New NIST FRVT Quality Assessment report seeks to improve face biometrics performance metrics
A second draft of the FRVT (Face Recognition Vendor Test) Quality Assessment Report has been published by the National Institute of Standards and Technology to help biometrics developers process images captured with lower quality.
NIST observes that despite marked improvements in the accuracy of biometric facial recognition, error rates remain significant, especially in applications with subpar photography conditions or in which stringent thresholds are set to ward off false positives. The quality of reference samples retained in databases is critical to performance, which is why quality scalar and quality vector are used as quality assessment tools.
In the quality assessment track, NIST evaluates algorithms that report scalar quality values to help improve the automated detection of poor-quality images. This gives it a much-reduced group of algorithms, and the 32-page report analyzes entries from the Lomonosov Moscow State University, the Universidad Autonoma de Madrid, China Electronics Import-Export Corp, Guangzhou Pixel Solutions Co Ltd, Paravision (formerly EverAI) and Rank One Computing, which is the only developer with two entries in the evaluation.
The various ISO/IEC standards relating to image quality, including the ISO-IEC-39794-5:2019 standard which includes ICAO-Portrait specifications and ANSI-NIST Type 10, have helped system accuracy, but several related standards are still in development.
The new report introduces new performance metrics, and NIST says that the area remains under-developed, so encourages the submission of new algorithms and comments to help improve the formulation and analysis of the problem.
The report examines different use cases and the predictive power of quality values for false negative matches, and also considers whether quality algorithms should predict false positives.