Study Of Facial Feature Point Detection For Multi-pose Face And Its Application In Eye Location

Face image contains various information which could be processed into more advanced information of human being,and these information can be widely used to understand human emotion,behavior and human-computer interaction.The location of facial feature point is the premise to get this information.However,the task of facial feature point detection is challenging due to the shape,size and other characteristics of the face region variability which be caused by pose variation.Therefore,in this paper,we study the problem of facial feature point detection under the multi pose change.A facial feature point detection algorithm for pose variation is proposed and applied at eye location in this paper.The main contents are as follows:(1)Facial feature point initialization for pose variation.In order to solve the problem that the initial position of the facial feature point is sensitive to pose variation,a facial feature point initialization algorithm based on facial orientation classification is proposed in this paper by analyzing the distribution of the feature points of faces in multi-pose.Firstly,for each face image in which the facial feature point will be positioned,the HOG(Histogram of Oriented Gradient)features of face image are extracted,and then train the random forest decision trees to predict its classify label.As a result,the different orientation selects the mean of the sample feature points in the corresponding training subset as its initialization value.(2)Extraction of shape-indexed features insensitive to pose change.In order to improve the robustness of facial feature point detection under multi-attitude variation,traditional framework of face feature point detection based on cascaded regression model is extend into three directions which include forward,left and right.Meanwhile,a new shape-indexed features which is insensitive to pose change is proposed based on the corresponding orientation.The features is called Average Pixel of Local Area feature Within-Class.Firstly,the feature points of key parts in the face are clustered,and the feature points in the class is randomly selected to compose the triangle template within the class composition triangle template set.Secondly,some reference points is selected in each region of the template,and the average value is calculated as the output in local region.Finally,the local region feature pair is obtained from feature correlation analysis as a shape-indexed feature.In this paper,the facial feature point initialization methods based on face orientation classification can effectively achieve classification in the different orientations.The average classification accuracy is 95.8%in CAS-PEAL Face Database and 93.8%in hybrid database of Helen-LFPW-300W.The average positioning error of the feature points is lower than the traditional random initialization method.At the same time,compared with the traditional features and statistical learning method,Shape-indexed features proposed in this paper has lower average positioning error of the feature points in hybrid database of Helen-LFPW-300W.When we apply our facial feature point detection algorithm to eye location,it has lower average positioning error comparing to statistical learning method.The experimental results show that the proposed eye location algorithm based on the feature point detection can solve the problem of pose variation.