Robust Single-Image Super-Resolution via CNNs and TV-TV Minimization

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Single-image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) input.
• Convolutional Neural Networks (CNNs) achieve state-of-the-art performance in SISR using large datasets.
• Current CNN-based methods lack consistency, as downsampling the generated HR image does not yield the original LR image.

Purpose of the Study:

• To address the inconsistency issue in CNN-based super-resolution.
• To propose a post-processing method that enforces consistency between HR and LR images.
• To improve the robustness of super-resolution to downsampling operator mismatch.

Main Methods:

• Post-processing CNN outputs using a novel optimization problem termed TV-TV minimization.
• Enforcing consistency between the super-resolved HR image and the original LR image.
• Evaluating image quality using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Main Results:

• TV-TV minimization significantly improves image quality, as measured by PSNR and SSIM.
• The proposed post-processing enhances the robustness of super-resolution against variations in the downsampling operator.
• Ensures that downsampling the output HR image precisely reconstructs the input LR image.

Conclusions:

• TV-TV minimization is an effective post-processing technique for single-image super-resolution.
• This method enhances both image fidelity and the practical applicability of super-resolution models.
• The approach offers a solution for operator mismatch problems in real-world super-resolution applications.