Data augmentation for facial expression based automatic pain assessment in equines

Abstract

Recognition of pain in equines is essential for their welfare. However, unlike humans, equines lack verbal communication and are reliant on external observers to assess their pain. Observers can assess equine pain using different pain scales, such as the Horse Grimace Scale, EquiFACS, or EQUUS-ARFAP. Training an observer is time-consuming, and observers often disagree on a diagnosis. This necessitates the need to automate the equine pain assessment process. In this work, we provide a system for pain assessment in equine faces based on the EQUUS-ARFAP scale. The proposed system consists of four steps, namely, automatic head orientation detection, automatic detection of facial regions, automatic pain detection for each facial region of interest separately, and automatic data generation. Our main contributions are a detailed analysis of the usage of Region of Interest (ROI) as the main representation of the assessment pipeline, instead of facial landmarks, and the deployment of synthetically generated data. We show improved pain classification on the publicly available Utrecht University Equine Pain Facial Dataset dataset and advance the state of the art in this problem. Part of the results produced in this thesis were published.