Abstract: With the rapid development of science and technology such as computer and artificial intelligence, there are a variety of intelligent image matching models based on deep learning. However, these models are mostly used for matching images with relatively regular texture features (such asportraits, building images,industrial parts, etc), and there is a lack of applicability research in map image matching with complex and diverse texture features. To this end, by taking three types of maps, namely the administrative division map, traffic map and thematic map as data sources, this study compares the map matching performances of the currently popular three image matching models, such as Superglue, COTR and GlueStick, by integrating semantic segmentation model (SAM). The visual matching performance, matching accuracy and matching efficiency results show that:①The boundary matching performance of GlueStick model is the best among the three types of maps, followed by Superglue, while COTR model has the worst matching performance.②Using SAM to extract map segmentation mask images is able to reduce peripheral features of the map, and further improve the matching performance of GlueStick and Superglue models where the accuracy are increased by 47.50% and 34.43%, respectively.③The matching efficiency of the COTR model is the lowest. While the matching efficiency of GlueStick model is lower than Superglue using the original map, their matching efficiency is comparable using the map segmentation masks. This study has important application values for contrast, recognition and review of map content.

Key words: map image matching, deep learning, Superglue, COTR, GlueStick