Super-Resolving Methodology for Noisy Unpaired Datasets.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Directly supervised learning for image super-resolution is often impractical due to real-world data acquisition challenges.
• Paired high-resolution and low-resolution datasets can be difficult to obtain and align accurately.

Purpose of the Study:

• To develop a practical methodology for image super-resolution under real-world conditions.
• To investigate the effectiveness of average denoising using multiple noisy images, including unpaired datasets.

Main Methods:

• Image noise reduction through averaging multiple noisy input images.
• Super-resolution reconstruction applied to denoised images.
• Evaluation using metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index map (SSIM).

Main Results:

• Increased number of images used for average denoising leads to improved super-resolution performance.
• Using 16 images for denoising outperformed using 4 images, yielding higher PSNR and SSIM values.
• Significant improvements were also observed in learned perceptual image patch similarity (LPIPS) and natural image quality evaluator (NIQE) metrics.

Conclusions:

• Average denoising is a practical and effective approach for image super-resolution in real-world scenarios.
• The proposed method demonstrates superior performance, particularly when handling unpaired and misaligned datasets.
• The findings highlight the benefit of incorporating more images for denoising to enhance super-resolution outcomes.