Image Feature Matching of the Moon Surface Based on Multi-task Learning

Abstract

Accurate and effective lunar surface feature matching is a key technology in lunar exploration projects and precise positioning of lunar surface targets. The traditional feature matching algorithm extracts and matches the feature points of the lunar surface image, because of its complicated calculation, single application, large descriptor dimension, long matching time, poor robustness to local similar regions, etc. There are certain limitations in the application scenarios in real life. Based on deep learning, this paper proposes an image feature matching based on multi-task learning. The method is based on the improved HF-Net network, adopting the lightweight MobileNetv3 network, and combining the transfer learning knowledge distillation method to achieve end-to-end line Feature point matching. Experiments show that this method can effectively reduce the complexity of calculation, improve the accuracy of feature matching, and has good robustness under different lighting conditions.