Challenges in Underwater Object Detection and Video Segmentation Using Deep Learning

Abstract

The diverse marine life, water distortion and lighting conditions make the process of object detection and instance segmentation for underwater video data highly challenging. In light of this dilemma, this paper investigates the development of machine learning (ML) models that are specifically used to carry out object detection and instance segmentation for detecting fishes and sharks in underwater videos. The research is crucial for applications in marine biology, environmental conservation, and underwater surveillance, as accurate detection and segmentation are essential for understanding underwater ecosystems, monitoring endangered species, and assessing human impacts on marine environments. Even though there have been many advancements in the field of computer vision, to date there is lack of research regarding the development and evaluation of specialized models for underwater object detection and instance segmentation. Therefore, this research aims at filling the gap by proposing methodologies and evaluating the performance of two computation models: for Object Detection and Instance Segmentation. The research methods employed include frame extraction, manual annotation, data preprocessing, data augmentation, model building, validation and testing. The (You Only Look Once) YOLOv8 architecture is utilized for both the object detection and instance segmentation Models. The findings indicate that while both models show promising performance, the instance segmentation model achieves superior accuracy in detecting and outlining individual fish and sharks. These results provide valuable insights and solutions for researchers and practitioners in marine biology, environmental monitoring, and underwater robotics, supporting progress in underwater exploration and conservation efforts.