High-Frequency Normalizing Flow for Image Rescaling.

Area of Science:

• Computer Vision
• Digital Image Processing
• Machine Learning

Background:

• Image rescaling is crucial for digital image transmission across devices with varying resolutions.
• Image downscaling inevitably leads to high-frequency information loss, making reverse upscaling an ill-posed problem.
• Current joint learning methods for image rescaling struggle to recover satisfactory high-frequency signals during upscaling.

Purpose of the Study:

• To develop an efficient image rescaling method that compensates for information lost during downscaling.
• To improve the recovery of high-frequency details in upscaled images.
• To propose a novel framework capable of learning the distribution of high-frequency signals.

Main Methods:

• Proposed High-Frequency Flow (HfFlow), an invertible framework utilizing a conditional flow on the high-frequency space.
• Introduced a reference low-resolution (LR) manifold to guide the upscaling process.
• Developed a cross-entropy Gaussian loss (CGloss) to align downscaled manifolds and recover details.

Main Results:

• HfFlow effectively learns and restores high-frequency signal distributions lost during downscaling.
• The proposed CGloss facilitates optimal upscaling solutions by aligning LR manifolds.
• HfFlow demonstrates superior performance in restoring rich high-frequency details compared to state-of-the-art methods.

Conclusions:

• HfFlow offers a robust solution for image rescaling, significantly improving high-frequency detail recovery.
• The framework's invertible nature and novel loss function address the ill-posed nature of upscaling.
• HfFlow shows potential for generalization to other image transformation tasks like colorization.