Image restoration is a fundamental problem in low-level vision with applications in photography, surveillance, and archival imaging. This review synthesizes deep learning–based advances from 2018–2025, covering CNN (Convolutional Neural Network), GAN (Generative Adversarial Network), and Transformer families. This paper presents a unified taxonomy by degradation type (denoising, deblurring, super-resolution, dehazing/draining, JPEG deblocking, inpainting) and by model class. Representative architectures are summarized with their core design choices, training objectives, and computational characteristics. The comparison of common protocols, datasets (e.g., DIV2K, SIDD, GoPro, REDS), and metrics like PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and LPIPS (Learned Perceptual Image Patch Similarity are given also). The review analyzes trade-offs between fidelity, perceptual quality, and efficiency. The identification of open challenges, robust generalization to real-world degradations, efficient high-resolution inference, reliable perceptual evaluation, and unified multi-degradation handling, and outline research opportunities are also considered. This review aims to serve as a unified reference for practitioners and researchers developing next-generation image restoration systems.
Abed et al. (Sat,) studied this question.