R. Montes Diez, C. Conde, Á. Serrano, L.J. Rodríguez-Aragón, and E. Cabello (Spain)
Computer vision, pattern recognition, face verification.
We present a face verification system. A 100 people face database has been created using a CCD video camera, with controlled illumination conditions and frontal upright face position. A Principal Component Analysis matrix has been computed with eight images per person, and only the 150 most important eigenvalues have been used. The results of PCA are fed into two classifiers (SVM and RBF), in order to perform a verification process in a control access. The algorithm proposed here allows to compute automatically the optimal acceptance threshold to divide a population of candidates into genuine or registered and impostors or non-registered. A Bayesian approach based on screening techniques has been considered, so that the user provides the economical cost for false acceptances and false rejections within the system. According to the ratio between these two costs, the optimal acceptance threshold is computed as the value that minimizes the expected total cost for both acceptances and rejections. Our experimental results show that our SVM classifier produces a lower false acceptance rate (FAR) for a given false rejection rate (FRR), and vice versa, than our RBF classifier. The FAR also appears to cancel for SVM for high security conditions.
Important Links:
Go Back
