Over the past decade, the use of Remotely Operated Vehicles (ROVs) in marine
research has grown significantly due to advances in computational power and
robotics. These tools now allow marine biologists to gather high-quality
underwater video footage for analyzing marine life. However, challenges such as
low visibility, light scattering, and colour distortion hinder accurate object
detection and classification in these environments. As a result, researchers
have turned to computer vision methods—particularly deep learning models like
YOLO—for real-time detection of underwater objects such as fish and corals.
While YOLO-based models have shown strong performance in underwater fish
detection, current research remains limited in scope. Most existing datasets
and models, including FishNet, FishInTurbidWater, and FishDETECT, are
fish-centric and do not account for the broader ecological diversity,
particularly marine vegetation. There is a noticeable lack of well-defined
datasets and ontologies for identifying and classifying underwater plants,
which are essential for comprehensive marine ecosystem monitoring. Efforts like
CoralNet and CATNet's MSID dataset have broadened species categories, yet
marine vegetation remains underrepresented.
To bridge this gap, we introduce FjordVision, a
hierarchical deep learning framework designed for detecting and classifying
both marine vegetation and fauna in Esefjorden, Norway. FjordVision includes
the Esefjorden Marine Vegetation Segmentation Dataset (EMVSD),
featuring over 17,000 annotated images with more than 30,000 labelled marine
objects. Leveraging YOLOv8 for instance segmentation and enhanced with a
taxonomically structured classification model, FjordVision improves on
traditional flat classification by categorizing objects into binary, class,
genus, and species levels. This approach delivers more ecologically relevant
insights, making FjordVision a vital tool for biodiversity monitoring and
marine conservation.
No comments:
Post a Comment