Research Of Robust 3D Facial Landmarking Techniques

As the basis of technology application on faces,3D face landmarking is widely used in face recognition,facial expression recognition,face matching,face alignment,facial shape analysis and so on.However,facial surface is intricate.In order to overcome the effects of pose changes and occlusion,more robust landmarking methods are eager to be proposed.Based on the previous work,a nose tip localization method is proposed by analyzing characteristics of nose tip,and a face multi-landmarking method is proposed by the application of deep models.The main work and contributions are as follows:1)A fast algorithm for nose tip localization is proposed,which is robust to pose variations.The accuracy of nose tip localization affects the performance of face preprocessing such as face cropping to a great extent.However,once the face has pose variations,nose tip cannot locate on the nearest place at positive direction,which seriously causes many traditional localization methods to be invalid.To solve this problem,this paper proposes a two-step algorithm: First,a sparse neighborhood set is constructed for each vertex.The distance between the domain point near the vertex and the plane of the local reference frame is extracted under local reference coordinates.After that,a new rotation invariance feature named local reference frame energy is obtained.According to this energy,it is possible to iteratively screen out candidate points that may be nose tips.Second,the divergence of each candidate points is calculated.The divergence can describe the degree of expansion and contraction of the vector field near the vertex.Because the degree of expansion at the nose tip is most in the surface of the face,so the vertex with maximum divergence value is identified as the nose tip.The algorithm in this chapter does not require any training process and classifier,and is very robust to attitude changes.2)A facial landmarking method based on denoising auto-encoder networks is proposed.There are often obstructions on the surface of the face,resulting in serious deviations in the localization of facial landmarks.The occluded face can be regard as a damage to the data.As a deep model,the denoising auto-encoder can extract features of the undamaged data from the corrupted data.A parallel denoising auto-encoder network is proposed based on the denoising auto-encoder.The network consists of four denoising auto-encoder trained by sub-regions.For each denoising auto-encoder,the localization accuracy of feature points is different on regions.The feature points are roughly predicted from the global feature.After that,all results predicted from four denoising auto-encoder are fused as coarse estimation.The fusion parameters can be obtained according to the degree of occlusion of the face.Then,the local feature is extracted from the coarse estimation,and the shape residual is predicted by the linear regression layer,so that the coarse estimation result is closer to the real position.In order to verify the effectiveness of the two proposed algorithms,several experiments were designed on the face database FRGCv2.0 and Bosphorus respectively.Compared with some current algorithms,proposed methods get the state-of-the-art performance.The experimental results demonstrate the robustness of the two algorithms.