Using deep learning models to decode emotional states in horses.

Area of Science:

• Animal behavior
• Machine learning
• Computer vision

Background:

• Accurate assessment of animal welfare is crucial.
• Understanding equine emotional states is vital for ethical riding practices.
• Developing objective methods for emotion detection in horses is an ongoing challenge.

Purpose of the Study:

• To develop and compare machine learning models for predicting emotional states in ridden horses.
• To investigate the impact of different image cropping strategies on model performance.
• To enhance model accuracy through techniques like transfer learning and fine-tuning.

Main Methods:

• Manual labeling of images for supervised learning.
• Data augmentation using Yolo and Faster R-CNN for cropped body and head datasets.
• Training and evaluating various convolutional neural network (CNN) models on different datasets.
• Application of transfer learning, fine-tuning, and interpretation methods (LIME).

Main Results:

• The cropped head dataset achieved the highest performance, with 87% accuracy, 79% precision, and 97% recall.
• CNN models trained on cropped datasets outperformed those trained on uncropped images.
• LIME interpretation method identified features consistent with expert annotations.

Conclusions:

• Cropped head images are highly effective for machine learning-based emotion prediction in ridden horses.
• CNNs, particularly when fine-tuned, offer a robust solution for automated emotion analysis in equines.
• Model interpretability methods like LIME can validate AI findings against expert knowledge.