Host: The Japan Society of Mechanical Engineers
Name : Proceedings of FDR2024 International Topical Workshop on Fukushima Decommissioning Research
Date : October 10, 2024 - October 13, 2024
In this paper, we describe the results of a feasibility study of a noise reduction method from images using deep learning technology for decommissioning work. Currently, remotely operated robots have been used for the decommissioning work at the Fukushima Daiichi Nuclear Power Station (FDNPS) due to the high radiation environment. we have been conducting research and development for providing clear images during operations by removing only noise from images containing noise to contribute to safe and secure decommissioning work. Since we do a feasibility study of the noise reduction method using deep learning, the main target is not the video, but rather images, which are components of the video.
We adopted the approach of building a learning model that can cope with various types of noise by training many noisy images in the deep learning process. In particular, a network called Noise2Noise was used in training to create a model to remove noise in the images. As a result of the noise reduction process, we confirmed the noise in the images was reduced. To quantitatively verify the effect of the noise reduction method applied to the images, a quantitative measure of image quality was calculated by Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE), which is a kind of No-Reference Image Quality Assessment. As a result, the BRISQUE scores of the images improved after the noise reduction process. This result suggests that the processing was achieved without any loss of image quality.
