Abstract: Surface sampling is an important approach in a lunar exploration mission. Intelligent sample identification will greatly contribute to enhance the efficiency and abilities of dealing with complex problems. The characteristics of the robotic arm camera (RAC) imaging are analyzed with the lunar surface excavation workflow. Imitating human identification process, the decoupled intelligent identification workflow is proposed, which has a hierarchical structure. And a class of deep convolutional neural network (CNN) is constructed for soil sample identification with deep learning method. The forward and backward relations of the network are deduced between image, feature and label space. The intelligent identification method is verified in ground surface excavation experiments. The simulation results indicate that the proposed method has preferable generalization in different illumination, scenes, processes, and sample shapes. The statistical error identification rate is less than 8.1%, and the average single image identification time is about 0.7 second.

Key words: Lunar surface sampling, Intelligent identification, Deep learning, Convolutional neural network