Detecção automática não invasiva de dados biométricos para peixes através de inteligência artificial

Abstract

The advancement of computing, including machine learning, has made it possible to optimize and minimize the consequences of the fish biometrics process while maintaining process efficiency and improving handling, reducing stress caused, and mitigating other symptoms that negatively impact the school and management. The aim of this work was to innovate the practice and methodology of the fish biometrics process by integrating knowledge from Artificial Intelligence, image detection, and object measurement. The project’s identification and prediction prototype was configured with the neural networks that make up YOLOv8, including the nano versions used in the initial tests and the large version, which is considered one of the most comprehensive YOLOv8 versions. The latter was used in the definitive training and adopted in the final prototype version. Additionally, a mathematical algorithm was implemented to estimate the size of the detected tilapia based on pixel size, converting measurements to provide the fish’s length in centimeters and an estimate of weight in real scales. After training, the model’s determined accuracy for masks and detection boxes was 74.5% and 81% in complex detections, and 97.4% and 97.8% in simple detections, respectively.