Advanced modelling and analysis for quality assessment and enhancement of underwater multimodal imageries.

Tolie, Hamidreza Farhadi

Abstract

Underwater sensing plays a crucial role in environmental protection and sustainable energy transitions, supporting marine ecosystem monitoring, resource management, and infrastructure development. However, visual perception in underwater environments is significantly restricted by light absorption and scattering, leading to colour distortion, contrast loss, and reduced visibility. Low-light conditions and turbidity further degrade image quality, limiting the acquisition of high-quality data. To address these challenges, multimodal sensing, which integrates optical and acoustic imaging, has been widely adopted. Despite its advantages, modality-specific degradations persist, resulting in low-quality data. This thesis introduces novel image quality assessment and enhancement techniques tailored for multimodal underwater imagery, improving perception in optical, Sound Navigation and Ranging (SONAR), and stereo vision sensors to support decision-making for both remotely operated and autonomous systems. A novel method is presented for evaluating underwater optical images by analysing edge structures, perceptual features, and colour dispersion. A directional Kirsch kernel-based approach captures scattering-induced degradation, while contour and saliency maps enhance object boundary emphasis. Channel-wise dispersion rates quantify colour distortion, and saturation and hue metrics measure colour purity and distinguishability, improving correlation with human subjective scores by 10% over state-of-the-art techniques. Additionally, a SONAR image quality assessment method is introduced that employs wavelet domain analysis to separate low-frequency noise from high-frequency object details for quantifying perceptual and utility quality, respectively. By extracting micro and macro-scale texture and contour features from decomposed components, this approach measures object visibility and identifiability, achieving an 11% higher prediction accuracy than existing methods. To mitigate proportional degradation and non-uniform colour casts in optical images, a deep neural network is designed that integrates inception modules and channel-wise attention mechanisms. Multi-scale feature extraction and channel-specific attention allow for improved assessment of colour loss and distribution, enhancing colour restoration by 2% in terms of structural similarity quality compared to prior methods. An integrated framework is further developed for automatic underwater object identification and distance measurement by refining stereo vision-based depth images and extracting object information from SONAR data. The fusion of stereo vision and SONAR data enables precise depth and range estimation, facilitating robotic manipulation and achieving an error rate of less than 1 centimetre. Extensive validation on benchmark datasets and in-lab experiments demonstrates significant improvements over existing techniques. These contributions enhance under-water inspection, condition monitoring, and maintenance by improving real-time positioning and sensing reliability. By advancing multimodal sensing technologies, this thesis strengthens underwater image analysis, enabling more effective robotic applications that support environmental sustainability and energy infrastructure management. To facilitate further research, source codes and collected data are publicly available at https://github.com/hfarhaditolie.

Citation

TOLIE, H.F. 2025. Advanced modelling and analysis for quality assessment and enhancement of underwater multimodal imageries. Robert Gordon University, PhD thesis. Hosted on OpenAIR [online]. Available from: https://doi.org/10.48526/rgu-wt-2988453

• Thesis Type: Thesis
• Deposit Date: Aug 25, 2025
• Publicly Available Date: Aug 25, 2025
• DOI: https://doi.org/10.48526/rgu-wt-2988453
• Keywords: Underwater image quality assessment; Image enhancement; Multimodal sensing; Sonar image analysis; Depth image refinement
• Public URL: https://rgu-repository.worktribe.com/output/2988453
• Award Date: Mar 31, 2025