Face biometrics accuracy over time boosted by adaptive reference gallery: EAB webinar
Adaptive face biometric systems have offered a way to improve the performance over time of systems that might otherwise suffer from degraded accuracy as time passes from the collection of the reference template. The latest lunch talk from the European Association for Biometrics (EAB) addresses the question of whether such systems are still necessary in the age of deep learning.
‘Adaptive Biometric Systems in the Deep Learning Era’ was presented by Giulia Orrù of the University of Cagliari’s Pattern Recognition and Applications (PRA) Lab.
Orrù starts by reviewing the characteristics of trustworthy biometrics systems; performance, fairness, security and explainability, and the impact on face biometrics systems of inter-class similarities and variations and variations over time.
Dealing with these challenges requires a good-quality reference template of the type produced in multiple enrollment processes, which are time-consuming, expensive, and require more cooperation than usual enrollment methods.
“Adaptive strategies are based on the detection of novel templates to replace or update” the template already contained in the gallery, Orrù explains.
A basic feedback loop is created to detect and ingest the novel biometric reference material, but this raises challenges around computational complexity, high processing time, high memory requirements and imposter insertion.
These challenges are addressed with a selection algorithm, which filters templates.
Adaptive systems can be either self-updating or co-updating, Orrù says. The former has evolved significantly over the past two decades, with the application of harmonic function and risk minimization techniques, and detection of changes in lighting conditions.
Currently, adaptation is carried out in supervised or semi-supervised modes, with references either based on a single template or built from several templates, or sets of references arranged for different biometric aspects in a multi-modal system.
Orrù presented a taxonomy of adaptive systems, with adaptation strategy in the center, and described online and offline adaptive methods.
Processing time can be kept constant by limiting the number of samples per subject in the gallery.
Orrù’s PhD research involved developing a novel classification-selection approach for adaptive facial recognition systems, which she briefly presented. These methods can be based on clustering templates to select those which return a match only for the correct subject, or editing the reference gallery based on the best matching score.
The technique, advantages and potential risks of several classification and selection methods were explained in some detail.
An analysis simulating long-term biometric system use required building a new dataset, as legacy datasets do not include temporal information that could be used to sequence their ingestion. The ‘Aphotoeveryday’ dataset is made up of 98 subjects, with photos captured over time, a few of them including images taken over a decade apart.
Different well-known face biometric models were used, and the acquisition period of the templates classified as short, medium or long. False matches and non-matches were measured, along with equal error rate and imposter percentage (erroneous sample insertion).
In a medium acquisition time scenario involving a young person, where the match scores for genuine and imposter probes begin to draw together after around 4 years, the application of an update system resulted in a notable improvement in biometric matching performance.
In a long timeframe example, biometric matching effectiveness fell apart within just a few years without an adaptive system, but the autoencoded template system again restored the performance of the biometric model.
In the end, a semi-supervised method was found to be mildly effective at preserving biometric performance, but each of the other adaptive methods held equal error rates low. This improvement was also reflected in the number of imposters included in the reference galleries.
Therefore, even in the age of face biometrics models based on convolutional neural networks and other deep learning techniques, adaptive systems for reference galleries can play a major role in improving performance over longer periods of time.