Vision Transformers in Image Restoration: A Survey.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) historically dominated computer vision tasks.
• Vision Transformer (ViT) architectures have emerged as powerful alternatives for image restoration.
• Both CNNs and ViTs offer efficient methods for enhancing low-quality images.

Purpose of the Study:

• To extensively study the efficiency of Vision Transformer (ViT) architectures in image restoration.
• To classify ViT architectures across various image restoration tasks.
• To detail outcomes, advantages, limitations, and future research directions for ViTs in image restoration.

Main Methods:

• Classification of ViT architectures for specific image restoration tasks.
• Analysis of ViT performance across seven key restoration areas: Super-Resolution, Denoising, Enhancement, JPEG Artifact Reduction, Deblurring, Adverse Weather Removal, and Dehazing.

Main Results:

• ViT integration into image restoration architectures is becoming standard practice.
• ViTs demonstrate advantages over CNNs, including better efficiency with more data, robust feature extraction, and superior learning of input characteristics.
• Identified limitations include a need for substantial data, higher computational costs, complex training, and lack of interpretability.

Conclusions:

• ViT architectures are highly effective for image restoration, often outperforming CNNs.
• Addressing ViT's current limitations, such as data requirements and computational complexity, is crucial for future advancements.
• Future research should focus on enhancing ViT efficiency and interpretability in the image restoration domain.