Research Of AAM Model For Key Points Detection And Application In Multi-pose Face Recognition

Under the condition of camera uncontrolled environment,poseture becomes the most challenging factor in the current face recognition and this problem has not been solved well.The main reason is because the key-points detection is not accurate enough.There exists dislocation when feature matches.The thesis is mainly focus on the AAM detection model to obtained key facial points and applying its massage to multi-pose face recognition.In this thesis,we firstly introduce LBP,Gabor feature extraction algorithm,and then analyze classification ability of key points and global feature in different poses.To get accurate position of facial key points,we propose an improved multi-pose face detection algorithm based on AAM and mixed tree location algorithm.This algorithm firstly expand pose expansion to AAM:the texture and shape models are established in thirteen different poses.Secondly,this algorithm combine multi-pose AAM with the corresponding hybrid tree model,then obtain the automatic key point detection model.It makes full use of the complementary of advantages and shortcomings in two basic methods above,which not only satisfy the AAM for requirements of the initial key points,but achieve accurate key points in multi-pose face automatically."On the basis of the works above,we proposed two methods to solve multi-pose face recognition.They all most use of local feature around the key points to replace traditional global feature as the face representation for its good classification capabilities when face image has big change in posture.The proposed method use circular statistical area on the key points to improve the adaptability of the algorithm for face rotation against to previous rectangular area.Although some local features can maintain unchanged across a certain posture range,face message appears self-occlusion when the pose changed greatly.To better solve the problem of face information loss in great posture,the paper put forward pose classification strategy according to the face posture parameters and employed the principle of"extracting half-face feature".The posture parameters obtained by the "X" template-based estimated method,which formed by five face key points.Experiments were conducted on LFW and CAS-PEAL-R1 face database.Compared to previous global method with same feature extraction algorithm,the proposed methods on LFW achieved 90.45%and 76.67%recognition rate under the condition of supervised training and without training respectively.They all improved more than 10%than the past the method On R1.Results show that our methods have good effect to multi-pose face recognition.