[M-GI34] Earth and planetary informatics with huge data management

[MGI34-P09] Super-resolution and colorization of small body images by Generative Adversarial Network

Rintaro Kishi¹, Ren Hashimoto¹, *Rie Honda¹, Kimikazu Tanase¹, Yasuhiro Yokota², Seiji Sugita³, Tomokatsu Morota³, Shingo Kameda⁴, Naoya Sakatani⁴, Eri Tatsumi⁵, Yuichiro Cho³, Toru Kouyama⁶, Manabu Yamada⁷, Masahiko Hayakawa², Hidehiko Suzuki⁸, Kazuo Yoshioka³, Chikatoshi Honda⁹, Kazunori Ogawa², Hirotaka Sawada² (1.Kochi University, 2.ISAS/JAXA, 3.Univ. of Tokyo, 4.Rikkyo Univ., 5.Instituto de Astrofisica de Canarias, Univ. of La Laguna, 6.AIST, 7.Chiba Inst. Tech, 8.Meiji Univ., 9.Univ. of Aizu)

Keywords:GAN, small body expoloration, super resolution, colorization

Super-resolution plays a big role in the choice of the landing site in the small body exploration. Conventionally, analytical techniques such as the reconstruction of higher resolution images using slightly shifted multiple images. In this study, we examined the method named SRGAN, based on Generative Adversarial Network, for estimation of a high-resolution image from the low-resolution image via learning and applied it to the images of asteroid Ryugu obtained by the ONC - T camera onboard Hayabusa2. The experimental result shows the improvement of images is confirmed visually and its calculation time is about 1sec for a 1024 x 1024 pixel image, which is fast compared with the conventional method.

In addition, we also examined the colorization of the single-band image by the method named pix2pix, which is also one of the methods of GAN. Multi-band images are utilized to speculate the type of surface material on the small bodies. However, obtaining multi-band images at the lowest altitude is difficult because the FOV of the cameras moves very fast relative to the surface. We also examined to generate pseudo color images from single-band images based on the learning of the pairs of single band images and pseudo color images. We will report the result of these experiments for asteroid Ryugu images obtained by Hayabusa2’s ONC-T.

Both methods are expected to help to make the decision-making process during the active small body exploration smooth by reducing the time required for analysis and improving the limited information at the maximum.