Vision-Language Models for Underwater Plastic Detection with Parameter-Efficient Fine-Tuning

Voladri, Prashanth Reddy (2025) Vision-Language Models for Underwater Plastic Detection with Parameter-Efficient Fine-Tuning. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (802kB) | Preview
Preview	PDF (Configuration Manual) Download (835kB) | Preview

Abstract

Marine plastic pollution is threatening the globe's oceans, but plastic detection underwater is challenging due to lighting variation, turbidity, and color distortions. The present research compares the vision-language plastic detection models underwater, such as the application of the Low-Rank Adaptation (LoRA) and the Weight-Decomposed Adaptation (DoRA) fine-tuning techniques applied to PaliGemma and PaliGemma2 architectures. Using 10,945 underwater images from JAMSTEC with comprehensive augmentation strategies, four model configurations were trained and evaluated on standard object detection metrics. Results reveal striking performance disparities: LoRA- adapted PaliGemma achieved 80.89%mean Average Precision at IoU 0.5 with balanced prediction density (1.03 predictions/image), while all other configurations failed catastrophically. DoRA adaptation proved unsuitable for both architectures, achieving less than 9% mAP, contradicting theoretical expectations. PaliGemma2 exhibited severe over-detection (8.9 predictions/image) regardless of adaptation method, suggesting architectural unsuitability for precise spatial tasks. While traditional detectors like YOLOv8 achieve marginally higher accuracy (91.2%), the PaliGemma+LoRA configuration offers unique advantages in semantic reasoning potential. The findings indicate that architectural simplicity and adaptation stability outweigh theoretical sophistication in the case of specialized detection problems, so vision-language models become viable alternatives to marine environment monitoring if properly configured.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Milosavljevic, Vladimir UNSPECIFIED
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	03 Jul 2026 11:38
Last Modified:	03 Jul 2026 11:38
URI:	https://norma.ncirl.ie/id/eprint/9468

Actions (login required)

View Item