HEA-GAN for high-fidelity low-light image enhancement and degradation-aware image augmentation using a hybrid deep learning and GAN framework

Abstract Poor vision and low-light conditions have a deleterious effect on the functioning of downstream computer-vision systems used in surveillance, autonomous driving, biomedical imaging, and remote sensing. Histogram equalization, Retinex based methods and manually crafted denoising methods are classical methods of enhancement that often do not preserve high-frequency structures, and can cause colour artifacts or other undesirable artefacts. In the last ten years, deep learning and Generative Adversarial Networks (GANs) have become good paradigms of low-light image enhancement, restoration, and synthetic data augmentation. The present manuscript suggests a hybrid deep-learning-and-GAN model capable of improving the image quality of low-light or low-quality images, and realistically augmenting the training data with the ability to control the illumination, noise, and blur levels. After that, we introduce the suggested architecture and loss design, and the training strategy and compare in a conceptual way its expected performance with modern state-of-the-art approaches on benchmark datasets like LOL, SID, ExDark, and DICM, in the metrics of PSNR, SSIM, LPIPS, and downstream detection. Recent experimental results have found that hybrid GAN trained models are capable of producing 1–3 dB PSNR and 0.02–0.05 SSIM extra improvements over convolutional base models and also produce more realistic textures and sturdier augmented samples. The proposed structure is a conceptual framework to be used for end-to-end enhancement and augmentation systems in demanding low-light conditions.