Geometrical facial feature selection for person identification

Automatic human identification is a hot topic with many applications, mainly in security. A common type of facial features used for the purposes of biometric recognition and identification includes geometrical characteristics. To this respect, the current study is performed on a proprietary dataset of 53 subjects (with several frames per subject) by utilizing 3D geometrical face features extracted from the Euclidean and geodesic intradistances between specific nodal facial points. The objective is to achieve high accuracy in the verification of the IDs of the users. Two feature selection methods based on information criteria are selected and benchmarked herein, i.e. the minimum redundancy and maximum relevance (mRMR) and the conditional mutual information with nearest neighbors estimate (CMINN). The repeated computations on several randomly selected training and test sets from the ensemble of frames give evidence for successful classification of the 53 persons based on a significantly reduced subset of features, where smaller cardinality of the subset is obtained by CMINN. The fact that both methods converge to the same level of classification accuracy when the feature subset increases, proves that high recognition rates can be achieved via only a very small fraction of features.