A new framework for 3D face reconstruction for self-occluded images

Abstract

Human heads are three-dimensional objects in 3D space with variations in position and in its structure. Consequently, 3D face modelling is largely acknowledged in face recognition application for uncooperative subjects. Structure from motion (SfM), 3D face reconstruction technique model a 3D facial shape by means of multiple 2D images sequence. In view of self-occluded 2D face image, this technique is susceptible to point correspondence error reducing its performance. To eliminate point correspondence error a matrix called shape conversion matrix (SCM) is appraised to obtain the true location of self-occluded facial feature points (FFPs). In the proposed system, a new SfM method called multi-stage linear approach is adopted. A novel face alignment algorithm called RASL is incorporated with the system. A more resourceful feature localisation technique called simultaneous inverse compositional algorithm is modified. A generalised polycube trivariant spline-based 3D dense mean model adaptation is integrated. By applying these methods, a proficient framework for robust 3D face reconstruction for self-occlusion is proposed in this paper.