Automated Facial Action Unit Recognition in Horses

Abstract: In recent years, with the development of deep learning and the applications of deep learning models, computer vision tasks such as human facial action unit recognition have made significant progress. Inspired by these works, we have investigated the possibility of training a model to recognize horse facial action units automatically. With the help of the Equine Facial Action Coding System (EquiFACS) created by veterinarians recently, our aim has been to detect EquiFACS units from images and videos. In this project, we proposed a cascade framework for horse facial action unit recognition from images. We firstly trained several object detectors to detect the predefined regions of interest. Then we applied binary classifiers for each action unit in related regions. We experimented with different types of neural network classifiers and found AlexNet to work the best in our framework. Additionally, we also transferred a model for human facial action unit recognition to horses and explored strategies to learn the correlations among different action units.