In animal breeding, the interest is in individual animal performances, which often involves the continuous monitoring and assessment of individual animals in terms of health, welfare, and production. Wearable sensors can be used to monitor and assess but are not easily applicable or preferred in poultry, given the production scale and animal size. Computer vision could provide a more versatile and easier scalable alternative. Computer vision entails image acquisition and subsequent transformation of images (high dimensional spatial data) by a computer to a more human interpretable format. Convolutional neural networks (CNNs) are state-of-the-art methods to transform images to useable data but require large manually annotated datasets of high quality to train the algorithm. Obtaining large high-quality datasets is, however, time-consuming and often constrained by budget. In this project, the aim is to develop and study methods that improve individual phenotyping of poultry using computer vision when there is limited quantity and quality of data for training. Through case studies in broilers and turkeys, four specific goals will be investigated: 1) the use of data augmentation to increase the quantity of data for detection of broilers by a CNN, 2) the use of a sensor-fusion approach between radiofrequency identification and video for individual recognition and tracking of broilers 3) the use of confident learning for the prediction of subjective turkey gait scores with a long short-term memory network (LSTM) to enhance data of limited quality, and 4) transfer learning across bird species (turkey to broiler) to increase the quantity of annotations in broilers. The outcomes of these case studies, and therefore of this PhD-project, will help in developing new camera-based sensor technologies that can assist in the continuous monitoring of group-housed animals which is useful for the improvement of breeding, and inadvertently for the improvement of management.
|Effective start/end date
|1/06/20 → …
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.