Irish and US researchers improve biometric age estimation with facial embeddings
A team of scientists from University College Dublin, Intel Ireland, and The University of Texas at San Antonio has improved biometric underage age estimation performance using an innovative method based on a regression-based model.
Dubbed Vec2UAge, the model was trained from the VisAGe and Selfie-FV biometric datasets using FaceNet embeddings extracted and used as feature vectors.
A third, unbiased dataset was then created to allow for balanced testing and validation.
Generally speaking, age estimation models rely on age-labeled, good-quality images, but accurate age annotations in facial biometric datasets are often inadequate, according to the new research.
“Certain age groups have few samples – particularly the underage age range,” the report reads. “Datasets for this age range are difficult to find due to legal restrictions and ethical implications.”
Yoti is working on building its own database for age estimation biometrics in significant part to address this same problem.
To circumvent this issue, the researchers chose the IMDB-WIKI and Adience datasets, respectively holding 500k and 26k age-categorized images.
During the experiment, the team also utilized various augmentation techniques to further expand the training dataset, then measured the deep neural networks’ learning rate (LR).
For context, LR here refers to how much the system should change the model in response to the estimated error each time data sets are updated.
A value too small may cause the training process to become too slow and eventually come to a halt, while a value too large may lead the system to an unstable training process or cause it to learn a sub-optimal set of weights too fast.
To find the optimal initial value for LR, the researchers then used a cyclic learning rate approach. They then evaluated the ensuing performance of the algorithms and concluded the new model achieved better performance than most algorithms today. The model reached a mean absolute error rate of 2.36 years.
“Current models usually attempt to tackle several challenging factors that affect the age estimation performance such as facial occlusion, non-frontal faces, brightness, contrast, quality, etc,” the report explains. “In our approach, a simpler challenge is addressed and a better performance is achieved.”
Moving forward, the team said it will continue to explore models to improve biometric age estimation performance, particularly through the use of a hyperparameter optimization framework for machine learning like Optuna.