3D Face Alignment And Reconstruction

With the development of computer technology,3D face models are more widely used in the applications, such as face recognition and computer animation. However, on one hand, there are various sources of3D face models which are not aligned and inconvenient to use. On the other hand, the capture of3D face data is still expensive and inconvenient. Hence in this thesis, new methods for3D face alignment and reconstruction are proposed.In3D face related applications, the input faces are usually required to be well aligned to a reference face. Because of the various orientations and scales, and the misleading non-face areas of input3D face data, automatic3D face alignment is still challenging in3D image processing field. In this paper, we present an automatic approach to align the input3D face through nose detection. Firstly, vertices of the input face are clustered according to the property of normal vectors. Secondly, a graph-based partitioning algorithm is proposed to further partition the input face into several patches and an SVM detector is trained to select the patches that belong to the nose region. Then, coarse alignment is achieved by approximating a3D affine transformation between the input face and the reference face. Finally, the ICP algorithm is used to align the face more precisely. Our method can deal with3D faces with arbitrary orientations and scales, and yield high-accuracy alignment between faces.In the first stage, we coarsely align the input face to a reference face by3D affine transformation based on nose detection. In the second stage, we refine the alignment by employing the ICP algorithm after re-sampling the input face. For re-sampling, a facial mask that comprises a set of pre-defined feature vertices is detected by utilizing the LCC coding scheme and ultimately, dense per-vertex and per-triangle correspondences between the input face and the reference face are established. Experimental results show that our method is efficient and achieves high accuracy.After preprocessing2D face images and corresponding3D face models, a coupled radial basis function network (C-RBF) can be trained. Meanwhile, a set of facial landmarks is defined and detected from each2D face image and the corresponding3D face model, respectively. Adaptive Appearance Model is used for2D facial landmarks detection and Local Coordinate Coding for3D facial landmarks detection. Finally, utilizing facial landmarks as local constraint, a3D face model is reconstructed from a2D face image.Experimental results show that our3D face alignment method is robust and our3D face reconstruction method is efficient.