Synthetic aging helps facial recognition algorithms navigate age gaps, says new paper

The question of aging is a particularly tough nugget for legislators to chew on when considering facial recognition or face biometrics for identity verification. There are subtleties in the debate, particularly when it comes to age assurance, which has been proven accurate on young faces, despite statements to the counter by UK regulators.

However, says a team led by researchers from the University of Galway, “it is noteworthy to mention that personal biometric characteristics change as time goes by.” Given that, it is hard to argue with the statement that “a robust age-invariant face recognition system is of crucial importance to address the challenges posed by aging and maintain the reliability and accuracy of facial recognition technology.”

This notion provides the motivation for a new paper that explores “the feasibility of utilizing synthetic aging data to improve the robustness of face recognition models that can eventually help in recognizing people at broader age intervals.”

The team’s method involved designing a set of experiments to evaluate cutting-edge synthetic aging methods, then testing the effect of age intervals on current deep learning-based facial recognition algorithms by combining synthetic and real aging data to perform “rigorous training and validation.”

“Experimental results show that the recognition rate of the model trained on synthetic aging images is 3.33 percent higher than the results of the baseline model when tested on images with an age gap of 40 years, which prove the potential of synthetic age data which has been quantified to enhance the performance of age-invariant face recognition systems.”

In simpler terms, for facial recognition systems that are not designed to factor in aging, the synthetic aging data helps improve accuracy.

The paper does note that in many scenarios, submitting a new photo at regular intervals is effective, as when renewing a passport. But it highlights outlier use cases such as tracking an individual who has been missing for years. And while the authors concede that “despite the advancements in synthetic age algorithms such as SAM, CUSP, and AgeTransGAN, these methods still exhibit limitations in preserving facial identity when compared to real world images,” they ultimately claim a win for their synthetic aging technique.

“As the final research outcome we concluded that leveraging synthetic age images for training purposes demonstrates robust results for face recognizer algorithms against the age gap promise.”

Free synthetic face dataset contains 10,000 identities

Those who want instant access to synthetic face data for algorithm training can download the face recognition dataset SynMulti-PIE, which has been made available for free by the Biometrics Security and Privacy group at the Idiap Research Institute.

In a post on LinkedIn, Sébastien Marcel, a senior member of the Institute, says the dataset “contains a total of 10,000 synthetic identities, with 18 variations (pose, illumination and expression) per identity.” It is available for download here.

Article Topics

biometric dataset  |  biometrics research  |  face biometrics  |  facial recognition  |  identity verification  |  synthetic data  |  synthetic faces