Deep Learning based Underwater Object Detection

Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs) equipped with an intelligent object detection system play a vital role in various underwater applications such as marine resource exploitation, marine environment monitoring, and marine cable protection. Deep learning based object detection methods have presented great performance advantages over traditional machine learning based methods. However, these deep learning based methods lack sufficient capabilities to handle underwater object detection (UOD) due to these challenges: (1) underwater images acquired in complicated environments suffer fromsevere distortion which dramatically degrades image visibility, objects in the underwater datasets and real applications are usually small whilst accompanying severe noise that greatly degrade the detection accuracy of UOD tasks. (2) well-annotated underwater data is not sufficient in terms of diversity and amount which highly influences the performance of deep learning models. (3) severely imbalanced data distribution and label noise distribtuion occur in underwater datasets, driving a deep learning model to be more biased towards the majority class.

In this thesis, we aim to address all these challenges, and develop robust deep learning systems to enhance and detect objects in complex underwater images. To achieve this goal, we firstly propose novel perceptual enhancement models to enhance the quality of underwater images. Secondly, we propose a novel Sample-WeIghted hyPEr Network (SWIPENET), and a robust training paradigm named Curriculum Multi-Class Adaboost (CMA), to address the noise and small object detection problems at the same time. Finally, to address the class imbalance problem, we propose a factor-agnostic gradient re-weighting algorithm (FAGR) that can adaptively fine tune the gradients of individual classes according to the distributions of their detection precision. We have evaluated the proposed methods by conducting extensive experiments on public datasets. Experimental results show the effectiveness of our methods for underwater image synthsis, image enhancement and object detection.