A preliminary study on visibility improvement of turbid underwater images for dismantling of nuclear facilities. (15th June 2021)

• Record Type: Journal Article
• Authors: Park, Seung-Kyu; Oh, Seong Yong; Shin, Jae Sung; Park, Hyunmin; Lee, Jonghwan
• Is Part Of: Annals of nuclear energy. Volume 156(2021)
• Publication Date: 2021-06-15
• Subjects: Video monitoring -- Underwater cutting -- Turbid underwater -- Deep learning neural network -- Visibility improvement -- Histogram-equalized image; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 621.4805
• DOI: 10.1016/j.anucene.2021.108207
• Languages: English
• ISSNs: 0306-4549

Abstract

Highlights: The visibility of blurred underwater images was improved by using an artificial neural network in a turbid underwater cutting environment. The conditional GAN neural network was adopted to improve the visibility of turbid underwater images. The neural networks were trained with images with the best visibility and images enhanced with software technology. Experiments demonstrated that the trained neural network provided significantly improved visibility in turbid images. Abstract: Video monitoring environment of underwater cutting objects is rapidly deteriorated due to the increasing water turbidity in proportion to cutting operation. Precise monitoring of the cutting area is one of keys to improve work efficiency. In this paper, we studied a visibility improvement technique based on an artificial neural network for monitoring objects in a turbid underwater cutting site. In order to improve visibility in turbid water, a deep learning neural network that learned two types of visibility enhancement process was adopted. The first type is to train to restore an ideal underwater image with the best visibility from a turbid underwater image, and the second type is to train to restore an image with improved visibility using the conventional visibility improvement technique. We adopted two types of training real images based on the GAN (generative adversarial networks) model for the corresponding turbid input images. The first training image is the histogram-equalized image of a clear underwater image. If the first training image does not exist, an image with improved visibility by using histogram equalization for the turbid input image itself was used as the second training image. Experiments demonstrated that the trained neural network provided significantly improved clarity in turbid images compared to that of the conventional improving technique.