09:00 〜 10:30
[PPS04-P12] Arbitrary-scale super-resolution of UVI images based on a recent deep learning technique
Both the Earth and Venus are believed to have appeared about 4.6 billion years ago. However, Venus today has no ocean and the upper atmosphere, unlike on the Earth, rotates at a speed of 60x faster than the surface. To reveal those differences by concrete evidence, the JAXA Venus explorer “Akatsuki”, was launched on May 21, 2010, and began its observation on Venus in 2015. However, because of Akatsuki’s elongated orbit, every observation image has different spatial resolution on Venus. Only few images are obtained with sufficient higher-resolution (HR), say, 15 km/pixel = 800 pixels disk size in every orbit, while many of Venus images have much lower resolution.
Here, image super-resolution is a task of enhancing the resolution of an image from lower-resolution (LR) to higher-resolution (HR), and then make the LR images show as much as possible the atmospheric features of Venus like HR images. Since the traditional image super-resolution models can only deal with a certain and discrete super-resolution scale determined by the number of post-up-sampling layers and require the dataset have the same scale parameter, images from Akatsuki, which can be different spatial resolution on Venus, need to be processed in an arbitrary scale image super-resolution model.
We introduced one of the arbitrary scale image super-resolution models, Local Implicit Image Function (LIIF), into the processing of Akatsuki dataset. Besides regular processing method which includes extraction of Venus desk, we divided Venus disk brightness by simple brightness model (Lambertian diffusion) to normalize large-scale pattern depending on solar incident angle. We also performed high-pass filtering for focusing on rather small scales cloud features of Venus atmosphere. Then, we designed cross-type down-sampling path by Gaussian blur and pyramid down-sampling to mimic image blurring in the actual optics of UVI. After that, we train the LIIF model and LIIF-GAN. As a result, our model help on super-resolution from LR images to HR images and reconstruction of the missing details of Venus atmosphere due to the uneven resolutions.
Here, image super-resolution is a task of enhancing the resolution of an image from lower-resolution (LR) to higher-resolution (HR), and then make the LR images show as much as possible the atmospheric features of Venus like HR images. Since the traditional image super-resolution models can only deal with a certain and discrete super-resolution scale determined by the number of post-up-sampling layers and require the dataset have the same scale parameter, images from Akatsuki, which can be different spatial resolution on Venus, need to be processed in an arbitrary scale image super-resolution model.
We introduced one of the arbitrary scale image super-resolution models, Local Implicit Image Function (LIIF), into the processing of Akatsuki dataset. Besides regular processing method which includes extraction of Venus desk, we divided Venus disk brightness by simple brightness model (Lambertian diffusion) to normalize large-scale pattern depending on solar incident angle. We also performed high-pass filtering for focusing on rather small scales cloud features of Venus atmosphere. Then, we designed cross-type down-sampling path by Gaussian blur and pyramid down-sampling to mimic image blurring in the actual optics of UVI. After that, we train the LIIF model and LIIF-GAN. As a result, our model help on super-resolution from LR images to HR images and reconstruction of the missing details of Venus atmosphere due to the uneven resolutions.