Impact of Image Translation using Generative Adversarial Networks for Smoke Detection

Computer vision is a top-tier domain of the technological world that is responsible for automating the visual systems from healthcare to self-driving vehicles. With a reputation for surpassing human intelligence, it can be implemented in various trigger systems like wildfire smoke detection where the emission of smoke as a result of wildfire is fairly unpredictable. Low contrast and brightness have a detrimental effect on computer vision tasks. We present a novel approach to detect forest wildfire smoke, using image translation for converting nighttime images to day time which eliminates the confusion between smoke, cloud, and fog. This translation aids the YOLOv5 object detection algorithm to detect the smoke with the same aptness irrespective of time and lighting conditions. This paper demonstrates that the object detection model performs better on the images translated to day time with a better confidence score as compared to the corresponding nighttime images.