New Trends in Image Restoration based on Artificial Intelligent Models: Analytical Study
S. N. Abed, A. Mahmoud Al-Jawher · Journal Port Science Research · 2025
Typical tasks include denoising, deblurring, super-resolution, dehazing/deraining, JPEG artifact removal, and inpainting each with distinct priors but shared challenges of balancing perceptual plausibility and signal fidelity.Recent works frame restoration as learning powerful priors that generalize across degradations and resolutions [1].Between 2018 and 2025, progress accelerated with deep learning.CNN-based methods (e.g., MIRNet, MPRNet, NAFNet) improved multi-scale feature aggregation and efficiency for diverse degradations[2][3][12].In parallel, GAN models advanced perceptual quality in super-resolution and face/real-image restoration (ESRGAN, Real-ESRGAN) [13][14].More recently, Transformers (IPT, SwinIR, Restormer, Uformer) brought stronger long-range modelling with windowed or efficient attention, achieving state-of-the-art results at high resolutions[1][8][15][16]. Evaluation relies on public benchmarks and metrics.Widely used datasets include SIDD and DND for real-image denoising, GoPro for motion deblurring, REDS for video SR/deblurring, and DIV2K for SR(Super-Resolution); common metrics are PSNR/SSIM for fidelity and LPIPS for perceptual similarity.These resources enable fair comparisons across tasks and architectures while revealing gaps between synthetic and realworld degradations[4][5][6][7].This review synthesizes these developments (2018-2025), offering: (i) a taxonomy by degradation type and model class (CNN, GAN, Transformer); (ii) a comparative summary of design choices and training objectives; (iii) a unified view of datasets, protocols, and metrics; and (iv) open challenges around robust real-world generalization, efficient highresolution inference, and reliable perceptual evaluationproviding a practical reference for designing next-generation restoration systems[8]. BACKGROUND AND TAXONOMY Problem FormulationImage restoration is commonly posed as an inverse problem(equation 1).The tasks include denoising, deblurring, super-resolution, dehazing/deraining, JPEG artifact removal, and inpainting; each defines a different 𝐻 (e.g., motion blur kernel for deblurring; scale/downsampler for SR, DCT quantization for JPEG, binary mask for inpainting).The evaluation typically balances fidelity (PSNR/SSIM) and perceptual quality (LPIPS) [6] [7].