2022 年 9 巻 2 号 p. 278-288
With the development of 3D scanning technology, complex objects in the real world can be recorded efficiently and accurately in the form of large-scale point cloud data. In the scanned point clouds, sharp edges such as contours and soft edges of undulating surfaces enable us to better understand the structure of the complex point clouds. In our previous work, we proposed a deep learning-based point cloud upsampling network that can generate new points to improve the point density in the edge regions, and then combined with our novel transparent visualization method which can effectively improve the visibility of the edge regions in the 3D-scanned point cloud. However, in our previous work, we mainly focused on sharp edges, and for soft edge regions, although the visibility can be improved to some extent, it can be further improved. In this paper, we optimize the upsampling network for soft-edge regions so that the characteristics of the soft-edge regions, such as the point density gradation, can be maintained while upsampling. Moreover, we apply the optimized upsampling network to real scanned data containing numerous soft edges to verify the performance of the upsampling network in the soft edge regions. Experimental results show that the optimized network can maintain the characteristics of soft edge regions while upsampling in the upsampling task for soft edge regions, thus improving the visibility of soft edge regions in transparent visualization more effectively.