Session: 06-14-02 Underwater Vehicles and Design Technology II

Paper Number: 79891

79891 - Artificial Underwater Dataset: Generating Custom Images Using Deep Learning Models 

Abstract 

The offshore industry's rapid developments in underwater operations have led to requirements for more sophisticated Underwater Vehicles. Many underwater operations such as inspection and monitoring of underwater structures are done with the help of Autonomous Underwater Vehicles (AUVs). Recently, the AUV industry has focused on exploiting the advances made in the area of Artificial Intelligence and particularly Deep Learning (DL), but despite the developments that have been made, the underwater sector has not yet fully adopted the DL models and algorithms when compared with other sectors such as Unmanned Aerial Vehicles (UAVs) and Autonomous Cars. This has mainly been because of the challenges of the underwater environment, such as low light conditions and blurred images. 

Autonomous Underwater operations using DL and Computer Vision (CV) methods currently face a significant lack of readily available underwater datasets, particularly datasets for specific tasks. Research has shown that publicly available datasets include only marine life images and the underwater environment, which in most of the inspection and monitoring cases of subsea infrastructure is not sufficient.  Furthermore, the lack of readily available and versatile underwater image datasets can affect the progress of projects at the initial stages of development. This study aims to investigate how to create a custom image dataset that will approximate the natural environment for the specific underwater application using images taken in controlled conditions and the current developments in DL and CV. Firstly, objects that can be found in underwater structures such as pipes, bolts, nuts, flanges and anodes were photographed, and a dataset created. Deep Learning Augmentation libraries were used to perform data augmentation on the custom dataset to train an object detection model. More specifically, the augmentation libraries performed random image saturation, exposure and noise, and random grey-scale transformations.   

A modified version of Generative Adversarial Networks (GANs) was used to generate underwater images by combining images taken in air and images with underwater environment characteristics. The model introduces a loss in the discriminator (Discriminator Loss), which is used as the classifier to determine whether the generated image is close to the underwater environment. Specifically, the CycleGAN model takes as the first input, the custom dataset, and then the second input is images of the underwater environment that can be found in publicly available datasets. The output images were used in an object detection model. Finally, the comparison of the results between the data augmentation using libraries and DL models, showed that the images generated using the CycleGAN models have features closer to the actual underwater environment and allow the creation of a custom underwater dataset specifically for the required application.   

The study shows that, for projects in the early stages of development, it is possible to generate a custom underwater dataset for specific cases where data are scarce and difficult to acquire, with results close to the real-world environment.   

Presenting Author: Ioannis Polymenis Newcastle University

Authors:

Ioannis Polymenis Newcastle University
Maryam Haroutunian Newcastle University
Rose Norman Newcastle University
David Trodden Newcastle University