Improved Mars Terrain Segmentation in Terms of Style Transfer and Network Architecture

Mars terrain segmentation facilitates full comprehension of Martian terrain and plays an important role in the autonomous planning of Mars rover missions. However, existing methods face several issues. Since Martian terrains vary greatly over different regions, segmentation performance tends to degrade severely because of differences from the training terrain. Furthermore, it becomes necessary to utilize unannotated data due to the difficulty of annotating data as required for the Mars terrain segmentation task. Finally, mis-segmentation of objects frequently occurs. To address these challenges, three key approaches are proposed in this paper: (1) We propose a new neural style transfer method named SA-CCPL, which converts images from source domain style to target domain style without losing the image details. (2) We implement a thing-class and stuff-class perception enhancement (TSPE) module, optimizing the network’s scene-comprehension capability. (3) An ensemble of multi-scale models (EMSM) pseudo-label construction method is proposed to improve pseudo-label reliability. Experiments demonstrate that our proposed method is effective in improving the performance of Mars terrain segmentation.