YOLO-Starfish: fish object detection learning complex underwater features

Area of Science:

• Computer Vision
• Robotics
• Marine Biology

Background:

• Underwater object detection faces significant challenges due to poor image quality, environmental variability, and fish concealment.
• Accurate fish detection is crucial for underwater robots to explore and understand aquatic ecosystems.

Purpose of the Study:

• To develop an advanced fish detection model and a comprehensive dataset for improving underwater object detection capabilities.
• To address the limitations of existing methods in handling the complexities of underwater imaging.

Main Methods:

• Proposes YOLO-Starfish, an enhanced YOLOv8 model incorporating the C2Star module for feature distribution and the Attention-driven Enhancement Module (ADEM) for channel imbalance.
• Introduces the Underwater Freshwater Fish Dataset (UFFD) with 16,904 images of 19 fish species in diverse underwater conditions.

Main Results:

• YOLO-Starfish demonstrates superior performance on underwater datasets (RUOD, UFFD) and the general object detection benchmark COCO2017.
• The C2Star module effectively mimics underwater optical degradation, while ADEM improves robustness against image channel imbalances.

Conclusions:

• YOLO-Starfish offers a robust solution for underwater fish detection, outperforming existing methods in challenging aquatic environments.
• The UFFD dataset provides a valuable resource for advancing research in underwater computer vision and marine robotics.