Assuming the input photo is focused on a person's face, if the person is wearing a surgical mask, most face recognition software fail to identify the subject's face.

Most facial landmark models are trained to identify at least the eyes and the tip of the nose (for example, dlib's 5 point landmark).

Is it possible to construct a model that is trained to identify a face based on only the eyes?

Edit: Sorry for my broken english, but by "eyes" I mean the periocular area. I am terribly sorry because english isn't my first language.

  • $\begingroup$ Try transfer learning from a network already proficient in face recognition on the entire face, and see how it performs. My guess would be yes, it is possible (but wont be as accurate). $\endgroup$
    – Recessive
    Feb 5 at 3:52
  • $\begingroup$ It could be possible. It might be inaccurate. But in some countries it might be illegal... See in Europe the GDPR $\endgroup$ Feb 5 at 8:00
  • $\begingroup$ @exius To clarify the terminology, facial recognition is accomplished using a photo or video of the face. When you say 'just with the eyes' you are no longer talking about facial recognition. Facial recognition is a biometric using an image of the face. Iris and retina recognition are biometrics using an image of the eye. A biometric is used to identify a person. $\endgroup$ Feb 5 at 16:34

Yes, it must be possible as retina scanners have been used as a method of personal identification for some time. The difference is, if you have a retinal scanner you are probably controlling the focal distance of the picture you are taking and using suitably high resolution. Your mileage may vary as these things decrease in quality.


The two main eye biometrics are iris recognition and retina recognition (aka retinal scan). These are not going to work from an ordinary photo of someone's face. I have used iris recognition at about ten feet away and this article claims it can be done at 40 feet!

Eye recognition, or identification of a person from an image of their eyes alone (i.e., without seeing their iris or retina), has such a high error rate that it is not done. You may find the following paper of interest:

Nawaz Ripon, K. S., Ershad Ali, L., Siddique, N., & Ma, J. (2019). Convolutional Neural Network based Eye Recognition from Distantly Acquired Face Images for Human Identification. 2019 International Joint Conference on Neural Networks (IJCNN), Neural Networks (IJCNN), 2019 International Joint Conference On, 1–8. doi:10.1109/IJCNN.2019.8852190

For more information on iris recognition see pp 10-11 of this tutorial, and pp 12-13 for retinal scan.

An excerpt from Retinal vs. Iris Recognition: Did You Know Your Eyes Can Get You Identified? by Danny Thakkar:

Retina recognition The posterior portion of human eye forms retina. It is made of a light sensitive tissue. When light passing through cornea and lens reaches retina, neural signals are generated and transferred to the brain via the optic nerve. Retina is a thin layer of tissue formed by neural cells. Capillaries responsible for blood supply of this layer forms a pattern that can be used for personal identification. This pattern of blood capillaries is believed to be unique in each individual due to huge possibility of variation how these capillaries run on the surface of retina. Since retina is located at the posterior portion inside the human eye, special equipment is required to scan this pattern. Retina recognition is one of the least deployed biometric methods because of high cost of the implementation and its highly invasive nature that may cause some user discomfort. Still, it is used is very high security applications like military and high level government access due to its accuracy and high level of security.

Retina recognition systems make use of low energy infra-red light to scan the retinal pattern. Blood vessels absorb infrared light while surrounding tissues reflect it. This reflection is detected by the retina recognition system and image of this pattern is captured. This image is further enhanced to make is usable for the recognition algorithm. Retina template is generated once the image is taken through recognition algorithm; this template is associate with a subject’s demographic data and stored. The process so far is called enrolment. The subject’s identity can be verified anytime by scanning a new retinal sample and matching it against the stored template.

Iris recognition Iris is the ring shaped colored portion in a human eye and is visible from outside with naked eye. It is made of muscle tissue that adjusts the size of pupil and controls how much light can enter the eye. Amount of melatonin pigment in iris is responsible for different colors that human eyes take. Folds in iris muscles throughout the ring create a pattern with great amount of details. Formation of this pattern is completely random and there is no rule how it will turn out in an individual’s eye. However, once this pattern is created during the foetal development, it stays the same throughout the life. An individual’s irises are unique and structurally distinct, even iris of same individual does not match. All these attributes make them good enough for personal recognition.

Details of iris can be captured with any high quality digital camera, however, modern recognition systems make use of near infrared (NIR: 700–900 nm) instead of visible light to capture details. Since iris recognition can be established with high quality camera and recognition software, it can be setup on any computing device; however, dedicated recognition systems are more common due to performance and security reasons. Iris recognition systems use a camera to capture details of the iris and this image is enhanced by the image enhancement algorithms. Once the image is usable enough, it is processed by the recognition algorithms, which extracts unique features to generate a biometric template. Associating identity data with this template establishes identity of the subject in question, which can be used for identity verification in future.

  • $\begingroup$ Thankyou for your answer. However I have some questions about your claim that eye identification is not done due to it having high error rate. In this paper and this paper having both around 90% accuracy at best. Why did you claim such a way or am I horribly missing something ? thankyou very much. $\endgroup$
    – exius
    Feb 6 at 10:29
  • $\begingroup$ My response is about real world applications and not research. 90% at best is very poor when compared to over 99% for facial recognition. For the very unusual, and temporary, situation we are in now (i.e., the COVID-19 pandemic) it sounds like eye recognition would be a solution but for what? Would you risk a 10% or higher error to use it at an ATM machine? What application could you possibly deploy this in before the pandemic is over? I can certainly see eye recognition being used to supplement a human in identifying people in special situations. One example would be for detective work. $\endgroup$ Feb 6 at 14:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.