Deep Learning (DL) has transformed computer vision, leading to significant
progress in areas like autonomous vehicles and industrial automation. However, its application in underwater environments remains challenging due to factors such as light absorption, scattering, and water turbidity, which degrade image quality and hinder DL model performance. This thesis addresses these challenges by enhancing DL-based computer vision techniques for underwater scenarios, with a focus on autonomous robotic litter collection from the seabed. The work targets key limitations such as data scarcity, visual degradation, and scene variability, and proposes domain-informed approaches to enhance model generalization. The contributions cover the full DL pipeline, starting with the design of representative training data that supports object detection in shallow-water conditions, providing a benchmark for training and evaluation of detection algorithms. A multi-robot system is developed that integrates aerial, surface, and underwater vehicles to perform collaborative litter detection and collection. The thesis presents the system design, deployment, and the role of computer vision in the operational workflow. To address image degradation, an automated framework is proposed for selecting image enhancement methods based on task-specific performance metrics. Furthermore, environment-specific neural networks are introduced to handle variability in turbidity and lighting. Generalization to Out-of-Distribution (OOD) data is further addressed through a hybrid classification framework that combines a Convolutional Neural Network (CNN) with a physics-based classifier using the Moving Horizon Estimation (MHE) framework. Their outputs are fused via Dempster-Shafer theory to enable decision-making in unfamiliar scenarios. Finally, domain-informed neural networks are proposed to integrate physics-based knowledge into the DL pipeline via knowledge distillation. This method improves generalization and reduces dependence on large labeled datasets. The proposed methods are validated through simulation and real-world deployments, demonstrating improved performance and adaptability. Together, these contributions provide an integrated framework for deploying DL-based perception systems in challenging underwater environments.

Marine litter pollution is a major environmental threat due to the widespread presence of plastics and their detrimental impact on marine life and human health. There is a need for autonomous systems with computer vision to help clean the oceans. This study compares the latest state-of-the-art You Only Look Once (YOLO) models YOLOv9 - YOLOv12 in an underwater object detection setting in terms of accuracy, computational speed, and architecture complexity. We specifically focus on the smallest versions of these architectures, due to the real-time constraints of the setting. Multiple underwater datasets are combined to obtain a wide representation of underwater conditions and marine objects. The findings provide valuable insights into selecting and optimizing object detection architectures for underwater litter detection, contributing to monitoring marine ecosystems and addressing marine pollution. This work can be used as a building ground for further improving underwater object detection systems.

Robust object detection is crucial for automating underwater marine debris collection. While supervised deep learning achieves state-of-the-art performance in discriminative tasks, replicating this success on underwater data is challenging. The generalization of these methods suffers due to a lack of available annotated data considering different sources of variation in the unstructured underwater environment and imaging conditions. In this paper, we present the Seaclear Marine Debris Dataset, the first publicly available shallow-water marine debris dataset annotated for instance segmentation/object detection. The dataset contains 8610 images collected using ROVs at multiple locations and with different cameras, annotated for 40 object categories, encompassing not only litter but also observed animals, plants, and robot parts. As part of the technical validation, we provide baseline results for object detection using Faster RCNN and YOLOv6 models. Furthermore, we demonstrate the non-triviality of generalizing the trained model performance to unseen sites and cameras due to domain shift. This underscores the value of the presented dataset in further developing robust models for underwater debris detection.

The majority of computer vision architectures are developed based on the assumption of the availability of good quality data. However, this is a particularly hard requirement to achieve in underwater conditions. To address this limitation, plenty of underwater image enhancement methods have received considerable attention during the last decades, but due to the lack of a commonly accepted framework to systematically evaluate them and to determine the likely optimal one for a given image, their adoption in practice is hindered, since it is not clear which one can achieve the best results. In this paper, we propose a standardized selection framework to evaluate the quality of an underwater image and to estimate the most suitable image enhancement technique based on its impact on the image classification performance.

Deep learning has enabled the rapid expansion of computer vision tasks from image frames to video segments. This paper focuses on the review of the latest research in the field of computer vision tasks in general and on object localization and identification of their associated pixels in video frames in particular. After performing a systematic analysis of the existing methods, the challenges related to computer vision tasks are presented. In order to address the existing challenges, a hybrid framework is proposed, where deep learning methods are coupled with domain knowledge. An additional feature of this survey is that a review of the currently existing approaches integrating domain knowledge with deep learning techniques is presented. Finally, some conclusions on the implementation of hybrid architectures to perform computer vision tasks are discussed.