An Ultra-Lightweight Fish Detection Model for Real-Time Aquatic Animal Monitoring on Embedded Platforms

Abstract

Simple Summary Monitoring fish without handling them can support biodiversity conservation, animal welfare, ecosystem assessment, and sustainable aquaculture. Camera-based monitoring reduces disturbance, but many detection models require computers that are too large or power-hungry for widespread use beside ponds, tanks, cages, or other aquatic observation sites. We developed an ultra-lightweight fish-detection model designed for small, low-power embedded devices. The model occupies only 3.0 megabytes and processes 19–24 frames per second on an embedded platform while retaining useful detection accuracy. It was evaluated using an in-house aquarium dataset containing challenging conditions such as variable illumination, turbidity, occlusion, and small or distant fish, and was also trained and evaluated on a difficult subset of a public fish-image dataset. The results show that real-time fish detection can be achieved with substantially lower computational demand. This approach may provide a practical visual-monitoring foundation for future applications in fish-population assessment, feeding management, animal-welfare monitoring, and aquatic conservation.