Research Of Face Alignment Based On Cascaded Shape Regression

Face alignment or facial landmark detection is an important issue in Computer Vision.It plays a critical role in the systems of face verification and recognition,facial expression recognition,facial attributes analysis,etc.Face alignment aims at automatically locating semantic facial landmarks such as eyes,nose,mouth and chin.It has made rapid progresses in recent years,several methods have reported close-to-human performance on benchmark datasets.However,it remains a challenging problem due to the complex variations in face appearance caused by pose,expression,illumination,partial occlusion,etc.This paper focus on the face alignment methods based on cascaded shape regression,and puts forward the corresponding improvement for some problems existing in popular face alignment methods.In this paper,the main research work is as follows:1)On the basis of the traditional explicit shape regression,the algorithm in this paper divides the face alignment task into two steps: four facial key points and the overall shape of regression based on coarse-to-fine manner.Reduce the computational cost of feature selection through feature pool to ensure to eliminate irrelevant features.According to the change of the alignment error in the training process,the number of training regressors is controlled by a "smart stop" method to speed up the convergence rate of the algorithm and improve the accuracy of the algorithm.2)Traditional face alignment methods mostly use the shape indexed feature,but these methods are weakly robust in the face of complex facial posture and expression.This paper adopts the relative-indexed way index pixel features under the set spatial dependency hypothesis.In principle,the following learning-based approach should be better.In order to learn the discriminative features,this paper uses the random forest to encode pixel-difference features into sparse binary features.Finally,the paper uses local shape regression to complete face alignment,and relative-indexed feature and local shape regression are based on the same spatial interdependency hypothesis.3)Thanks to the deep convolutional neural network powerful ability of nonlinear modeling,the first level of the cascaded shape regression associates the face alignment with the associated tasks head Pose,gender,and occlusion in the one depth convolution neural network.The second level of the cascade regression uses the first level prediction results to refine each component landmarks by using the shallow network of the same structure.This method which makes full use of multitask learning and cascaded shape regression from coarse-to-fine achieves a high-precision by s trengthening the non-linearity of deep convolution networks.In this paper,the benchmark datasets are used to verify the performance of the algorithm,and the algorithms are evaluated from the averaged alignment error,the failure rate,and the cumulative error distribution,etc.