MÉTODOS DE ROTULAÇÃO DE IMAGEM PARA CONTAGEM DE PEIXES

Abstract

Fry counting in aquaculture is a crucial task for production management, and automation via Artificial Intelligence emerges as a promising alternative to traditional methods. This study aimed to evaluate the impact of two different image labeling methods—polygons and bounding boxes—on the performance of the YOLOv8 detection algorithm for counting Nile tilapia (Oreochromis niloticus) fry. For this purpose, approximately 5,000 images were captured in a controlled environment, with densities ranging from 10 to 50 fish per container. The results showed a clear superiority of the polygon labeling method, which maintained precision and recall above 97% across all evaluated densities, with a mean Average Precision (mAP) of up to 99%. In contrast, the bounding box method showed a sharp decline in performance at high densities, with the recall dropping to 80.78% in the 40-fish scenario. It is concluded that for the automatic counting of fry under overlapping conditions, polygon annotation is essential to ensure the robustness and high precision of the model, being the most suitable method for applications that require reliable results.