Classifying Faces with Non-negative Matrix Factorization
David Guillamet, Jordi M. Vitria · 2002
This paper addresses the well-known problem of recognizing faces under several unfavorable situations. We have analyzed situations with changes in expression, in illumination and occlusions such as faces wearing sunglasses or scarfs. We have introduced the use of the Non-negative Matrix Factorization (NMF) technique in the context of classification of face images and we have directly compared performances of NMF and Principal Component Analysis (PCA) using a well-known face database, the AR, that contains a large number of individuals taken under several conditions. Moreover, these results have also been compared to two leading algorithms, one template based and the other feature based, noticing that NMF is able to improve them when using a high dimensional space. In addition, NMF has been used with some distance metrics as L1, L2 or correlation in order to determine the best one for such problem. We have discovered that the correlation metric is the most suitable one for our problem.