In this study, I wanted to look at the accuracy of facial recognition software when used on images of a person at different ages of their lives. This presents value in simple tasks, like quickly sorting and identifying old pictures, to more serious cases, like identifying missing or kidnapped children.
I used the face_recognition library, by ageitgey (link and credit on main page). From his page, “Built using dlib's state-of-the-art face recognition built with deep learning. The model has an accuracy of 99.38% on the Labeled Faces in the Wild benchmark.” This library mapped and created a 128-dimension array of facial measurements of the eyes, nose, lips, and chin; points specifically chosen as “landmarks” which are visible on every face, not (as) easily obscured, and remain proportional for large portions of our lives.
Tests required a collection of “known” faces (the people to search for) and “unknown” faces (the pictures to be searched). In the known faces, I used samples of individuals when they were young and old.
My brother, Watson, showing sample pictures used of the same person at different ages
Then, I searched large collections of pictures with faces in them. Some pictures contained faces the software “knew”, some contained people it had never seen before. Some pictures were of single people, and others contained large groups.
After testing multiple people, over periods of time from several years to several decades, there were many visible trends.
First, recognition of infants is highly unreliable. With much pudgier, more rounded faces, the distinctive 'landmark' features are much harder to find and match to older pictures of the same person. The faces of infants were also often mis-matched to other infants pictures. In addition to inaccuracy, human faces change so rapidly that the infant template becomes next to worthless after the first year to year and a half of life.
Second, facial recognition seems to work best within subsets of years. In my findings, my data showed that a facial scan would only work for about 10-12 years before it started producing highly inconstant results. So a picture of a 2 year old will match to all of that persons' later pictures up to about age 12 or 14. After that, a subject has matured to the point where their 'landmark' features no longer match to those same features a decade prior. However, after someone reaches early adulthood (around ages 17-21 in my trials) their face can be matched until very late in life, when sags and wrinkles begin to distort the 'landmark' measurements once again.
A facial scan of a child around age 2 will matched with pictures taken once per year
Around age 15, however, the original scan rarely matched with later pictures
The original facial scan of the father, however, worked throughout the entire 22 year series of pictures
And last, match quality suffers greatly when used with large sample sizes. The larger the number of known faces to map, as well as the larger the body of sample pictures to search, makes the number of mismatches and false-positives rise exponentially. When looking for a specific child's picture among hundreds of unknown faces doest the work of clearing out the pictures that obviously lack the child, but still returns a large number of pictures that don't feature the child. Different camera angles, zoom levels, photo quality, and just plain looking like someone else is far too common to produce meaningful results.
Although it would be lovely if an e-z pass camera could pick up the face of a missing child as he whizzes by on the highway, the technology just isn't there yet. There remains too much variation from both the known face and the unknown sample to search to produce results with enough certainty to deem credible. Research is beginning to show promise in the field of 3D facial mapping and recognition, which accounts for many of the shortcomings of my study. With the countless applications of instant identification, I believe this area will be a major focus of study for decades to come.