DOLPHIN DETECTION USING AN ENHANCED LIGHTWEIGHT YOLO ARCHITECTURE

Abstract

Dolphin detection plays an important role in marine ecosystem monitoring, species conservation, and behavioral analysis. However, visual identification in underwater environments faces challenges such as light refraction, water turbidity, and dynamic sea conditions. This study proposes a deep learning-based dolphin detection approach by modifying the YOLOv8 architecture to produce a lightweight yet accurate model. The modifications include reducing the number of channels in the backbone and neck, as well as simplifying the SPPF block, thereby reducing the model parameters from 3.01 million to 1.83 million and the computational complexity from 8.2 GFLOPs to 7.2 GFLOPs. A specialized dolphin dataset consisting of 5,493 labeled images, collected from underwater and surface conditions, was developed to train and evaluate the model. Experimental results show that the proposed model achieves 67.1% mAP@50 and 45.8% mAP@50–95, outperforming YOLOv8-Nano and other lightweight YOLO variants. Additionally, the model demonstrates better runtime efficiency, with a latency of 49.2 ms and 20.38 FPS, making it suitable for real-time implementation on resource-constrained devices. Overall, this research presents a more efficient and accurate dolphin detection solution, while also providing a specialized dataset that can support further research in the field of computer vision-based marine conservation.