Comprehensive Study on Underwater Object Detection Based on Deep Neural Networks: Application of YOLOv8 and YOLOv9 in Marine Debris Recognition

This article conducts an in-depth study on underwater object detection, using YOLOv8 and YOLOv9 models with 32 experimental configurations on the TrashCan 1.0 dataset. It evaluates the impact of model variants, category distributions, and learning rates on detection performance, providing technical references for marine environmental protection and underwater robot vision.

Research Background and Significance

Oceans cover more than 70% of the Earth's surface, but marine debris pollution has become a global issue—over 8 million tons of plastic waste enter the oceans each year, threatening ecosystems. Traditional underwater monitoring relies on manual diving or simple camera equipment, which is inefficient and costly. Underwater environments face challenges like light attenuation and water turbidity, making conventional computer vision algorithms less effective. Deep learning technologies, especially YOLO series models, offer new possibilities to solve these problems.

Experimental Design and Methodology

The study conducts 32 groups of experiments, examining three variables: model variants, dataset structure, and learning rates:

• Model Selection: Multiple variants of YOLOv8 (n/s/m/l) and YOLOv9 (t/s/m/c);
• Dataset Configuration: The TrashCan1.0 dataset is divided into three categories (debris, animals, ROV) and four categories (adding plants), including training/validation/test images;
• Training Parameters: All models are trained for 100 epochs, with learning rates set to two levels: 0.01 and 0.0001.

Key Findings and Result Analysis

1. Positive correlation between model capacity and performance: YOLOv9c performs best on the three-category dataset, while YOLOv8l leads on the four-category dataset;
2. Impact of category count on detection difficulty: Models perform better on the three-category dataset than the four-category one—simplifying categories can improve accuracy;
3. Evaluation metrics: High-capacity models have advantages in precision, recall, and mAP metrics, with YOLOv8l and YOLOv9c showing outstanding performance.

Practical Application Value

• Marine Environmental Protection: Automated detection systems can be deployed on AUVs/ROVs to achieve large-scale, long-term monitoring with low cost and wide coverage;
• Underwater Robot Vision: Provides a technical foundation for autonomous navigation, target tracking, etc.;
• Expanded Fields: Can be applied to underwater archaeology, marine biology research, underwater facility inspection, etc.

Research Limitations and Future Directions

Limitations: The dataset is not included in the code repository, requiring users to obtain it independently, which increases the threshold for reproduction; model weights may need to be provided separately or trained.

Future Directions: Explore lightweight models to adapt to embedded devices; study multi-modal fusion (sonar + optical images); develop online learning mechanisms to adapt to different underwater environments.

Conclusion

This study comprehensively evaluates the performance of YOLOv8 and YOLOv9 in underwater object detection through systematic experiments. It finds that model capacity is positively correlated with detection performance, and controlling category complexity has important implications for practical applications. The research results provide references for the development of underwater computer vision systems and contribute to marine environmental protection.