3D Facial Modeling Using Image Processing

Abstract

This research work proposes new artificial neural network algorithms for face image segmentation and 3D facial modeling. Ruth s face database is used as a reference for modeling 3D face. Two orientation of a face is taken at 90o. These two images are segmented. The segmented information is projected onto Ruth s face to form a 3D facial model. Contextual clustering algorithm is used for extracting features from the facial images. The features are Mean, Summation and Contextual value (Vcc). The Vcc is compared with a threshold to assign segmentation values for the image. Training patterns are formed from the extracted features to train the back propagation (BPA), Radial basis function (RBF) and Echo state neural network (ESNN). At the end of training final weights are stored into a database. The final weights are used to segment facial images. The receiver operating characteristics curve, accuracy and sensitivity of the implemented algorithms are discussed. newline1. The average peak signal to noise ratio (PSNR) of the segmentation algorithms are CC=27.36 db, BPA is 24.32db, RBF is 38.37db, and ESNN is 42.3db. newline2. The average peak signal to noise ratio (PSNR) of the facial modelling by BPA is 23.27db; RBF is 30.76db, and ESNN is 33.71db. The PSNR values of the proposed ANN algorithms for modelling is less when compared to the PSNR of the segmentation outputs of the algorithms. The reason for the low values are due to mismatching of the chin, cheek and forehead profiles as the size of the Ruth s face is smaller than the size of the faces of persons used in this research work. newline newline