Recovering Realistic Details for Magnification-Arbitrary Image Super-Resolution

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Implicit Neural Representations (INRs) enable continuous image representation by mapping coordinates to RGB values.
• Existing methods can super-resolve images from low-resolution inputs but often produce blurry results lacking fine details.

Purpose of the Study:

• To introduce Implicit Pixel Flow (IPF) for modeling coordinate dependencies between blurry and sharp image distributions.
• To enable the recovery of perceptually-pleasant details in arbitrary-resolution single image super-resolution.

Main Methods:

• IPF assigns coordinate offsets to pixels near blurry edges, replacing original RGB values with those from neighboring pixels for sharper edges.
• Convolutional neural networks extract continuous flow representations, and multi-layer perceptrons build the implicit function for pixel flow calculation.
• A novel double constraint module is proposed for stable and optimal pixel flow searching during training.

Main Results:

• The method successfully converts blurry INR distributions to sharp ones by modifying coordinate-to-pixel relationships.
• Experimental results on benchmark datasets demonstrate the restoration of sharp shape edges and textures.
• This is the first method to achieve perceptually-pleasant details in arbitrary-resolution single image super-resolution.

Conclusions:

• IPF effectively enhances image quality by generating sharper details and textures.
• The proposed method advances the field of single image super-resolution, particularly for arbitrary magnification.
• IPF offers a novel approach to refining continuous image representations for improved visual fidelity.