D2BGAN: A Dark to Bright Image Conversion Model for Quality Enhancement and Analysis Tasks Without Paired Supervision

This paper presents an image enhancement model, D2BGAN (Dark to Bright Generative Adversarial Network), translate low light images bright without a paired supervision. We introduce the use of geometric and lighting consistency along with contextual loss criterion. These when combined multiscale color, texture edge discriminators prove provide competitive results. performed extensive experiments using benchmark datasets visually objectively compare our observed performance on real-time driving that are subject motion blur, noise, other artifacts. further demonstrated enhanced can be profitably used in image-understanding tasks. Images processed technique obtain best or second average scores for three different quality evaluation methods Naturalness Preserved Enhancement (NPE), Low Light Image (LIME), Multi-Exposure Fusion (MEF) datasets. Best also obtained LOw-Light (LOL) test set Berkeley Driving Dataset (BDD) D2BGAN. Face detection tasks DarkFace dataset show mAP (mean Average Precision) improvement from 0.209 0.301 improves 0.525 finetuning techniques adopted.