Research On Facial Feature Detection Technology Based On Hierarchical Ramdom Forests

The technique of facial feature detection is to detect the position of some previously-defined points in face images, mainly the position of five sense organs. Since facial feature detection can provide some important basic data for face analysis research, it plays an significant role in the field of face analysis. At present, national and international scholars have put forward a large amount of methods in aspects of facial feature point detection, among which the detection accuracy of frontal faces in highresolution images has exceeded that of the human annotated accuracy. However, there are still great difficulties in facial feature point detection to no constraint imaging conditions. The main reasons lie in that changing factors that influence face images lead to the accuracy decrease of facial feature point detection. These changing factors include complex background, change of face orientation and facial expression, occlusion of ornaments beside faces(hats, sunglasses, face covers, ect.), change of illumination condition and so on. Aimed at the difficulties in facial feature point detection to no constraint imaging conditions, the thesis makes improvement on classical random forest algorithm and proposes building hierarchical random forest model to perform facial feature point detection.The hierarchical random forest model has two layers, the first of which is the random forest used for estimating face orientation whereas the second of which is the conditional regression forest used for detecting facial feature point. During the process of training, firstly, in order to estimate face orientation, the first-layer random forest is created by learning textural characteristics of each image patch in the complete training set. Then the conditional regression forest is created by learning the spatial relationship between each image patch and facial feature point in the training subset classified according to face orientation. Since the regression forest is created in conditional training subset, the decision tree doesn't need to take care of all facial changes both in outside look and forms when growing. During the testing process, the first-layer random forest is used to estimate the face orientation of the tested images and then the secondlayer conditional regression forest is created according to the calculated face orientation rate. A voting collection is formed by all the leaf nodes the tested image patches can reach in the second-layer conditional regression forest. So as to remove the effect caused by local occlusion of the face or global occlusion of facial feature point, the thesis puts forward improved hough voting method based on the geometric constraint conditions about the position of facial central point and each feature point. At last, the final position of the facial feature point is obtained by combining all voting elements with the mean-shift algorithm. The research result indicates that the method put forward in this thesis preforms well in the human face dataset. The detection accuracy is close to the human annotated accuracy under the unconstraint conditions.