A Study On Face Alignment Using Local Hough Voting Method

Face alignment is a computer vision technology for identifying the geometric structure of humanface pictures and he plays an important role in the face processing system. Many face processing post-tasks require the face alignment. However, the face is a kind of complicated three-dimensional non-rigid natural structure target, and the facial features are vulnerable to face pose, facial expression andthe external light or effect of noise. So, all of this increase the difficulty of the accurate face alignment.Consequently, robust accurate face alignment research in reality is of great significance.First this paper makes a thorough review analysis on the methods of face alignment in recent years.And then we focus on a broad class of object detection model: part based models. The proposed methodof ours uses the fitting methods of one of the constellation model: Active shape model for reference.Meanwhile, we use the voting method of one of the non-constellation model: Implicit shape model forreference. Additionally, we combine the two ideas together and improve the precision of face alignmentto a certain degree. The main contributions of this work are as follows:1. We introduce the probability output discriminative appearance model. Generally many re-searchers use the generative models in the ordinary face alignment, and they make the assumptionthat the local facial appearance follows the single mode Gaussian distribution. However, the fa-cial appearance is vulnerable to the effect of external light and facial expression. The single modeGaussian distribution is hard to depict the distribution of the facial appearance.We improve thisby introducing the probability output discriminative appearance model and improve the searchaccuracy of the local feature point to a certain extent.2. We introduce the voting model based on the anchors. When dealing with the face feature points,we detect the facial points which easy to detect firstly. And then generate the voting model ac-cording to the spatial information between the feature point and the anchor point.3. In order to further improve the precision and robustness of our model, we fuse the appearancemodel and voting model based on the multiple output ridge regression method.In order to illustrate the efficiency of our method, we test and analysis it on thePUT XM2VTS BioID and Talking face database separately. The results show that our method im-proves the precision of face alignment to a certain extent and achieves the desired objectives.