Machine learning methods for automatic cheiloscopy on facial images

Human identification through lip recognition (i.e., cheiloscopy) has recently been studied by several authors. Most researchers use lip shape features, which leads to discriminative information loss by not taking into account local features. Others use local feature descriptors - like SURF and SIFT - to match images of lips, losing information of the general structure. The main contribution of this paper is the adaptation of classic facial recognition techniques to lip photography recognition. These methods use a global features approach: they consider images as a high-dimensional vector and then apply a dimensionality reduction algorithm, trying to preserve as much discriminative information as possible. In this paper, several dimensionality reduction techniques are combined with different machine learning classifiers and their recognition rate is determined empirically. Additionally, a novel data augmentation technique is proposed to correctly train the different classification models, even when only one picture per subject is available. Light normalization and lip alignment problems are also addressed, so the cheiloscopical analysis chain is described in its entirety.