Automatic Detection of Lunar Rocks Using Single Shot Multibox Detector

Lunar rocks are the important targets with high scientific values for the sample return missions on the Moon. However, the manual identification of rock species in navigation images requires a large amount of time and a wealth of experience. Developing the autonomous capacity to classify extraterrestrial rocks could greatly improve the efficiency of sample collection for the manned or unmanned program to process massive images. Traditional image processing methods such as image edge detection cannot achieve the function of identifying rock types and recognizing the position. SSD (Single Shot MultiBox Detector) is a popular and robust object detection network that can detect, localize and classify objects. In this work, the SSD with deep transfer learning and VGG-16 architecture is used for automatic detection and classification of three lunar rocks. Moreover, the lunar rock images from NASA’s Lunar Sample Compendium are chosen as the training datasets, and manually labeled based on Labelimg. Through the study, the fine-tuned VGG-16 pre-training network shows a good performance in the classification of lunar rocks comparing with other base models. Overall, the proposed method may achieve a good accuracy in identifying rock types. It may provide a feasible solution to assist an autonomous sample collection task for the future space exploration.