Online Quality measurement of face localization obtained by neural networks trained with Zernike moments feature vectors

Abstract

Quality measurement of face localization using neural networks is presented in this communication. First, neural network was trained with Zernike moments feature parameters vectors. Coordinate vectors of pixels surrounding faces in images were used as target vectors on the supervised training procedure. Thus, trained neural network provides on its output layer a coordinate's vector (p, Theta) representing pixels surrounding the face contained in treated image. In second stage, another neural network, trained using TSL color space of images, is used to give a measure quantifying the quality of the localization obtained in the first stage. Experiments of the proposed method were carried out on the XM2VTS database.