3D facial reconstruction using micro and macro level features
Rajeshkannan Sundararajan, Prakash Veeraswamy Radhakrishnan, Saravanan Govindarajan, P. Ezhilarasi · Research Square · 2022
Abstract A fundamental problem in computer visions and graphics, which has various applications such as animation and face recognition, is the reconstruction of 3D face models using 2D images. The main reason due to which this poses a challenge is the loss of information during camera projection. The initial step of this process is to compute a coarse estimation of the 3D face of the target by fitting a parametric face model to the input image based on an example. Afterwards, a smooth deformation that captures the medium scale facial features is applied to the coarse face model in order to enhance it. The lighting and reflectance parameters from the enhanced model is also estimated following which, an analysis-by-synthesis framework for face recognition with variant PIE (pose, illumination and expression) is proposed. In order to reconstruct a personalized 3D face model from a single frontal image with normal illumination and neutral expression, an integrated micro and macro level feature matching technique is introduced a robust probabilistic method for the estimation of macro features in Bayesian estimation theory, with a prior Markov random field model for the assigned disparities. This process, with the appropriate boundary conditions, is also used to estimate disparity in problematic regions of stereo pairs, such as occluded areas and non-textured (homogeneous) regions.Subsequently, to characterize the face subspace, realistic virtual faces are synthesized based on the personalized 3D face. Ultimately, the process of face recognition is conducted based on these virtual faces.