Research On Face Alignment Algorithm Based On Supervised Descent Method

Face alignment is a process of automatically locating facial landmarks on an facial image.It is a key step in face recognition and the basis of many facial information processing tasks,such as face tracking,expression recognition,head pose estimation,etc.Because of many complex factors,such as occlusion,expression,illumination,the face alignment task still faces great challenges.Cascade regression has become one of the most popular and advanced methods of face alignment due to its excellent performance in precision and speed,and the Supervised Descent Method(SDM)is a classical cascade regression method,which is evolved from Newton method and has been well applied in face alignment.However,SDM algorithm still has room for improvement.The SDM algorithm is improved from two aspects in this thesis.The main works are as follows:(1)When SDM algorithm adopts different features,different results will be produced,and the influence of different features in each iteration is also different.Therefore,the SDM algorithm is optimized by multi-feature selection in this thesis.In each iteration of training,three features are used for training and shape updating respectively,and the mean square error between the coordinates of three updated shapes and the coordinates of real shape is calculated.Which value is lowest,the corresponding feature will be selected in the iteration.Experimental results indicate the effectiveness of our method.(2)In some complex conditions,such as exaggerated expression or excessive head pose,because of the difference between the initial shape and the real shape is too large,SDM is unable to achieve good results.To solve this problem,a coarse-to-fine SDM(CFSDM)method is proposed in this thesis.This method predicts the approximate coordinates of the facial landmarks with a simple Convolutional Neural Network(CNN)in advance.Then SDM will take the coordinates as its initial shape's coordinates and continue the subsequent alignment.Experimental results show that the proposed method can get higher accuracy and achieve better alignment results than the original algorithm under some complex conditions.