Suitable Approach for Body Segmentation of Broilers Based on Deep Learning

Abstract Smart technologies in modern poultry farms enable precise monitoring of broiler performance through computer vision and artificial intelligence, supporting data-driven farm management and improved production outcomes. This study evaluates three deep learning models Mask R-CNN with MobileNetv2, YOLOv8-large, and SAM for broiler body segmentation. Accurate broiler body segmentation plays a crucial role in modern farm management tasks, including weight estimation, health monitoring, and performance tracking, ultimately contributing to more efficient and sustainable poultry production. A dataset comprising 1122 top-view images of Arian broilers was collected over 13 different days during the growth period after building a suitable data acquisition platform. Models-specific modifications and customizations were implemented to enhance training and evaluation. YOLOv8-large achieved the highest segmentation accuracy (99.5%) and efficient training within 50 epochs, while also delivering real-time processing speeds of 33 frames per second, suitable for embedded applications. Mask R-CNN exhibited rapid convergence within 100 epochs; however, its performance was constrained by the lightweight MobileNetv2 backbone. SAM demonstrated high accuracy and smooth segmentation outcomes by applying a region of interest (RoI) approach, although its high computational requirements and slower processing speeds restricted its practicality for real-time deployment. Overall, YOLOv8-large combined high accuracy, fast inference, and low resource demands, positioning it as the most suitable model for real-time broiler segmentation. These results underscore the potential of deep learning-based solutions to enhance scalability, efficiency, and precision in poultry farming.