Automatic Framework Of 3D Face Recognition

Face recognition is one of the most popular research fields at present. Recently, most of the researchers on face recognition are based on 2D face images. Because of the influence of illumination, pose variation and expression, the improvement of recognition accuracy of 2D face recognition is greatly impeded. This makes it still difficult to build a robust face recognition system. 3D model holds more rich information than 2D image, so implementing face recognition on 3D face model is one of the effective approaches to tackle the present problems.This thesis proposes a precise, fast, and highly effective automatic framework of 3D face recognition, it contains two parts: preprocessing, feature detection and recognition. The preprocessing part composes two steps: crop the 3D face model and estimate the pose of the face, and the feature detection and recognition part contains extract the symmetry profile of the face, detect the feature points and face recognition. In the preprocessing part, by cropping the 3D face model, acquiring the 3D face mask and estimating the pose of the face, we solve pose normalization. Pose normalization plays an important role in the following recognition. In the feature detection and recognition part, we improve the existent algorithm of extracting symmetry profile and the ICP method, and obtain the symmetry profile of the 3D face geodesic mask high efficiently and precisely, this not only can help us detect the feature points precisely, but also provide us a method of face recognition. In the face recognition phase, based on the ICP method, we take the symmetry profile as a classifier, take the Mean Hausdorff distance between the 3D face models after registration as their similarity, we finally finished the 3D face recognition.The input data of this thesis is triangulation mesh model of 3D face. For a arbitrary input 3D face model, we first extract its least square fitting plane, then take the plane as reference plane, find the nose tip of the 3D face model. In order to eliminate the influence of the expression of face, we compute the geodesic distance from every point of the 3D face model to the nose tip point. Take the nose tip point as origin point, we retain all the points in the 3D face model, the geodesic distance from the nose tip point of which is less than 110 mm, all these points finally form the 3D face geodesic mask S .Then we estimate the pose of the 3D face model. We first apply PCA to the cropped 3D face geodesic mask S , 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 vertical direction of the mask, y-axis represent the horizontal 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 new 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 xz coordinate plane as the mirror plane, and acquire the mirror representation S m of the original 3D face geodesic mask, then we apply ICP method to S and S 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. After doing this, we can obtain the symmetry profile of the 3D face geodesic mask high efficiently and precisely.The symmetry profile of the 3D face geodesic mask pass through many distinctive facial features: the forehead, the bridge of the nose, the nose itself, the philtrum, the mouth and the chin. Of these, the nose is the most robust geometrical feature on the symmetry profile (and on the entire face). It is least changed under different expressions and contains clearly distinguishable points. We can use the structure of the symmetry profile to extract feature points on it. 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 point pt , of which the distance from l is the largest, and the point on the symmetry profile as nose bridge point pb , of which the distance from l is the first smallest in the direction of head forehead, and the point on the symmetry profile as philtrum point p p, of which the distance from l is the first smallest in the direction of head mouth. Formulation as follows: As we don't know the coordinate of the 3D scanner and the pose of human face, so the vertical direction of the symmetry profile is unknown. If | pt ? p1 |< | pt ? p2|, then pb = p2 , p p= p1 ;else pb = 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. In this way, we determine the vertical direction of the symmetry profile.The three feature points p_t,p_b,p_p on the symmetry profile uniquely determine a coordinate, take pt as origin, V_tb = ( p_b - p_t )/ ||p_b - p_t|| represent the y-axis, V_tp= ( p_p - p_t )/ ||p_p - p_t||, the cross product between V_tb and V_tp represent x-axis, and z-axis decide by right hand rule. We call this coordinate the intrinsic coordinate of the 3D facial mask. In the process of comparing the similarity between two 3D facial masks, we first extract the intrinsic coordinate of each one, and then rotate them to a uniform coordinate before comparing the similarity between the two masks.At last, we consider comparing the similarity between two 3D facial masks. As the shape and the vertex number of the facial mask is different, the registration between the symmetry profiles and the masks is partial matching, so we consider the Hausdorff distance metric. The Hausdorff distance is a method for computing the similarity between two point cloud, the advantage of it is that it permit partial matching between two surfaces, that is to say, the Hausdorff distance can compute the similarity between two surfaces even when the two surfaces are not fully match. The standard Hausdorff distance is tend to infect by the contaminative point, in this thesis, we adopt the Mean Hausdorff distance(MHD) proposed by Dubusson and Jain, because its simplicity and steady.The symmetry profile provides us a method for face recognition. In fact, in our system, we use the symmetry profile of the 3D facial mask as a reject classifier. In the 3D face database, we not only save the 3D facial masks, but also save the symmetry profiles of each mask. When input a 3D face model, we first crop it to get a 3D facial mask, estimate its pose, and then extract its symmetry profile, we compute the MHD between this profile and the profiles in the database as their similarity after apply ICP to them. Only the masks correspond to the profiles in the database of which the MHD is less than a prior given threshold can participate the next process.After the symmetry profile recognition, we then consider the similarity between two 3D facial masks. We first apply RI CP and t ICP obtain from the last step to the whole mask, this can guarantee a good relative position between the two masks, and then apply ICP to them, the MHD between the two 3D facial masks after ICP is take as the similarity between them. And in this way, we finally finished the recognition of the 3D face.