Evaluation of Eating Posture Detection Artificial Intelligence Models Trained on Different Supervised Datasets and Prototyping of Real-Time Feed Intake Estimation Tool for Broilers.

This study investigated the impact of supervised dataset content on the accuracy of object detection Artificial Intelligence (AI) model for detecting eating postures (EP) in broilers. The findings confirmed that (1) more individuals in EP did not necessarily improve the AI model's detection accuracy, and using still images with different numbers of such individuals could enhance the AI model's ability to detect EP. (2) Including both EP and postures similar to eating postures (SEP) was found to improve detection accuracy; additionally, adding still images that contain only SEP also enhanced detection accuracy; moreover, accuracy was higher when EP and SEP were included in the same still image than in separate images. (3) Ensuring an equal proportion of still images containing different numbers of individuals in EP and in SEP was necessary. Furthermore, a tool to estimate broiler feed intake in real time was prototyped. Using this tool, eight feed intake estimations of 0.5 h were conducted for five broilers under group management, resulting in the mean absolute error (MAE) of 4.3 g and the mean absolute percentage error (MAPE) of 13.3%, indicating good estimation accuracy.