Abstract
In aquaculture, accurately counting fish fries is a prerequisite for fish population management and marketing. However, due to the fry overlapping and occlusion issues, the manual counting method is time-consuming, and the counting result might be inaccurate. Therefore, we utilized the computer vision technique to develop a lightweight fish fry counting model to address this issue. First, we constructed a diverse dataset containing images of largemouth bass in varying numbers, captured under different lighting conditions and water depths. Then, we optimized the YOLOv5s model by channel pruning to reduce the model size, while maintaining the detection accuracy. Through extensive experiments, we examined the effect of different pruning rates on the model performance and compared the pruned YOLOv5s model with state-of-the-art detection models through the evaluation criteria like precision, recall, mean average precision (mAP), model size, GFLOPs, and detection speed. In addition, we investigated the impact of environmental factors, such as lighting conditions and water depths, on the detection performance of the pruned YOLOv5s model. The experimental results demonstrated that the YOLOv5s model with a pruning rate of 15% achieved over 90% accuracy and 13 FPS in the dense and complex scenes, which met the practical requirement for the fry counting task. In addition, we also identified that the pruned YOLOv5s model would achieve the optimal performance under the white illumination and shallow water depth setting. In conclusion, this study provided an efficient solution for fish fry counting, and the proposed model was expected to be applied in the real-world application.
