Neural Network Cascade for Facial Feature Localization

摘要

We present here a complete system for the localization of facial features in frontal face images. In the first step, face detection is performed using Viola &: Jones state of art algorithm. Then, a cascade of neural networks localizes precisely 28 facial features. The first network performs a coarse detection of three areas in the image corresponding roughly to left and right eyes and mouths. Then, three local networks localize, in these areas, 9 key points per eye and 10 key points on the mouth. Thorough experiments on 3500 images from standard databases (Feret, BioID) show the detector accuracy, its generalization ability and speed.