|
Vol.17 No.2previousAA184 |
|||||
| Academic Articles | |||||
| Regular Paper | Vol.17 No.2 (2025) p.15 - p.24 | ||||
Method for Creating Large Datasets for Deep Learning to Improve Image Depth Accuracy |
|||||
|
Masahiro MURAYAMA1,*,Yuki HARAZONO1,Hirotake ISHII1,Hiroshi SHIMODA1 and Yasuyoshi TARUTA2 |
|||||
|
1 Graduate School of Energy Science, Kyoto University, Sakyo-ku Yoshidahonmachi, Kyoto-shi, Kyoto 606-8501, Japan 2 Fugen Decommissioning Engineering Center, Japan Atomic Energy Agency, 3 Myojin-cho, Tsuruga-shi, Fukui, 914-8510, Japan |
|||||
| Abstract | |||||
|
High quality depth images are required for accurate 3D modeling of a facility. However, depth images captured using a typical commercially available RGB-D camera include much noise. Recently, methods using deep learning for depth enhancement have been developed. As described herein, we developed a novel method to create a high-quality dataset by generating high-quality depth images with pixel-wise depth enhancement, which is less affected by camera pose estimation errors. Furthermore, our method improves the quality of the entire dataset by post-processing suitable for our depth enhancement process. Comparison with the dataset created using the existing method showed that datasets created using the proposed method are suitable for training a network for depth enhancement. Depth images taken inside the Fugen Decommissioning Engineering Center are processed by a network trained on the dataset. The network completed the missing areas more correctly and removed the noise while maintaining the detail shapes. |
|||||
| Keywords | |||||
| dataset creation, deep learning, depth image, noise removal | |||||
|
Full Paper:
PDF
|
|||||
