LumiNet: Multispatial Attention Generative Adversarial Network for Backlit Image Enhancement
Samprit Bose, Sahil Nawale, Dhruv Khut, Maheshkumar H. Kolekar
Abstract
Backlit image enhancement is a crucial task in improving the quality and visibility of the underexposed regions in an image caused by the difference in illumination between the background and foreground. Traditional methods struggle to effectively handle the dynamic range compression required for backlit image enhancement and fail to properly balance the exposure between the background and foreground areas. In this paper, we propose a novel backlit image enhancement architecture using a modified UNet and a unique 1x1 discriminator-based conditional generative adversarial network. Our approach incorporates custom hyperparameters, a tailored loss function, and bilateral guided upsampling for efficient enhancement of large images. We evaluated our model on the BAID dataset and achieved superior results in various image evaluation metrics like PSNR, NIQE and ΔE°, demonstrating its effectiveness in enhancing both backlit and low-light images. We have also compared the computational efficiency and resources used by our method with other state-of-the-art methods. Extending our work, we have incorporated our model into a pipeline for enhancing backlit traffic images and videos, which is then used to detect the license plates of vehicles with improved accuracy.