AI Case Study

NASK researchers train convolutional neural networks to confirm horse identity based on eye images

Researchers from the Biometrics Laboratory Research and Academic Computer Network (NASK) develop a method to identify horses using eye scans, achieving a good accuracy rate. Deep convolutional neural networks are used to analyse both iris and periocular features as a method to quickly and accurately confirm racehorse identity.


Basic Materials


Project Overview

"Iris recognition has been shown to work with horse irides, provided that algorithms deployed for such task are fine-tuned for horse irides and input data is of very high quality. In our work, we ex- amine a possibility of utilizing deep convolutional neural networks for a fusion of both iris and periocular region features. With such methodology, ocular biometrics in horses could perform well without employing complicated algorithms that require a lot of fine-tuning and prior knowledge of the input image, while at the same time being rotation, translation, and to some extent also image quality invariant.

Periocular biometrics, on the other hand, employs features of the entire eye region, such as the location and shape of eye corners, eyelids, eyelashes, eyebrows, fragments of the nose, etc. Compared with iris recognition, periocular recognition often requires less constrained image acquisition conditions, such as imaging in visible light, at- a-distance, or on-the-move. However, these methods are far less capable of delivering excellent recognition accuracy. Our contribution is thus to make use of both iris and periocular features for horse identification. Employing convolutional neural networks for this task seems like a natural approach, as they are capable of taking a whole, unprocessed image as an input and and predict a class label by hierarchical feature extraction and classification."

Reported Results

The researchers achieved an equal error rate (EER) - the point at which the rate of false identification acceptances and rejections is the same - of 9.5%.


"Two variants of a deep convolutional neural network has been constructed for the task: HorseNet-4 and HorseNet-6, with four and six convolutional layers, respectively. The network receives an unmodified (apart from 4× downsampling) input eye image, which is then processed through several convolutional layers (Conv2D) with a 3 × 3 and 5 × 5 kernels, each of them followed by a max pooling layers (MaxPooling2D) with a pool size of 2 × 2 with stride of 1. Next, a fully-connected layer (Dense1) is put after the last Conv2D+MaxPooling2D+BatchNormalization set of layers. Finally, a second Dense2 layer with Softmax activation function outputs the estimated probability that the input sample belongs to one of N classes, where N denotes the number of animals. Batch normalization is applied after each pooling layer. Dropout was applied after the first Dense1 layer, with a probability of dropping any given unit set to p = 0.5. Dropout is a technique used to fight against network overfitting due to neuron co-adaptation (units reliance on other units) and employs random removal of a selected portion of neurons and their connections during training to prevent over-adaptation of the network. The model was trained using the Stochastic Gradient Descent (SGD) optimizer with learning rate of 0.01, decay of 1e−6 and momentum of 0.9."


R And D

Core Research And Development


"Identification of race horses is crucial for animal identity confirmation prior to racing. As this is usually done shortly before a race, fast and reliable methods that are friendly and inflict no harm upon animals are important."



"During the course of this study, we have collected a novel database of images representing eyes of horses. Twenty eight animals had their eyes photographed: 14 mares (50%), 10 stallions (35.7%) and 4 geldings (14.3%) aged 1 to 24 (with an average of 9 years old). The largest group were Arabian horses (15 animals, 53.6%).
For iris image collection horses had their eyes photographed using a specialized veterinary device Pupil- LR (by SIEM Bio-Medicale) originally purposed for oph- thalmological examinations, but at the same time capable of producing high-quality near-infrared photographs of the eye and its surroundings.