Study On Key Technologies Of 3D Face Modeling Based On Morphable Model

Human face is the most important part of human being to recognize identity.It plays an important role in information exchange and emotion expression,which motivate many researchers to pay a great attention to the depiction and description of human face.With the development of the computer vision and computer graphics technology,the perceive ability to face and expression of the human vision system causes researchers to consider how to let the computer vision system have the same power.So,many researchers start to study the human face.As the most important technology of the human face study,3D face modeling technology becomes the study hot spot.As the research work goes further,3D face modeling technology gradually plays an increasingly important role in many ap-plication fields,such as information countermeasure,face recognition,human-computer interaction and movie animation,etc.The outlook of 3D face modeling technology be-comes more wider,and whose application values become bigger.In the 3D face modeling technologies,morphable model based technology is one of the most important 3D face modeling technology which has the greatest effect of 3D face modeling.Morphable model can be used in many areas such as face modeling,face recognition,human face image analysis,expression clone and facial animation,etc.Nev-ertheless,the creation procedures of morphable model are very complex,which hinder its widely usage.According to these complex problems which hinder the widely usage of morphable model,this paper carry out the research works on the problems of how to detect the face area of the human face samples and how to achieve dense correspondence between human faces.The main works and innovations of this thesis are examined in the following.(1)Nose tip detection for 3D faces based on multi-angle energyThis paper proposes a novel method to detect nose tip based on multi-angle energy.The method does not rely on curvature,needs not training,can handle large rotations,occlusions,expression changing and the scale variance of human faces.The core of the method is the multi-angle energy(ME)idea proposed by author.Utilizing the multi-angle energy,the ME value of each point is computed and sorted in descending order.Then the first m points in the descending order list are obtained and hierarchical clustering method is used to cluster these points.In the first h largest clusters,one point with the largest ME can be found.For all scales of the scale-space,a series of such points which are treated as nose tip candidates can be gotten.For these candidates,hierarchical clustering is applied again.In the obtained largest cluster,the mean value of ME is computed.The ME of nose tip will be closest to the mean value.(2)3D face localization based on endocanthion detection and based on ellipse fittingTo localize the face area in 3D face samples,this paper proposes two methods to localization 3D face,one is based on the endocanthion detection,another is based on ellipse fitting.In the first method,the endocanthion is detected by utilizing a modified multi-angle energy.Then the distance between nose tip and endocanthion is determined.A sphere whose radius is based on such distance and whose center is on the nose tip is created.3D face localization can be achieved by using the sphere.The second method turns the problem of 3D face localization into 2D field.The method presents ways to determine the facial midline and the iso-curves which are parallel and perpendicular to the facial midline.During analyzing the facial midline and the iso-curves,the paper defines a modified resultant force and its signed resultant force strength,then use them to locate the positions of a series of elliptical boundary points.Using these elliptical boundary points can fit an ellipse in the 2D representation of the 3D face scan by solving an over determined system of equations in the least square sense.Finally the 3D face localization can be achieved by using the fitted ellipse.(3)3D face registration based on three level fittingThis paper proposes a 3D face registration method based on three level fitting.Such method first constructs the conformal term,regular term,correspondence term and sparse term,and using them to construct the optimizing cost function.The 3D face registration procedures are divided into three level:rough non-linear fitting,fine linear fitting and final detail fitting.After these three level fitting procedures,3D face registration can be achieved.To reduce the converge time and let the result of 3D face registration close to dense correspondence,this paper also proposes a 3D face sparse correspondence method based on Mobius transformations.Such method turns the problem of finding sparse corre-spondence of 3D faces in Euclidean space into the Mobius space,utilizing Mobius voting to achieve sparse correspondence between 3D faces.(4)Establish 3D face dense correspondence based on face landmarksThis paper first proposes a novel method to detect landmarks automatically for high resolution 3D faces without learning or training,solely using geometric information.To detect landmarks,global and local constraints for each landmark are extracted by consid-ering relative positions of landmarks and differential geometric features,then landmarks are detected by combine these constraints.After the 3D face landmarks detecting proce-dure,the registered 3D faces are treated as 3D face samples,the correspondences between 3D face landmarks are used to construct the landmarks term and is combined into the 3D face registration optimization cost function.Using the new optimization cost function,the 3D face registration based on three level fitting is performed again.With the constraints of the landmarks' correspondence,3D face dense correspondence is established.