Learn to Train: Improving Training Data for a Neural Network to Detect Pecking Injuries in Turkeys

SIMPLE SUMMARY: Injurious pecking against conspecifics in turkey husbandry is a widespread, serious problem for animal welfare. Evidence suggests that bloody injuries act as a trigger mechanism to induce pecking. Thus, continuous monitoring of the herd should be ensured to allow timely intervention in this type of behavior. The aim of the present study was therefore to develop a camera-based warning system using a neural network to detect injuries in the flock. The data for the network were provided by images on which human observers marked existing pecking injuries. Then, a network was trained with these human-labeled images in order to learn to detect pecking injuries on other unknown images from the same domain. As the initial agreement on the injuries detected by human observers and the trained network was unacceptable, various work steps were initiated to improve the data that were used to train the network. Finally, the aim of this process was for the network to achieve at least a similar ability to mark injuries in the images as a trained human observer. ABSTRACT: This study aimed to develop a camera-based system using artificial intelligence for automated detection of pecking injuries in turkeys. Videos were recorded and split into individual images for further processing. Using specifically developed software, the injuries visible on these images were marked by humans, and a neural network was trained with these annotations. Due to unacceptable agreement between the annotations of humans and the network, several work steps were initiated to improve the training data. First, a costly work step was used to create high-quality annotations (HQA) for which multiple observers evaluated already annotated injuries. Therefore, each labeled detection had to be validated by three observers before it was saved as “finished”, and for each image, all detections had to be verified three times. Then, a network was trained with these HQA to assist observers in annotating more data. Finally, the benefit of the work step generating HQA was tested, and it was shown that the value of the agreement between the annotations of humans and the network could be doubled. Although the system is not yet capable of ensuring adequate detection of pecking injuries, the study demonstrated the importance of such validation steps in order to obtain good training data.