Assessing face image quality with LSTMs
Biometric authentication using ngerprints or face recognition is making its way into
the mainstream, and there is an urgent need to make these authentication methods as
secure and reliable as possible. One way to achieve better performance with a biometric
authentication method, is to introduce a quality estimation step early in the pipeline, so
that unsuitable, or low-quality samples can be rejected.
While existing work predominantly focuses on algorithms for detecting specic properties
of the face images, we investigate whether machine learning techniques can provide a
general way to estimate overall face image quality.
We train a selection of neural network types, and discover that a type of Recurrent
Neural Network (RNN) called Long Short-Term Memory (LSTM) can reliably estimate
face image quality, with better performance than the bespoke algorithms.