Improving YOLO Performance on Low Light Images using Synthetic Data Generation

Low-light conditions significantly degrade the image quality, limiting the information available for detection. Even humans struggle to perceive the important details due to poor lightning, and since computer vision mimics human visual perception, it faces similar challenges. This issue is especially critical in surveillance systems, where time is of the essence and it can be life saving. Many researchers have given their solutions and have worked towards resolving the issue of object detection in low light images. The method proposed in this paper consists implementation of YOLOV8 and YOLOV9 for this purpose. Two types of datasets were generated one is an augmented dataset made by applying filters and transformations to stimulate night time scenes. Second is the synthetic dataset generated using DALL.E 3 and Runway. Comparison was made for both versions and YOLOV9 outperformed using Synthetic Dataset with 92 mAP which shows that it can be used in the case of low light conditions.