Dehazing Techniques for Navigation in Underwater Environment

Abstract

This thesis focuses on developing modules that can be integrated to develop a semi-autonomous underwater vehicle aimed to assist human operators in performing task-oriented operations in a challenging underwater environment. Autonomous underwater vehicles use visual, inertial, haptic, and other sensory data to augment the perception of human operators and help with the generation of motion control signals during the vehicle’s navigation. As previous studies have shown, our system demonstrates the effectiveness of the above-mentioned control methods in harsh underwater conditions that exhibit low visibility and high turbidity. The effectiveness is particularly relevant in scenarios with a high quantity of suspended particles.

The solutions proposed in this thesis involve extensive use of machine learning techniques. The high computational and memory demands of the existing learning-based systems pose challenges for real-time information-rich visual and sensory data processing. This is particularly relevant in resource-constrained platforms like underwater robots, where efficiency and energy optimization are crucial for long-duration operations. Inspired by biological neural systems, neuromorphic computing offers an alternative to conventional methods. Spiking Neural Networks (SNNs), with their energy-efficient spike-based computations, present a promising approach, and their applications in underwater robotics are explored in this thesis.

This thesis presents three different SNN-based modules for a semi-autonomous underwater vehicle. Accordingly, three specific research objectives are addressed in this research, leading to three key contributions: (i) the development of an energy-efficient SNN-based underwater visibility enhancement algorithm, (ii) the proposal of a lightweight underwater dehazing model with a smaller number of parameters, combining physical image formation principles with a spiking transformer, and (iii) devise a hybrid CNN-SNN-based pose estimation module using multimodal visual-inertial data.

The primary contribution presented in this research is the design of UIE-SNN, a 19-layered spike-driven visibility enhancement algorithm to address the challenges of visual perception in underwater conditions. UIE-SNN achieves comparable performance to existing conventional methods while reducing energy consumption by 85%. Comparative evaluation shows that UIE-SNN is 3.5× to 9× more energy-efficient than existing UIE methods.

The second contribution is the introduction of snnTrans-DHZ, an underwater dehazing framework, integrating a learning-based spiking transformer model with a physics-driven image formation model to improve underwater image clarity while maintaining computational efficiency. The snnTrans-DHZ model, with only 0.57M parameters, could yield a PSNR score as 21.6773 dB and SSIM value of 0.8795 on the UIEB dataset.

The third contribution is the development of NeuroVIO, a multimodal hybrid VIO framework aimed at addressing the challenges in pose estimation of an underwater vehicle. NeuroVIO is a hybrid CNN-SNN framework for multimodal visual-inertial odometry estimation. It leverages the strengths of both architectures to achieve energy-efficient performance in challenging underwater conditions. NeuroVIO uses a CNN-based module to extract the required features from successive image frames. The extracted features are then converted to their equivalent spike representations. These spike signals are then concatenated with the inertial features extracted using the SNN-based feature extractor. The concatenated features are then fed to an SNN-based regression head to estimate the six-dimensional pose of the vehicle.

The solutions presented in this thesis are modular and integrated to form an energy-efficient, scalable framework for the semi-autonomous navigation of underwater vehicles. These techniques contribute to advancing underwater robotics by developing energy-efficient methodologies for environmental perception and localization in challenging environments.

Date of Award	20 May 2025
Original language	American English
Supervisor	JORGE Dias (Supervisor)