Broiler Live Weight Estimation through Image Processing and YOLO-based Deep Learning

This study aims to estimate the live weight of broilers using image processing and deep learning techniques. The proposed method is designed to optimize management processes in broiler production systems, reduce labor requirements and operational costs, minimize animal stress caused by direct human contact, and serve as an effective alternative to traditional manual weighing practices. The study was conducted at a commercial broiler farm located in Kahramanmaraş, in the Mediterranean region of Türkiye. An automatically controlled measurement enclosure was constructed to capture broiler images with minimal human intervention. Digital cameras mounted at the top of the enclosure recorded images of broilers that spontaneously entered the enclosure. By using these images, live weight estimation was carried out in two stages: the first stage involved morphological image processing in the MATLAB environment, while the second stage focused on deep learning-based modeling for prediction. During the image processing stage, multiple regression analysis was performed using the actual weights obtained through manual weighing and the estimated weights derived from image-based measurements. The analysis resulted in an adjusted R² value of 0.97 and a standard error of ± 131 g (P<0.01). The mean absolute error (MAE) was calculated as 84.4 g, while the mean relative error (MRE) was found to be 7.6%. In the deep learning stage, the YOLOv8 model was trained for 150 and 500 epochs. Notable improvements in both accuracy and generalization capability were observed after 500 epochs. Under these conditions, the model achieved a high mean Average Precision (mAP) of 0.969, with substantial increases in precision, recall, and F1-score across all 17 predefined broiler live weight classes. Furthermore, regression-based performance indicators were approximated from the class-based predictions to enable a quantitative assessment of weight estimation accuracy. Based on this indirect evaluation, the proposed model achieved an estimated MAE of 37.0 g and MRE of 4.43%. Overall, the findings suggest that the proposed framework has strong potential for adaptation to live weight prediction in other livestock species.