Quality-driven end-to-end restoration system for underwater images using a self-supervised convolutional neural network

• Autor(a) principal: Walysson Vital Barbosa
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/44155
• topic: image restoration; underwater vision; convolutional neural networks; image quality metrics; Computação – Teses; Processamento de imagens -- Tecnicas digitas – Restauração e conservação -Teses; Visão subaquática – Teses; Redes neurais convolucionais – Teses

Advances in technology have allowed humans to delve into the depths of Earth and to study the outer space, even if our resources are not sufficient to help us answer all questions about each one of these environments. The underwater environment is one of those places, which has been vastly studied in past years due to the increasing use of underwater research locations. However, there are a few reasons why studying this environment is challenging. In most cases, structures located underwater cannot be moved out of this medium as they can lose their properties and be damaged. Moreover, images taken in these environments have very poor quality in comparison to images from out of water places. The water medium causes various effects during the image acquisition process. Rays of light are scattered and absorbed as they travel to the camera. This thesis proposes an underwater image restoration method based on convolutional neural networks and image quality metrics, the former being considered universal function approximators. Features extracted from the original underwater image are applied to the inverse image formation model in order to recover the original image radiance. No labeled data is needed as the network is trained based only in the quality metrics computed using the original and restored underwater images. In 60% of the cases, our proposed methodology performs better than the techniques applied to the improvement of underwater images, taking into consideration the UCIQE metric. Additionally, two underwater image datasets are presented, which were acquired on a planned process, focusing on underwater image restoration purposes.