8th of July was a great day for SkyBiometry not only because it was sunny summer day, but also because we have released a new version of algorithm.
Based on deep neural networks, the new algorithm provides 5 times higher accuracy in identifying full frontal faces and 10 to 15 times higher accuracy for unconstrained facial recognition.
“With this new version our development team focused on face recognition in real-world, unconstrained environments”, said Dr. Justas Kranauskas, project lead for SkyBiometry. “We achieved a ten-fold accuracy improvement on faces captured in lower resolution, with complex illumination, expressions and head rotations. This enabled us to offer a new face verification component which greatly simplifies user authentication by face, especially in mobile applications, while also enabling the face recognition algorithm to be used for complex 1:N identification”, Dr. Kranauskas added.
Improvements to the face recognition algorithm have resulted in much higher accuracy in facial identification compared to the previous version, based on False Rejection Rate (FRR) at the same False Acceptance Rate (FAR) value. This not only improves the user experience by resulting in fewer errors, it makes the product significantly easier to use and apply to a much broader range of face recognition applications, such as conducting automated facial image searches in large databases without the need for manual review.
Faster face detection and more accurate estimation of facial attributes, including gender, smile, closed eyes, open mouth, glasses and dark glasses are also included. What is more, number and position of facial points are changed (watch the picture below). The new facial landmarks detection and tracking capabilities are more robust in a wider range of facial poses.