Researchers have developed an artificial intelligence system capable of detecting and classifying different types of eye blinks in horses, a development which could eventually support research into equine pain, stress and emotional state.
The study, led by researchers from Aalborg University and the Swedish University of Agricultural Sciences, compared three computer-vision methods for identifying full blinks, half-blinks and periods when the eye remained open.
Subtle changes in horses’ facial expressions can provide useful information about their experiences. Blink-related facial movements have previously been associated with pain and stress, but they can be difficult and time-consuming for people to identify, particularly when reviewing video recorded at normal speed.
Automating part of this process could help researchers analyse larger amounts of footage and reduce the considerable time required to annotate equine facial movements manually.
Testing three approaches
The researchers evaluated three different methods:
- a YOLOv12 model which analysed individual video frames
- an optical-flow method which detected movement around the eye
- a VideoMAE model designed to analyse short sequences of video
The models were trained and tested using footage involving six horses of different breeds. A publicly available set of 12 videos, containing two videos from each horse, was used for testing.
The YOLO model was trained using 13,206 manually labelled frames. Each frame was classified as showing an open eye, half-blink or full blink.
When the task was simplified to identifying whether a blink had occurred, without distinguishing between full and half-blinks, the YOLO and VideoMAE models both achieved a macro-F1 score of 0.926 when evaluated using video clips.
Macro-F1 is a measure that considers how well a system performs across each category, rather than allowing the most common category to dominate the overall result.
For the more difficult three-category task, the video-based YOLO approach performed best, achieving a macro-F1 score of 0.898.
Half-blinks were difficult for people too
The researchers also investigated how easily people could identify the same movements.
Six participants without specialist annotation experience were asked to classify 22 short video clips as showing a full blink, half-blink or no blink. Overall classification accuracy was 78.8%, with agreement highest for full blinks and open eyes.
Half-blinks proved considerably more difficult. Agreement for this category was 73%, with some half-blinks incorrectly identified as full blinks.
The researchers said this reflected the subtle and sometimes ambiguous nature of partial eyelid closure.
When tested on the same 22 clips, the YOLO model classified every clip correctly. However, the researchers noted that these clips had been deliberately selected to avoid ambiguity, meaning the result should not be interpreted as evidence that the model will perform perfectly across all horses or real-world conditions.
A possible tool for researchers
The findings suggest automated blink detection could become a useful tool for identifying sections of video that require closer examination.
For example, a system could flag possible blink events for a trained observer to review rather than replacing professional assessment altogether. This could make the process of creating annotated datasets faster and help support future research into equine facial expressions.
The technology is not yet ready to diagnose pain or determine a horse’s emotional state independently. Blinking can occur for many reasons, and welfare assessment requires consideration of the whole horse, its behaviour, health and surrounding circumstances.
The study also had several important limitations. It used footage from only six horses, while half-blinks remained challenging to classify. The different methods also faced practical difficulties, including reduced sensitivity to subtle movement, limited temporal detail and sensitivity to changes in recording conditions.
The authors said larger and more varied datasets will be needed, alongside work examining a wider range of pain-related facial movements.
While still at an early research stage, the study demonstrates how computer vision may eventually assist scientists and veterinary professionals in recognising subtle changes that can easily be missed by the human eye.
