Дослідження сучасних методів поліпшення якості фотозображень, зроблених у поганих умовах, з використанням штучного інтелекту

Abstract

This work examines AI methods for enhancing images captured in poor conditions – low light, noise, blur. It reviews major deep-learning approaches, including CNNs, GANs, and Retinex-inspired models. A hybrid attention-based UNet is highlighted, reaching about 28.4 dB PSNR and 0.88 SSIM on benchmark datasets. The work compares supervised, zero-shot, and unsupervised strategies, noting their key trade-offs. It also discusses PSNR/SSIM and the gap between metric scores and real downstream-task performance. The main goal is to identify the most effective real-time enhancement methods for edge devices