Regression based joint subspace learning for multi-view facial shape synthesis

Multi-view facial image synthesis is an important issue in computer graphics, 3D facial image reconstruction and accurate face recognition. In this paper, we propose a regression based joint subspace learning method (RJSL) for automatic multi-view facial shape synthesis. This method synthesizes multiview facial shapes from one input facial image. In conventional joint subspace learning based multi-view facial image synthesis, the coefficients estimated by using the input image is directly used for multi-view facial image synthesis. In our proposed method, the coefficients are estimated by a regression method based on the coefficients of the input facial image. We first construct a original multi-view facial database. The different view image pair (e.g. 0 degree and 15 degree, 0 degree and -15 degree) are connected as a joint vector for corresponding subspace learning. The training data are divided into two groups: one for joint subspace learning and another one for regression of coefficients. And our proposed method trains by shape information and texture information. In this paper, the shape information is expressed by feature points. And the texture information is expressed by luminosity values of normalized facial image. Our proposed method uses the luminosity value as depth information. In experimental step, this method trains pseudo 3-dimentional shape information (x, y-axises: feature points, z-axis: luminosity values). Our proposed method realizes accurate multi-view facial shape synthesis by these two contributions.