Deep learning for crescent detection and recognition: Implementation of Mask R-CNN to the observational Lunar dataset collected with the Robotic Lunar Telescope System

The ability of the human eye to identify a crescent depends on its apparent object contrast versus the sky background, and inaccurate assessments are common when identifying it. The use of telescopes and cameras to monitor the crescent moon is becoming increasingly important as technology advances. Thus, in this study we developed an automated moon detection system with deep learning and integrated for the robotic telescope OZT-ALTS with an infrared camera. By utilizing a deep learning method called Mask R-CNN, we have created infrared camera software with the goal of identifying and recognizing the crescent moon. The result shows, a total of 3,202 manually annotated moon images were used for deep-learning-trained models. We tested several combinations of training hyperparameters and image distribution numbers. The results show that the crescent detection issue can be resolved using a Mask R-CNN. Using the top-performing Mask R-CNN configuration, the trained model achieved a mean averaged precision (mAP) at the intersection over union (IOU) of 0.5, with a 99% for the extreme condition of a young crescent concealed by clouds and a 99% for the normal case for each moon phase. We also show that such systems can be utilized as a framework for future monitoring, detection, and recognition of the young crescent and all moon phases.