Expression Invariant 3D Face Recognition

As a hot topic in pattern recognition field,face recognition technology has achieved greate development,which has been gradually infiltrated into people's lives.Due to limited information,it is difficult to further enhance recognition rate in two-dimensional human face which affected in pose,light and other conditions.3D face recognition has caused a lot concern.It can overcome the problem of illumination and pose using shape information.However,the expression changes usually cause that inter-class distance is larger than outer-class distance,the expression is still a hot issue to be resolved in 3D face recognition.In view of the above problems,this paper studies the expression invariant 3D face recognition technology.In steps of preprocessing,feature extraction,and matching,design 3D face recognition algorithm.To 3D face model with facial deformation,extract a robust face description,and implement effective matching.The main contents of this paper are as follows:.Firstly,3D face database preprocessing.Standardize the 3D face database by smoothing and pose correcting with automatic correspondence method.After the above operation,the standardization work of 3D face is finished.Secondly,this thesis proposes a simple feature model.Construct the geodesic ring in the average face,select rigid points according to prior knowledge and determine the topological relationships among points.So all 3D facial feature models can be determined automatically.Then we introduces a robust feature description for 3D extraction.Utilize Gabor wavelet to extract rich point neighborhood feature which is stable to facial deformation.At the same time,using fast marching to compute the geodesic distance to describe relationships among points.So we obtain the feature model which is similar to the elastic model.Lastly,finish the recognition in elastic matching and convolutional neural networks.In one hand,measure the similarity in respective points and edges,compute the weight by experiment,to implement recognition based on elastic matching.On the other hand,take the above features as the input of the convolutional neural networks,by train we can obtain an appropriate network.Experiments are verified that proposed method can obtain the higher accuracy than other method and robust to facial deformation.At the same time,the whole method is simple and does not need preprocessing work with high quality.