Semantic Segmentation of Martian Terrain Based on Dual-Branch

Abstract:
With the rapid development of computing, neural network technology has been widely used in various fields. Recently, the aerospace field has gradually become a symbol of scientific and technological progress in various countries. In the aerospace field, the autonomous movement of the planetary vehicle is the most important step in Mars exploration. How to accurately identify the planet’s topography is the main factor in the advancement of the planetary vehicle. The training of Mars terrain segmentation is inseparable from the labeling of Mars terrain and landforms. The labeling boundaries are often not very accurate due to various reasons. At the same time, the real terrain ground samples of Mars are extremely limited. These cause problems for the division of Mars terrain. In this paper, a dual-stream network is designed for the semantic segmentation of Martian terrain. One branch recognizes the accurately labeled internal area of the object, and one branch recognizes the inaccurate boundary area. The two work together to achieve a perfect Martian terrain segmentation. The scheme proposed in this paper is compared with the more popular U-net, etc., and our scheme shows superior performance.