Facial Landmark Localization For Face Recognition

Facial landmark localization is an important part of face recognition system, whose result can also be applied to other fields, such as driver fatigue detection, automatic face tracking and facial animation. However, facial landmark localization is susceptible to facial expressions, face poses, partial occlusion and illumination conditions.Taking all factors into consideration, we propose a new approach to localize facial landmarks based on a face detection algorithm. Its crucial idea is "from few to many, from coarse to fine". At the same time, we would localize inner points and contour points separately to reduce the mutual interference between them. The main contents of this paper are described as follows:Firstly, based on the relevant knowledge of CNN, the direct contact between facial high-level features and coordinates of facial landmarks is established by DCNN. Thus, we can predict the positions of two eye centers, nose tip and two mouth corners simultaneously on a pre-processed face image.Secondly, in order to mitigate the effects of unstable face detection and face poses, h, the parameters of affine transform from predicted five points to average five points, should be solved. Then we use h to transform the average shapes of inner points and contour points. The transformed shapes are initial positions for localizing facial landmarks precisely.Thirdly, according to SDM algorithm, we train the descent directions and biases from initial positions to ground positions making use of simplified SIFT feature. The training result and the SIFT feature at initial positions are used to gradually optimize the initial positions, which are derived from predicted five points. After several iterations, the precise coordinates of 49 inner points and 17 contour points are obtained separately. That is to say, based on the predicted five points we can localize 66 facial landmarks accurately. Then we achieve the goal of this paper, which is what we want.Finally, experiments on CAS-PEAL-R1, FERET face database and a validation set show that our approach improves the localization performance in terms of both accuracy and robustness. And the accuracy rate of points around the eyes on CAS-PEAL-R1 face database is 99.23%.