More work needed on super-resolution to adapt selfie biometrics to poor conditions
Convenience and immediacy are great reasons for financial companies to offer selfie-acquired face biometrics, but the security they deliver depends on clear images, which the front-facing cameras on mobile devices can only take in good conditions.
As it happens there are plenty of researchers and companies looking into a computer vision technique called super-resolution for applications including biometrics.
Seeing the trend, three university researchers, from Sweden, Malta and Spain, evaluated super-resolution models for their ability to reconstruct face and iris images.
They found that traditional super-resolution techniques do not necessarily correlate with better match rates, because the specific structure of biometric images is different than the generic images super-resolution systems are often trained on. The changes super-resolution systems make, however, are not optimal for biometric processing.
With this in mind, the researchers set out to adapt super-resolution techniques to the specific needs of iris and face biometrics applications.
The team reports “promising performance,” but only under optimal, controlled conditions. They cited non-frontal views, facial expressions and lighting changed as factors in the wild that trip algorithms.
It is difficult to train algorithms because, as the authors point out, there is less than an abundance of poor-quality facial image databases.
Developers have tried unsuccessfully to get around that short coming by down-sampling normal images. But down-sampling does not create artifacts or pose variations.
They propose finding a way for image reconstruction schemes to properly align images despite the blurring that can occur with lower-resolution images.
Otherwise, a solution might have to wait until device and software makers are convinced a high-resolution front-facing camera will increase sales.