Determining whether an unknown face belongs to a gallery of faces is known as face recognition. To solve this problem, both a face model and a measure of similarity between faces are needed. To model faces, we can mainly distinguish two approaches: those based on a facial image, and those based on geometric features of the face.
One of the best solutions for face recognition is based on a principal component analysis (PCA) of the facial image. However, as the database of faces increases in size, the computational efficiency of these algorithms decreases. Motivated by this, and by the fact that geometric methods give a lower dimensional vector representation of a face, the goal of the work presented in this paper is to use a geometric method to filter the face database, to reduce its size before applying another method for exact recognition.
The authors present two different models of faces: one based on specific points or landmarks of a face, and another based on ratios of distances between the landmarks. In both models, PCA is applied to reduce the dimension of the geometric representation of a face. The authors also propose two similarity measures corresponding to each model, with which they obtain the best results: refined Procrustes distance for the landmark-based model, and eigenvalue-weighted cosine (EWC) for the ratio-based model. The methods are tested on two well-known databases, and the results are presented, in an extensive and detailed way, to justify the conclusion that both methods are appropriate for use with the filtering application, and in exact face recognition. The methods also show robustness with regard to changes of expression and age in faces.