Core Concepts
Deep learning has enabled significant advancements in single-image super-resolution, with numerous methods continuously pushing the state-of-the-art forward. This survey provides a comprehensive overview of deep learning-based SISR techniques, categorizing them based on their specific design targets.
Abstract
This survey provides a thorough overview of deep learning-based single-image super-resolution (SISR) methods. It first introduces the problem definition, research background, and significance of SISR.
The survey then covers related works, including benchmark datasets, upsampling methods, optimization objectives, and image quality assessment methods. It categorizes deep learning-based SISR methods into three main groups: Simulation SISR, Real-World SISR, and Domain-Specific Applications.
For Simulation SISR, the methods are further divided into three subcategories based on their design targets: Efficient Network/Mechanism Design, Perceptual Quality, and Additional Information Utilization. The survey discusses the key contributions and innovations within each subcategory, such as residual learning, global and local residual learning, perceptual losses, and the use of additional information like edge priors.
The survey also presents reconstruction results of classic, latest, and state-of-the-art SISR methods to help readers understand their performance. Finally, it discusses remaining issues in SISR and outlines future trends and directions for the field.
Overall, this survey provides a comprehensive and structured overview of the rapidly evolving field of deep learning-based SISR, which can help researchers better understand the latest advancements and inspire future research.
Stats
PSNR is the most widely used metric to evaluate image reconstruction accuracy, which is highly correlated with MSE.
SSIM measures the similarity between two images on a perceptual basis, including structures, luminance, and contrast.
LPIPS and DISTS are popular metrics used to measure the perceived differences between images, reflecting the sensitivity of the human eye.
NIQE is a completely blind image quality assessment method that does not require any training data.
The Perception Index (PI) is a combination of the no-reference image quality measures Ma and NIQE, used to evaluate perceptual quality.
Quotes
"Recently, deep learning (DL) has demonstrated better performance than traditional machine learning models in many artificial intelligence fields, including computer vision and natural language processing."
"DL can transfer the SISR task to an almost end-to-end framework incorporating all these three processes, which can greatly decrease manual and computing expenses."
"This target-based survey has a clear context hence it is convenient for readers to consult."