Researchers led by Tali Boneh-Shitrit et al. (2022) have introduced the first controlled-experiment dataset of canine emotional facial expressions, featuring 29 Labrador Retrievers exposed to two distinct emotional contexts: positive anticipation and negative frustration. These emotional states were elicited in a structured lab environment, with facial movements recorded and analyzed using both manual coding and artificial intelligence techniques.
Two AI approaches were tested to classify the dogs’ emotions. The first used a Dog Facial Action Coding System (DogFACS)-based method involving a decision tree classifier built upon identified facial movements. The second employed a deep learning model trained directly on video data without manual coding. While the DogFACS approach achieved over 71% accuracy, the deep learning model reached a significantly higher accuracy of 89%.
A notable contribution of this study lies in its focus on explainable AI. The decision trees mirrored known DogFACS emotional correlates, providing transparent, rule-based insights. Meanwhile, the deep learning model used heatmaps to visualize the network’s attention, sometimes highlighting facial regions consistent with DogFACS variables—such as ear and muzzle tension. These attention maps may also uncover hidden visual cues that are imperceptible to the human eye but critical for emotion recognition.
This research advances our understanding of canine emotional communication and lays the groundwork for future AI-based welfare monitoring tools. It also underscores the importance of combining data-driven techniques with interpretable models in studying animal emotions.
Source: Tali Boneh-Shitrit, Matan Feighelstein, Andrea Bremhorst, Shiran Amir, Tzion Distelfeld, Yael Dassa, Shira Yaroshetsky, Stefanie Riemer, Ilan Shimshoni, Daniel Mills, Ariel Zamansky. “Explainable automated recognition of emotional states from canine facial expressions: the case of positive anticipation and frustration.” Scientific Reports, Volume 12, 2022-12-01.
