Nondestructive estimation method of live chicken leg weight based on deep learning

Abstract

In the broiler-breeding industry, phenotype determination is critical. Leg weight is a fundamental indicator for breeding, and noninvasive testing technology can reduce damage to animals. This study proposes a broiler leg weight estimation system comprising a weight-estimation model and computed tomography (CT) acquisition equipment. The weight-estimation model can automatically process the scan results of live broiler chickens from the CT acquisition equipment. The weight-estimation model comprises an improved you-only-look-once (YOLOv5) segmentation algorithm and a random forest fitting network. The segmentation head was introduced into the YOLOv5 network, combined with a multiscale attention mechanism and an atrous spatial pyramid pooling architecture, and a new network model, YOLO- measuring chicken leg weight (YOLO-MCLW), was proposed to improve segmentation efficiency and accuracy. Morphological parameters were extracted from the obtained mask image, and a random forest network was used for fitting. The experiments show that the system exhibited an average absolute error of 7.27 g and an average percentage error of 4.82% in tests on 50 individual legs of 25 broiler chickens. The prediction R2 of broiler chicken legs can reaches 88.98%, the segmentation intersection over union result reaches 95.45%, and 37.04 images are processed per second. This system provides technical support for the part determination of broiler chickens in commercial breeding.