MADNet: Marine Animal Detection Network using the YOLO platform.

Area of Science:

• Marine Biology
• Computer Vision
• Artificial Intelligence

Background:

• Underwater computer vision faces challenges in marine animal detection for population monitoring.
• Existing detectors suffer from large parameters, high computational costs, and low accuracy, hindering deployment on low-power devices.

Purpose of the Study:

• To develop an efficient and lightweight computer vision model for marine animal detection in underwater environments.
• To address limitations of current detectors in terms of accuracy, speed, and computational cost.

Main Methods:

• Developed MADNet using the YOLO framework, integrating both anchor-based and anchor-free techniques.
• Employed a network structure comprising CBS, C3b, Bottleneck, SPPFr, and C3 modules.
• Evaluated MADNet against YOLOv5n, YOLOv6n, YOLOv7-tiny, and YOLOv8n on diverse underwater image datasets.

Main Results:

• Anchor-free methods demonstrated superior performance compared to anchor-based methods.
• MADNet achieved an overall performance score of 27.8%, significantly higher than YOLOv8n (20%), YOLOv6n (18.9%), YOLOv5n (17.8%), and YOLOv7-tiny (15.6%).
• MADNet exhibited improved accuracy, speed, and reduced computation cost and training time.

Conclusions:

• MADNet is a lightweight and effective solution for detecting marine animals in challenging underwater conditions.
• The model's performance indicates its suitability for real-time population monitoring and biological data collection.