Sign In

Communications of the ACM

ACM News

Distinguishing Identical Twins

View as: Print Mobile App Share:
Identical twin sisters Sun Yutong and Sun Yumeng graduating from Harvard Graduate School of Education last May.

The challenges of using facial recognition systems to identify identical twins are amplified by their similarity, although like parents, facial recognition systems are beginning to be able to tell them apart in certain circumstances.

Credit: South China Morning Post

Facial recognition is evolving, challenging human capability to identify individuals. Can it accurately identify identical twins?

The answer is yes, and no.  

The challenges of using facial recognition systems to identify identical twins start with the same problems as identifying individuals, such as image quality, position of the head, facial expression, lighting, and superficial changes such as make-up and the addition or removal of beards and other facial hair. Those challenges are amplified by the apparent identicalness of twins, although like parents, facial recognition systems are beginning to be able to tell them apart in certain circumstances.

Jeremy Dawson, an associate professor specializing in biometrics at West Virginia University, has built datasets of identical twins' images for facial recognition research purposes. He describes algorithms that break down the principle components of a facial image, such as mouth, nose, and eyes, to create a template of the image and encrypt the data. Dawson says, "If images are high-quality and components of a face are broken down into a template, it is possible to see minute differences between identical twins."

There is a caveat, however; while it is possible to match an image of an identical twin against a small database of images, as the database grows, performance deteriorates as there is a greater chance that there are other facial images with similar features. Says Dawson, "If an image group is large, there is an impact on the performance of any identification."

Another approach to facial recognition uses spatial orientation of facial features. For example, a picture of eyes and everything around them can be transformed into spatial frequency (the level of detail present per degree of visual angle). This is the basis of linear discriminant analysis, based on an idea suggested by Sir Ronald A. Fisher in 1936 and used to find the subspace representation of facial images. Again, using this technique it may be possible to distinguish identical twins.

Arun Ross, professor of computer science and engineering at Michigan State University, breaks down facial recognition into three levels. Level one includes the shape of a face; level two includes specific features such as eyes, nose, and mouth, and level three incorporates more precise detail, such as freckles, scars, or tattoos. Says Ross, "Using multiple feature sets and level three features, identical twins can be identified, but systems will still make errors."

Having twice visited the Twins Day Festival in Twinsburg, OH, the largest annual gathering of twins in the world, Kevin Bowyer, Schubmehl-Prein professor in the department of Computer Science and Engineering at the University of Notre Dame, Indiana, collected facial, fingerprint, and iris data with the goal of distinguishing identical twins.

Like his peers, Bowyer says high-quality images and high-performing algorithms can go some way to distinguishing identical twins, although the algorithms need to be able to register fine details that are reliably different. The challenge here is transient detail, which Bower says can include such minute detail as scabs on the face, which can be covered up. In these types of cases, Bowyer says it is hard to catch twins substituting for each other. "Facial recognition most probably won't work if identical twins set out to defeat the system."

Facial recognition systems often use several techniques and fuse the results. For example, machine learning and neural networks allow systems to look at full images and subsets of data included in templates. Similarly, several algorithms can be agglomerated to create datasets of certain aspects of a face and, over time, learn what features to extract to support accurate facial recognition.

While vendor facial recognition systems are in place to match individuals to database images for purposes such as border control, law enforcement, and workplace security, they have limitations when the individuals are identical twins. Elke Oberg, marketing manager at Dresden, Germany-headquartered Cognitec, which markets FaceVACS facial recognition technology, says the algorithms underlying the FaceVACS product line can tell identical twins apart to the point even if they have very small differences in features that wouldn't reqister quickly on the human eye; when twins are absolutely identical, however, the technology fails.

Government agencies asked about the ability of their facial recognition programs to distinguish identical twins for this article declined to respond, perhaps because this is a problem they have yet to solve.

Looking at current schemes, U.S. Customs and Border Protection (CBP) is demonstrating facial recognition exit technology at eight U.S. airports, but have made no mention of how the system will deal with identical twins.

The U.S. Federal Bureau of Investigation (FBI) Next-Generation Identification program is expanding from a previous fingerprint system to incorporate other biometrics such as palm prints, iris scans, and facial recognition. Facial recognition may be used to support identification of potential offenders, but does not appear to be used in isolation for high security purposes.

Similarly, and perhaps tellingly, the Australian Department of Immigration and Border Protection's plans for a contactless passenger identification system to go beyond the SmartGate electronic border processing system currently in place and capable of matching the face of a passenger with the image stored in the microchip of their e-passport, will rely on both facial recognition technology and fingerprint scans.

In the consumer market, Apple's September 2017 introduction of the iPhone X with a facial recognition system called Face ID to unlock phones looked promising, but the fine print says children under 13 should not use the technology, as they could be too similar to each other to protect phones from intruders. From that, it is possible to extrapolate that the technology also may not work well with identical twins.

The overwhelming conclusion, for the time being at least, is that facial recognition systems working in isolation will not distinguish identical twins infallibly. More reliable solutions are iris scans and multiple fingerprints, none of which are identical, or perhaps one day in the future, facial images based on DNA.

Sarah Underwood is a technology writer based in Teddington, U.K.


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account