Livestock Fish Larvae Counting using DETR and YOLO based Deep Networks

Counting fish larvae is an important yet demanding task in aquaculture. In this work, we evaluate four neural networks, including convolutional and transformer-based architectures, in different sizes, in this counting task. We present a new annotated image dataset with less data collection requirements than preceding works, with images of spotted sorubim and dourado larvae. We also combined the neural networks with a tiling technique, in order to improve the counting performance, and performed a comprehensive experimental evaluation on the methods. We achieve a MAPE of 4.46% (± 4.70) with an extra large real-time detection transformer, and 4.71% (± 4.98) with a medium-sized YOLOv8. The results indicate the feasibility of using deep learning to count spotted sorubim and dourado larvae, which may become an important tool to enhance the efficiency of aquaculture farms that deal with these species.