Automated 3D Face Recognition

Face recognition technology is the focus research of the biometric identification .Because of the influence of illumination, pose variation and expression, the effect of 2D face recognition has been greatly hindered. 3D face model can provide more information than 2D face image, some of the extracted features are invariant under a rigid transformation, and uneasily influenced by the make-up and light. The use of 3D surface registration algorithm can overcome posture changes well. Facial movements through the synthesis of 3D model ,to some extent ,can overcome the changes in expression. So 3D recognition is a relatively robust face recognition algorithm, and will receive extensive attention.This paper presents an automated method of 3D face recognition. Becauseof change of the posture and there is redundant information, every mask need doing a few preliminary processing. In the preliminary processing we will complete estimate the initial nose tip, cut face model to remove redundant information and calibration the posture of the face model. Removing the redundantinformation can reduce the subsequent calculation, at the same time can reduce the noise points and will reduce the difficult of feature extraction. In the later work, we will extract the symmetry profile first, and then extractthe feature points on the symmetry profile including the nose tip point, the nose bridge point, the philtrum point, the midpoint of the mouth and chin. Except the feature points on the symmetry profile, we also extracted the feature points in the main organs on the face such as, four eye corners, two mouth corners, two external nose wing corners and zygomatic. After all feature points have being extracted, we use the triangles consisted of the featurepoints to design the first layer refuse classifier. The face model which can pass the first level refuse classifier we will use the symmetry profile to design a second layer refuse classifier, and select the average Hausdorff distance as the similarity measure . We only apply the global ICP to which have passed the two-storey refuse classifier. We still choose the average Hausdorff distance as the similarity measure. The face model, in the library, which has the smallest average Hausdorff distance to the face model being recognized is the ultimate outcome of the identification.After inputting a 3D triangle mesh of a face, we use the basic fact that the nose tip has a maximum projection in some direction ,and divide the first two quadrants of the XZ coordinate plane with a step of 2 degree. To every angle, we will apply a rigid transformation on the face model. We transform the face model to this pose, and select the point which has the largest Z value as the initial nose tip candidate. Then the point which has the largest Z value in all of the candidates will be selected as the initial nose tip. After the initial nose tip having been extracted, we will make it as a source point to cut the face model with a very radius in a appropriate distance measure. As the human face is no-rigid, the expression change will cause the displacement of the organs in the face. In order to make the important information in the face no to cut out, we select the geodesic distance as the distance measure. After many experiments we choose 110mm as the radius. That is, all the points in the face model which geodesic distance to the initial nose tip less than 110mm will be compose the 3D geodesic mask F.Then we estimate the pose of the 3D face model. We first apply PCAto the cropped 3D face geodesic mask , acquire three principal component of the mask: the vertical direction of the mask, the horizontal direction of the mask and the normal of the least square fitting plane of the mask. The three direction determine a coordinate, which take the barycenter of the mask as the origin point, X-axis represent the horizontal direction of the mask, Y-axis represent the vertical direction of the mask, Z-axis is decide by the right hand rule. We take this coordinate as the first pose estimate of the 3D face model. We solve pose normalization by rotating the 3D face geodesic mask to this coordinate.In order to obtain the symmetry profile of the 3D face geodesic mask, we first extract the symmetry plane of the 3D face geodesic mask. We take the YZ coordinate plane as the mirror plane, and acquire the mirror representation F_m of the original 3D face geodesic mask, then we apply ICP method to F and F_m , and take the least square fitting plane as the symmetry plane of the 3D face geodesic mask, which pass through all the midpoint of the final corresponding pairs of point. The intersection between the 3D face geodesic mask and its symmetry plane is the symmetry profile of the mask. In this process, we improve the existent algorithm of extracting symmetry profile and the ICP method. At first, we extract the symmetry plane of the simplified 3D face geodesic mask, which can largely reduce our computation, and then in the ICP method, we use the k-d tree data structure and prohibit the boundary of the mask to participate the registration, and this can reduce the probability of ICP converge to local optimization.We join the two end points of the symmetry and obtain a line l, and take the point on the symmetry profile as nose tip pointp_t, of which the distance from l is the largest, and the point on the symmetry profile as nose bridge point p_b, , of which the distance from is the first smallest in the direction of head forehead, and the point on the symmetry profile as philtrum point p_t, of which the distance from l is the first smallest in the direction of head mouth.If ||p_t -p₁|| < ||p_t-p₂|| , then p_b = p₁, p_p = p₂; because the distance from the nose tip to the nose bridge is bigger then the distance from the philtrum to the nose tip.We sort the points on the symmetry profile in the direction from nose tip to nose bridge, that is the points on the symmetry profile are sorted in ascending order according to Y value. Start form the point with the minimum Y value, the first point with the maximum distance to l is the chin. To extract the feature points of the organs on the face model, our algorithmmainly based on curvature and statistical model. Firstly ,choose an appropriate statistical model and give the feature point an initial position accordingto the common characteristics of human. To every statistical model there is an unique reference point, and the reference point often choose the point which has been extracted and close to it. When exact the inter eye corners, we choose the nose bridge as reference point. When exact the outer eye corner ,we choose the inter eye corner in the same side as reference point. When exact the zygomatic .we select outer eye corner in the same side as reference point . When exact outer corner ,we choose mouth mid point as reference point . When exact nose wing ,we choose the philtrum and inter eye corner in the same side as the reference point. This scheme can reduce the search domain and reduce calculation. In the proceeding of feature exact based curvature we also use the Z value and curvature of the centroid of a triangle. Then give them some weights to obtain a multiplex value. This value can describe the character of the feature points well. Finally, we get the feature exacting algorithm.Then we use the triangle which is made up by the feature points to design refuse classifier. In this paper , we use the position in the space ,the shape and size of the triangle to measure the degree of similarity. We use the dihedral angle and the Euclidean distance between the correspondent point to measure the similarity of the position in the space of the tow triangle, and use the inner geometric value to measure the similarity of the shape. To the size, that just use the length of the edge of a triangle, if the tow triangles are very similar to each other in the first tow hands. Before we calculate the similarity of tow triangles we should transform them to their intrinsic coordinates respectively.The face model which has passed the first level refuse classifier ,we will use the symmetry profile to design a second layer refuse classifier. And select the average Hausdorff distance as the similarity measure . When the average Hausdorff distance between the two symmetry profiles smaller than a very value, we will apply a global ICP registration to the two face models. We will apply a global ICP registration to the two face models when the one in the library can pass the two refuse classifiers .At first we use the rigid transformation which obtained in the last step to get a coarse registration .This scheme can provide a good initial position to make the ICP algorithm convergence. We will use the average Hausdorff distance as the similarity measure too. The face model, in the library, which have the smallest average Hausdorff distance to the face model which is recognized is the ultimate outcomeof the identification.