Research On Occlusion And Pose Robust Facial Landmark Localization

Face alignment can be used for 3D face reconstruction,face animation,face recognition,expression analysis,virtual makeup,craniofacial surgery,fatigue monitoring,etc.,with a wide range of application background and great research value.The current algorithm has a large room for improvement in key point positioning accuracy under occlusion and large pose.Based on the above problems,this paper proposes an occlusion robust facial landmark localization algorithm and a pose robust facial landmark localization algorithm.Finally,a face-driven animation system is realized by using a common monocular camera.The main innovations and contributions of this peper are summarized below:1.An occlusion robust face keypoint localization algorithm is proposed.A three-stage network structure design is proposed for occlusion problems,including feature extraction part,first stage coarse positioning module,second stage fine positioning module and the three-stage binary coordinate regression module,each part designed a unique loss function for dealing with face keypoint positioning under occlusion.The test results of the public dataset show that the algorithm can better locate the key points of the occlusion area compared with the traditional face key location algorithm based on binary coordinate regression.2.A pose robust face keypoint localization algorithm is proposed.The depth-separable convolutional network is used to quickly obtain the corresponding 3DMM parameters and three-dimensional face key points from face images.It covers various kinds of face pose and gets satisfactory result.Moreover,a new parameter initialization strategy is proposed for the problem of error accumulation in cascade regression.3.A face-driven animation system is realized,which consists of face detection,3DMM parameters regression and Unity3D synthetic system.Experiments show that the system proposed can accurately capture facial expressions and generate vivid animation videos.The animation driving system based on the fast face detection algorithm and the pose robust 3D face key point localization algorithm implemented in this paper has a frame MRF objective function is adopted for view selection where new forms of data term and rate of about 20?30fps,which can satisfy the requirements of real-time driving.