
• To enhance transparency and understanding of the image reconstruction process in super-resolution.

Main Methods:

• Introduced a frequency division module and a step-wise reconstruction method to process images based on frequency components.
• Developed a frequency division loss function to ensure specialized reconstruction modules (ReM) operate at specific image frequencies.
• Designed an interpretable subpixel upsampling process by deriving its inverse and creating a displacement generation module.

Main Results:

• The proposed network, even without the frequency division loss, achieved state-of-the-art performance in qualitative and quantitative evaluations.
• Incorporating the frequency division loss significantly improved network interpretability and robustness.
• A minor decrease in PSNR (0.48 dB) and SSIM (0.0049) was observed with the frequency division loss, indicating a trade-off between performance and interpretability.

Conclusions:

• The developed interpretable frequency division SR network offers a transparent alternative to traditional black-box models.
• This framework is particularly beneficial for sensitive applications like medical imaging where understanding the reconstruction process is vital.
• The study demonstrates a viable method for balancing high performance with enhanced interpretability in deep learning-based super-resolution.