A comprehensive review of deep learning-based single image super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image super-resolution (SR) is crucial for enhancing image detail in computer vision.
• Significant progress in SR over two decades, particularly with deep learning.
• Classical SR methods laid the groundwork for modern techniques.

Purpose of the Study:

• To provide a comprehensive survey of recent single-image super-resolution (SR) progress.
• To detail deep learning-based SR methods and compare them with classical approaches.
• To categorize SR methods and discuss challenges, metrics, and datasets.

Main Methods:

• Classification of SR methods into classical, supervised learning, unsupervised learning, and domain-specific.
• Review and evaluation of state-of-the-art deep learning SR models.
• Introduction to SR problem, image quality metrics, datasets, and challenges.

Main Results:

• Deep learning methods have significantly advanced image super-resolution capabilities.
• Categorization provides a structured overview of diverse SR techniques.
• Evaluation highlights the performance of leading SR models like EDSR, CinCGAN, MSRN, Meta-RDN, RBPN, SAN, SRFBN, and WRAN.

Conclusions:

• Deep learning represents the forefront of image super-resolution research.
• Future research should address open problems and explore emerging trends in SR.
• This survey offers a valuable resource for researchers in the field of image super-resolution.