Face Alignment In-the-Wild

Face alignment,also knowns as fical landmark detection,is a focus issue in computer vision.It is a task to locate fiducial facial landmarks,such as eye corners,nose tip and mouth corners in a face image.Face alignment in the wild means enhancing the robustness of face alignment alogtithms to large head poses,various expression and lightning as well as partial occlusion.It can provide accurate and semantic information of face shape,and help achieve geometric image normalization and feature extraction.Therefore,face alignment becomes an indispensable part in such face analysis tasks as face recognition,facial expression analysis,human-machine interface and 3D face modeling.In practice,due to the differences in facial expressions,head poses,illuminations and partial occlusions,the issue of face alignment still confronts great challenges.Under this background and with studies of current face alignment alogrithms,we propose four different alogrithms to handle face alignment with large poses,various expression and lightning as well as partial occlusion,and enhance their robustness in the wild.(1)The regression-based face alignment alogrithms are all in need of an initialized shape,and the error of initialized shape directly affects the convergence and accuracy of these algorithms.In order to construct a more accurate initialized shape,we propose face alignment on Local-Shape-Based Combined model(LSBC)and face alignment by Coarse-to-Fine Shape Estimation(CFSE).With coarse estimations of pose and expression,LSBC can choose proper local shapes to construct a combined shape,which reduces the initialized error and improves the performace of face alignment.In CFSE,the main landmarks will be detected firstly and then used to help locate the entire shape.Then,it constructs an independent head pose classification model based on Convolutional Neural Network to estimate and classify head poses.With the classification result and the detected landmarks,a more accurate shape will be constructed,meantime the initialized error will be further reduced.Validations on three public face alignment datasets show that the proposed two algoritms are more robust to poses and expressions.(2)The multi-view model based face alignment algorithms firstly partition the optimization space into multiple domains with homogeneous descent,and then the domain-specific regressor will be used to update the shape belonging to this domain to approximate the ground-truth shape.This domain splitting strategy has the following two disadvantages: 1)Due to differences in testing and training,the splitting strategy lacks a probability of fault-tolerance.2)They update shapes without considering their inter-complementary and regularisation role which limit the performance of face alignment.Hence,we propose face alignment algorithm by Component Adaptive Mechanism(CAM).A probability-based fern classifier is adopted in the partition of the optimization space into multiple domains of homogeneous descent,and a training strategy based on dominant set approach is used to train a stronger domain-specific regressor.Then a component adaptive mechanism is used to fuse results of different domain-specific regressors to further update the shape.Validations on three public face alignment datasets show that the proposed two algoritms are more robust to poses and expressions.(3)The proposed face alignment algorithm by cascaded regression and de-occlusion is used to handle partial occlusion problems.As any part of the face can be occluded by arbitrary objects,it is difficult to locate the occlusion accurately.Since the facial occlusion is usually very complicated in real scenes and face analysis tasks largely depend on regions around landmarks,it is difficult and unnecessary to recover the whole face.Therefore,we only recover the occlusion region around landmarks(image patch),and the recovered image patches will be combined with the original image to obtain a new one.Then this new high resolution image will be used as input of deep regression network to get more effective regressor.By cascaded deep regression and face de-occlusion,the landmark is located and recovered progressively.Each stage benefits from the previously recovered face image and the final performance of face alignment is improved.Validations on three public face alignment datasets show that the proposed algoritms are more robust to partial occlusion.