Computational cost of anchor-based algorithms.

It is necessary to overcome the real-life challenges encountered in detecting marine animals in underwater bodies through computer vision for monitoring their populations and biological data. At the same time, the detectors for such tasks are limited by large parameters, high computation costs, low accuracy, low speed, and unfriendly deployment in low-power computing devices due to their large size. To tackle these problems, MADNet was developed using the YOLO framework, incorporating both anchor-based and anchor-free techniques. The structure of MADNet includes CBS, C3b, Bottleneck, SPPFr, and C3 modules, and it was evaluated against YOLOv5n, YOLOv6n, YOLOv7-tiny, and YOLOv8n with consistent application methods on various open-source underwater image datasets. Using the computation cost, trained time, loss, accuracy, speed, and mean absolute error (MAE) as performance evaluation metrics, the anchor-free methods performed better than the anchor-based methods. Similarly, the overall performance score for MADNet was analyzed at 27.8%, which is higher than 20% for YOLOv8n, 18.9% for YOLOv6n, 17.8% for YOLOv5n, and 15.6% for YOLOv7-tiny. As a result, MADNet is lightweight and effective for detecting marine animals in challenging underwater scenarios.