Tracking pecking behaviors and damages of cage-free laying hens with machine vision technologies

Feather pecking (FP) is one of the primary welfare issues in commercial cage-free hen houses as that can seriously reduce the well-being of birds and cause economic losses for egg producers. After beak trimming is highly criticized in Europe and the USA, alternative methods are needed for pecking monitoring and management. A possibility for minimizing the problem is early detection of FP behaviors and damages to prevent it from spreading or increasing as feather pecking is a learned behavior. The objectives of this study were to develop a machine vision method, testing the performance of new models in tracking the pecking behaviors of hens and potential damages in the cage-free facilities and improve the detection accuracy of the model. Two YOLOv5 based deep learning models, i.e., YOLOv5s-pecking and YOLOv5x-pecking, were developed and compared in tracking FP behaviors of laying hens cage-free facilities. According the performance based on a dataset of 1924 images (1300 for training, 324 for validation, and 300 for testing), YOLOv5x-pecking model had a 3.1 %, 5.6 %, and 5.2 % higher performance in precision, recall, and Map than YOLOv5s-pecking model, respectively. However, YOLOv5s-pecking model size is 80 % smaller, and thus used 75 % less GPU memory and 80 % less time in model training than YOLOv5x-pecking model for the same dataset. Therefore, YOLOv5s-pecking model was considered with superior performance. This study was among the first to apply YOLOv5 models to track problematic behaviors of cage-free hens. The model provides a basis for developing a real-time automatic model for tracking pecking damages in commercial cage-free houses to protect the health and welfare of hundreds of millions laying hens.