Research On The Detection Of Facial Feature Points Based On HOG Feature And Gradient Boosting Decision Tree

Face recognition is an important research topic in the field of computer vision,and its purpose is the identification and verification by the feature information of face.At present,it has been widely used in various fields such as national defense security,social stability and so on.Facial feature point detection is an important step in face recognition,which directly affects the efficiency of face recognition.In addition,coordinates of facial feature points provide important geometric information for analysis and recognition of facial expression.Facial expression usually express people's joys and sorrows.The relationship between the coordinates of facial key points can help identify facial expression.Facial feature point detection is to automatically detect the salient points in the face images,such as the nose,mouth and other organs or face contour,which is based on face detection.With the development of computer vision,the techniques for facial point detection have become an independent research topic,and is becoming more and more mature.Its realization is a complex process,especially in complex natural scenes,the light,facial color,pose,shelter and other accessories such as glasses or other factors make it become more difficult.Usually,in order to reduce the difficulty of detection,people need to conduct image denoising at first,and then all the face images are normalized in the image preprocessing phase.There are many popular algorithms for face landmark detection,where Active Shape Model(ASM)is one of the most classical face feature points detection algorithms.ASM is a feature points extraction algorithm based on Points Distribution Model(PDM).Many scholars proposed some improved methods based ASM,which have been achieved very good result.But ASM is very dependent on its initial shape model.Random forests is an ensemble learning method for regression or classification.In the problems of classification or regression,random forests has the advantages of high detection precision,fast training speed,good result even if some data is loss.People often use random forests or combining random forests with other algorithms to learn image features for improving the accuracy of face landmark.However,random forests will over-fitting in some problems with the relatively large noise of classification or regression,resulting in poor result.Therefore,in order to further improve the accuracy of face landmark,this paper proposes a method that combining Gradient Boosting Decision Tree(GBDT)and facial features.GBDT is a machine learning technique used to solve regression and classification problem,which will produces a prediction model in the form of an ensemble of weak prediction models.The iterative process of GBDT is similar to that of Boosting,But GBDT can be used to optimize the loss function that can be differentiated.In this paper,we use GBDT to train classification models using the HOG,LBP,Gabor features and their fusion and cascade features in a large scale in order to find the best feature set.Experiments show that HOG is the best feature.Therefore,this algorithm named H-GBDT.And then compare the H-GBDT algorithm with SO-RF and Face++.The results of comparison show that H-GBDT algorithm has very good detection performance,especially in the frontal face images.The main contributions of this paper are as follows:1.Proposing a detection algorithm of facial points based on GBDT and HOG.2.Annotating 2000 images among LFW,and each image include 20 feature points.The meaning of feature points coordinate is the same as that in the BioID data set.3.Extracting Gabor,LBP,HOG,their fusion and cascade features from images on the BioID,LFPW,LFW datasets,and than engaging in large-scale experiments to prove the true that the HOG feature is the best feature for GBDT.The detection results of facial points with H-GBDT will be compared with those of SO-RF and Face++.In this paper,the error of detection can be controlled within 5%.Especially in the BioID data set,the error can be controlled within 3%.In addition,the advantage of this H-GBDT is easy to understand and realize.