Single image super-resolution with multiscale similarity learning.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Learning-based super-resolution (SR) methods effectively generate high-resolution (HR) images using external training sets.
• The performance of these methods is highly dependent on the consistency between training data and the target low-resolution (LR) images.
• Incompatible high-frequency details in training sets can negatively impact SR results.

Purpose of the Study:

• To propose a single image SR approach that mitigates the reliance on external training data.
• To leverage multiscale self-similarities within the LR image itself for detail synthesis.
• To improve the accuracy and robustness of super-resolution reconstruction.

Main Methods:

• Learning multiscale self-similarities from the LR image by observing patch repetitions across and within scales.
• Establishing HR-LR patch pairs using the initial LR input and its down-sampled version.
• Utilizing neighbor embedding for LR-HR relationship estimation.
• Employing a gradual magnification scheme and accumulating results for iterative upscaling.
• Applying nonlocal means and a nonlocal prior regularization term for edge preservation and artifact suppression within a reconstruction-based SR framework.

Main Results:

• The proposed method achieves compelling SR recovery, outperforming state-of-the-art baselines.
• Quantitative and perceptual evaluations demonstrate the effectiveness of the approach.
• The method successfully synthesizes missing high-frequency details by exploiting internal image redundancies.

Conclusions:

• The proposed single image SR method effectively reconstructs high-resolution images by learning internal multiscale self-similarities.
• This approach offers a robust alternative to methods requiring external training datasets.
• The technique demonstrates significant improvements in both quantitative metrics and perceptual quality of super-resolved images.