Lunar Crater Detection based on Grid Partition using Deep Learning

Although a digital elevation model (DEM) is often used for crater classification, the DEM has low resolution and thus may not reveal small craters. This research showed that a small crater found in images of the lunar south pole region, taken by the Narrow Angle Camera of the Lunar Reconnaissance Orbiter, had higher resolution than DEM. This research adopted a convolutional neural network (CNN) for crater classification. However, CNN had the problem of not being able to classify multiple craters at once due to its fixed length output. In this research, by dividing the images into grids, equaling output size/input using a semantic segmentation technique, and detecting the position of each crater in each grid using an object detection technique, we could classify each crater even if appearing in only one pixel from a large image with a numerical value and high accuracy rate. The recall rate and maximum precision rate was 80.5% and 77.3% respectively.