Research And Application On Technologies Of Face Alignment

Recently,technologies about face processing has a great progress,more and more intelligent equipments that based on the face recognition technology are infiltrated into our daily life and improve life quality significantly,for example,in finance,public security,entertainment and social network,etc.Face alignment is an essential step of face recognition and can be applied into some related technologies,such as,facial beautification,facial landmark tracking,impression recognition,age estimation and so on.Face alignment aims to locate some specific landmarks(points on the eyes,nose,mouth and cheeks)and the outlines of a face are drew by using these landmarks,then we can get an intuitive impression from these outlines.With the increasing number of various applications,people not only want the applications to be more fast and more accurate,but also can be used in various complex scenarios,such as bad illumination,heavy occlusions,large pose that far from frontal face.Unfortunately,accuracy of face alignment is unsatisfied in these situations.To deal with the above problems,we studies the cascade shape regression framework for face alignment and its application,the contributions of our paper are as follows:1.A salient guided face alignment model is proposed.Both common problems(occlusion,pose and illumination)of face alignment and the initial problem of cascade shape regression model are considered,we propose a salient guided shape regression method for face alignment.Because the salient landmarks can express semantic information of a face and have strong robustness to the above common problems,we apply cascade shape regression to locate salient landmarks firstly.Then,we search the faces which have similar salient landmarks with the input face from the training database and compute a new face by combining these similar faces.The computed face is regarded as the initial face for the regression and the final result is obtained.Moreover,both the distances between two points and between two face shapes are considered,a new evaluation metric is proposed for face alignment.2.A face alignment method under large pose is proposed.In the uncontrolled environment,people may have a variety of head gestures,especially,in the situation of a wide angle of the side face,the alignment algorithms cannot give satisfied results.In order to solve this problem,we design a new feature and a scheme that is used to select good features.The new feature is applied in regression decision tree and faces with large poses are augmented to diversify the training data.Our approach contains two stages,new feature and data augmentation are applied to detect the salient landmarks precisely in the first stage,feature selection constraint is applied in the second stage to boost the final result.Experimental results show that our method is fast and high precision and have a big improvement over the baseline model.3.A feedback cascade shape regression framework for face alignment is proposed.Cascade shape regression models are sensitive to the hidden landmarks and the unified initialization can make the method falling into local minima.To deal with the above problems,we design a new pipeline,that is salient-to-inner-to-all,to progressively detect the facial landmarks.By applying the feedback strategy,the preliminary results are fed back to the initial results of the salient point positioning,and then the progressive process is repeated to obtain the final results.We propose a pose-invariant shape retrieval method to assist the transitions from salient landmarks to inner landmarks and from inner landmarks to all landmarks.Results on two public databases show that our method improves the initialization and gets results close to the best deep learning algorithm.