An SVM-PCA-based method of human face ROI localization

Human face Region of Interest (ROI) localization is a prerequisite step for face recognition and lip reading application. This paper describes a ROI localization method based on Support Vector Machine (SVM) and Principle Component Analysis (PCA). We use simplified skin color model to segment input images. In the estimated face area, we first make rough classification with two-class SVM and then we localize the facial ROI by minimizing PCA reconstruction error, also called Distance From Feature Space (DFFS). Results are presented on images from Chinese Audio-Visual Speech Database (CAVSD).