Research Of 3D Facial Landmarking Techniques Based On Local Feature

As an important part of 3D face recognition,3D facial landmark localization is widely used in the region of face alignment,face recognition,facial expression recognition,facial shape analysis,segmentation and labeling of facial parts,facial region retrieval,facial mesh reconstruction,face synthesis,face animation,etc.In this paper,we have done research on the local feature that distinguishes landmarks from other points based on analysis and summaries of the existing 3D facial landmark localization algorithms.We propose two approaches to 3D facial landmark localization that respectively detection 7 and 14 facial landmarks.The main work and innovations in our paper are as follows:1)An algorithm for 3D facial landmark localization based on two-step keypoint detection is proposed.In the algorithm,landmark localization is devided into two subproblems solved independently,keypoint detection and labeling.And a two-step keypoint detection method is proposed.First,coarse positions of landmarks are located by applying the supervised descent method on depth images.Extracting the neighbourhood of landmarks' coarse position as keypoint region.Second,an approach which combines multiple local descriptors is proposed for keypoint detection by filtering out the subset of facial points within the keypoint region.At the stage of keypoint labeling,combinations of candidates are generated from keypoints and candidates in the combination which fits the facial landmark model best are label as the landmarks.2)An algorithm for automatic localization of landmarks on 3D face by using a feature fusion based constrained local model is proposed.To make better use of local feature in 3D facial scans,a 3D mesh feature based constrained local model is proposed.The model merges10 local descriptors of 3D facial mesh and classifiers are trained for every landmarks as responses of the local model.To locate the facial landmarks,a modified method for model fitting is proposed.First,coarse positions of landmarks are achieved by applying supervised descent method on depth images and used as the initial positions of landmarks in the proposed model.Second,model parameters are updated iteratively by using regularized landmark mean-shift algorithm based on a sequence of windows with descending width and landmarks are finally aligned.A series of experiments are designed for algorithms proposed in this paper based on FRGC v2.0 and Bosphorus dataset.Algorithms are evaluated thoroughly and compared with several state-of-the-art approaches.The experimental results strongly demonstrate the effectiveness of the algorithms proposed in this paper,which lay the foundation for the better application of 3D facial landmark localization.