Low Light Image Enhancement for Autonomous Vehicle Applications

The proposed work aims to improve the quality and visibility of low-light images for autonomous vehicle applications in poor lighting conditions. Image enhancement in poor-lighting conditions is essential in improving the visual quality of low-light images for a diverse spectrum of image processing field of applications including night-time photography, autonomous-driving, underwater, security-surveillance systems and medical-imaging. Yet, capturing high-quality images in low-light conditions becomes challenging as the images are often blurry, noisy, low-contrast and with considerable noise. The described challenges create difficulties in tasks including object detection, scene-visibility and tracking. The images of road scenes in low-light environment often exhibit lack of natural-details, potentially increasing the risk of accidents. Enhancing low-light images during night and foggy conditions will enhance visibility and improve driver's overall experience. However, existing research on low-light image enhancement (LLIE) has major gaps in recovering fine details from nighttime driving-scenes. In the study, the challenge of low-light image enhancement (LLIE) is addressed by employing Zero-Dce (Zero-reference deep curve estimation) model to enhance night-time driving images from the dark-zurich dataset. The zero-dce model enhances low-light images by preserving structural details and reducing noise making it highly suitable for real-world autonomous driving applications. Experimental results demonstrate a superior performance of the proposed zero-dce network exhibiting significant improvements in contrast and detail-recovery of images leading to an accuracy of 81.96% when tested on test images from dark-zurich dataset, demonstrating it's effectiveness compared to other networks. Furthermore, a comparative study with other state-of-art models such as mirnetv2, basic-cnn, lime and traditional histogram-equalization methods have been evaluated in the study on the dark-zurich dataset, highlighting the proposed zero-dce netork's superior performance in enhancing night-time driving scenes. The findings obtained suggest that the proposed approach contributed to a more reliable method of LLIE in autonomous-driving applications.