3D Face Recognition Based On MeshSIFT Features And Convolution Neural Network

Subject to the limitations of data acquisition hardware,in the past few decades,face recognition is mainly studied in the two-dimensional image,which provides illumination,posture and occlusion constraint conditions.Two-dimensional image recognition has achieved good recognition results,and some technology have been commercialized.In a complex environment,the illumination condition of 2D face recognition will seriously affect the recognition results.Because of its complete preservation of the face space information,the 3D face recognition reflects the true shape of the face,and the real world on the biological simulation in the eyes of the person,not sensitive to illumination,pose,occlusion and other factors,it is an effective supplement to the research of 2D face recognition.This paper analyze the current situation of face recognition,put forward to use the convolution neural network model and collect low-resolution 3D database to verify the applicability and robustness of the algorithm on low-cost devices.The main tasks are as follows:Firstly based on the characteristics of the local descriptor algorithm meshSIFT,a matching method based on angle weighting is proposed to improve the recognition rate.Compare with the method base on quantities,the method base on angle weighting can solve the problem of misjudgment even in same matching quantities.The recognition rate was improved 3%in the Bosphorus database Rank1.Then,the concept of similar matrix map is put forward,and the 3D face recognition is transformed into binary classification problems.Taking the advantage of binary classification,the normalized similarity matrix map is trained and identified by using the convolution neural network model.The recognition rate of the algorithm is 99.46%in the Bosphorus database Rank1.In order to verify the applicability and robustness of the algorithm on low-cost and low-resolution 3D devices,Kinect2SZU database was collected by Microsoft Kinect 2^nd-generation devise to test for 3D face recognition algorithm.Comparing with the open data KinectFaceDB set collected by Kinect 1^st-generation devise,the accuracy of each algorithm is slightly improved.But if comparing with the Bosphorus database which was collected by high-precision scanning equipment,the accuracy of each traditional handmade feature algorithm is greatly reduced,less than 45%.The experiment shows that the resolution of 3D data has a great influence on the recognition accuracy.Compared with other traditional handmade feature methods,the convolution neural network algorithm proposed by us has much better robustness.This method has been obtained 70.19%and 80.16%accuracy rate on Kinect FaceDB and Kinect2SZU database respectively.