INTELIGÊNCIA ARTIFICIAL NA CONTAGEM DE ALEVINOS DE TILÁPIA Oreochromis niloticus

Abstract

Tilapia counting is of great importance in aquaculture systems. It allows us to evaluate and monitor the population, control reproduction, estimate production and control the environment so that there is no overpopulation, avoiding possible diseases and saturation of the environment. One of the difficulties in training neural networks is labeling images, which is often done manually. The aim of this study was to evaluate the efficiency of automatic labeling for identifying and counting tilapia fry using artificial intelligence. 1000 images of tilapia fry (10 ± 1.2 g) were collected in a controlled PVC environment with a white background. The images were collected as videos and then transformed into photo images for labeling. The roboflow platform was used for automatic labeling, and models with 100 and 150 epochs were tested. The metrics mAp 0.5, recall and precision were evaluated. It was found that mAp@0.5 was 98.2% for 100 epochs, recall 97% and precision 93.5%. For 150 seasons, the results were mAp@0.5% 98.1%, recall 95.4% and precision 93%. It can be concluded that automatic labeling can be used to train neural networks for counting tilapia fry, and that 100 seasons was more efficient than 150 seasons.