Performance Analysis of Enlighten GAN on Low-Light Enhancement and Denoising: Fid-Bau Portal

Performance Analysis of Enlighten GAN on Low-Light Enhancement and Denoising

Panwar, Moomal / Gaur, Sanjay B. C.

Deep learning-based techniques have achieved remarkable extraordinary performance in the field of image enhancement and restoration. These techniques are still competitive in the absence of paired training data despite their exciting performance for the above-mentioned areas. One such issue is investigated in this paper for enhancement of the low-light image, where taking a photo of the same scenic situation in both dim and normal light at exact time is actually extremely challenging. Retrained enlighten GAN (REGAN), a very successful network that is unsupervised and can be trained without dim-light and normal-light picture pairs but performs admirably on a variety of test images taken from various real-world scenarios, was a great answer to this issue. It suggests regularizing rather than guiding learning with ground truth data, the unsupervised training uses the data retrieved from the image itself. But if this model is given a training set of specially created low-light photos that can be made by adding desired noise and darkness, it was found that the suggested strategy performs better than prevailed methods. Extensive experiments on the five real datasets MEF, LIME, NPE, SCIE, and DICM show that the enhanced images of our method are closer to the actual one. Because of the tremendous flexibility, REGAN is proved to be more powerful tool to upgrade real-world photos from many domains.