Facial Landmark Localization Algorithm From Coarse-to-Fine

Facial landmark detection has been a popular research topic over the past few years due to its importance in various face analysis applications such as face recognition,facial expression classification,and age estimation.Recently,there have been extensive improvements of the face alignment algorithms,but is still facing challenge that facial images with large variations on face expression,head pose,illumination,and partial occlusions.To solve these problems,We present a novel face alignment framework based on coarse-to-fine.Coarse is coarse-localization:obtained by DCNN(Deep Convolutional Neural Network).Fine is fine-localization:use coarse-localization as initial shape,then refine the shape in a cascaded manner.In this paper,we propose a better learning based approach for face alignment based on two insights:initial shape extraction and regression parameters selection.Considering the self-study character of DCNN without initial shape,we use DCNN with few number layers to obtain initial shape for facial image.It overcomes the disadvantages of complicated calculation for deep network with many layers,and the efficiency of face alignment has been improved.During the process of network training,we use many optimization methods to get better result for initial shape such as weight initialization and parameter optimizing.We find the shape learned by network is not able to meet the requirements of accuracy.So we use it as initial shape for regression,the experiment result indicates that our approach significantly outperforms mean-shape initialization.Then we use improved K-means algorithm to determine shape candidate sets in order to fine-tuning the shape in it.More over,We use SIFT feature extraction,comBining with the approved SDM(Supervised Descent Method)regression method,update the shape of face stage by stage to get closer to ground truth until converge.We choose 3 as our stage number.In this paper,We evaluate the performance of the proposed method on different data sets.The experimental results on benchmark databases such as 300W,LFPW and HELEN demonstrate that our proposed method outperforms previous work for facial landmark detection.The average localization error rate on 300-W is 6.34%,our approach outperforms the comparative methods such as ESR(Explicit Shape Regression)(7.58%),traditional SDM(7.5%),and robust to large variations on face expression,head pose,illumination,and partial occlusions.