Face segmentation using projection pursuit for texture classification

Abstract

Frontal face images are segmented into 7 regions using only sum and difference histograms as pixel information, without any a priori knowledge. In the training phase, a decision tree is created using a projection pursuit algorithm: in each step, the optimal one-dimensional projection is chosen by a simulated annealing process according to a projection index, and classes are isolated by a decision boundary that maximizes class separability, until the end nodes contain only one class each. Satisfactory qualitative and quantitative results were obtained and presented. © 2012 Springer-Verlag.