Image-Based Underwater Liquid Leak Detection and Transfer Learning

This paper addresses the critical challenge of liquid leaks in the oil and gas industry by leveraging advanced computer vision and deep learning methodologies. The objective is to develop practical models for detecting underwater objects with low image quality in adverse conditions. We train and test CNN detectors using Facebook's Detectron2 Faster R-CNN. The model was evaluated on a custom dataset of underwater oil spill videos, focusing on detection accuracy and processing speed. The results demonstrated that even using images of smoke in the sky as training made it possible to detect the underwater oil leak accurately.