关键词: 月面重建, 深度感知, 立体匹配, 深度学习, 轻量化网络结构设计

Abstract: Optical images acquired in lunar surface scenarios suffer from significant weak texture, low contrast, and local shadow interference, which severely affect the stereo image matching accuracy of lunar rovers. This paper proposes a confidence-guided monocular-stereo collaborative stereo matching method for lunar rover images. Built on shared geometric-semantic encoding, the method unifies monocular structural priors and stereo matching in the same representation space, and enables adaptive cross-modal interaction through confidence-gated modulation. On this basis, it further combines disparity-hypothesis-driven stereo reasoning with a coarse-to-fine refinement mechanism, which effectively reduces feature redundancy and cross-modal interaction cost while preserving the accuracy of geometric estimation. The method is trained and quantitatively evaluated on the LuSNAR simulated lunar surface dataset, and is directly tested on images captured by the Chang'e-4 rover CE-4 panoramic binocular camera. The results show that the proposed method achieves an average endpoint error EPE of 0.43 and a D1 error of 1.42% on the LuSNAR validation set, outperforming domestic and international comparison methods. It also shows strong scene adaptability and cross-domain generalization in real-data experiments on CE-4 images. In addition, the efficiency analysis shows that, compared with the default baseline configuration, the proposed method reduces the number of model parameters from 346.39M to 36.69M, FLOPs from 42.51G to 11.45G, and forward inference time from 4.163s to 0.153s. These results indicate that the proposed method significantly improves computational efficiency while maintaining matching accuracy, showing its potential for in-orbit applications.

Key words: lunar surface reconstruction, depth perception, stereo matching, deep learning, lightweight network architecture