TLDR: A study successfully implemented YOLOv10-nano, a lightweight deep learning model, for real-time fish detection in Indonesian marine ecosystems, particularly Bunaken National Marine Park. The model achieved high accuracy (mAP50 of 0.966) and efficiency (2.7M parameters, 29.29 FPS on CPU) by training on DeepFish and OpenImages V7-Fish datasets. This makes it suitable for practical marine monitoring and conservation on resource-constrained devices.
Indonesia, home to the incredibly diverse Coral Triangle, boasts some of the world’s richest marine ecosystems. Protecting these vital environments requires effective monitoring of fish populations, a task traditionally hampered by time-consuming methods and the need for expert knowledge. This challenge has spurred the development of automated solutions, and a recent study highlights a significant advancement in this area: the application of a lightweight deep learning model called YOLOv10-nano for real-time fish detection in Indonesian waters.
The research, conducted by Jonathan Wuntu, Dwisnanto Putro, and Rendy Syahputra from the Dept. of Electrical Engineering at Sam Ratulangi University, focuses on leveraging the cutting-edge YOLOv10-nano architecture. YOLO, which stands for “You Only Look Once,” is a renowned framework in artificial intelligence for its ability to detect objects in images and videos in a single, rapid processing step. The YOLOv10 version, released in 2024, brings several enhancements, including an improved CSPNet backbone for better feature extraction, a Path Aggregation Network (PAN) for combining features from different scales, and a Pyramid Spatial Attention (PSA) Block to help detect objects of various shapes and sizes. These improvements make YOLOv10 more computationally efficient and accurate than its predecessors.
The study specifically aimed to implement YOLOv10-nano for marine fish detection and counting in Indonesian waters, using test data from the Bunaken National Marine Park. To train the model, two primary datasets were utilized: DeepFish and OpenImages V7-Fish. The DeepFish dataset provided a diverse sample of fish from various mangrove habitats, which is crucial for understanding the region’s biodiversity. However, it was found to be less effective for detecting fish near coral reefs. To overcome this limitation and enhance the model’s robustness, the OpenImages V7-Fish dataset, which includes more challenging images, was incorporated into the training process.
The results of this research are highly promising. YOLOv10-nano demonstrated high detection accuracy, achieving a mean Average Precision (mAP50) of 0.966 and mAP50:95 of 0.606. What makes these figures particularly impressive is that they were achieved while maintaining low computational demands, with only 2.7 million parameters and 8.4 GFLOPs. This efficiency is critical for real-world deployment, especially in environments where high-performance computing resources might be limited. Furthermore, the model delivered an average inference speed of 29.29 frames per second (FPS) on a standard CPU, confirming its suitability for real-time applications on less powerful devices.
While the OpenImages V7-Fish dataset alone yielded lower accuracy, its inclusion was vital in complementing DeepFish, significantly improving the model’s ability to detect fish in complex natural marine environments, such as those found near coral reefs. This combined approach addressed the limitations of using a single dataset and made the model more adaptable to diverse underwater scenes.
Also Read:
- Enhancing YOLO’s Reliability for Underwater Object Detection
- Deep Learning Illuminates the Microscopic World: A Look at Enhanced Imaging
In conclusion, this study successfully demonstrates the significant potential of YOLOv10-nano for efficient and scalable marine fish monitoring and conservation efforts in Indonesia. Its ability to perform accurately and rapidly on CPU-based systems opens new avenues for practical applications, including underwater monitoring, biodiversity assessments, and supporting marine conservation initiatives in data-limited regions. For more details, you can read the full research paper here.
