Face Recognition Based On Image And Depth Information Fusion

With the rapid development of artificial intelligence technology,face recognition plays an important role in access control,intelligent transportation,video monitoring and other fields.Face recognition is still very challenging due to different methods of spoofing attacks,illumination changes and the requirements of the actual scene system.This thesis uses the method of convolutional neural network and geometric features to detect the face authenticity.Research and use face information fusion in different modes to obtain features that can better represent face,so that face recognition system can better detect face authenticity and improve face recognition accuracy.The research content of this topic has a positive significance.Aiming at face liveness detection,this thesis proposes a two-stage liveness judgment method based on face depth map and color map information.Based on the results of face key point detection,this thesis studies the liveness detection based on face depth map,and constructs a spatial plane model to determine whether the attack comes from the plane mode.Based on face color image,a lightweight convolutional neural network two classification model is constructed to judge the face authenticity of color image.In this thesis,the advantages of the first-order model for the plane attack with good results and the two-stage neural network model for the judgment of the color map are combined,so that the results of face detection can accurately judge the authenticity of the face through two stages of detection,and try to further optimize the algorithm,so that the second-order activity detection algorithm can give correct results for the attack modes in different scenarios.Aiming at the face feature extraction algorithm,this thesis proposes a three-dimensional face feature extraction model based on the two-dimensional face feature Lightcnn algorithm.The face recognition algorithm needs to have better robustness under different illumination changes.The 2D face recognition has achieved good recognition accuracy,but the robustness of the algorithm model needs to be improved.This thesis first proposes a network model algorithm that combines facial features extracted under different modalities.The new features are characterized by the fusion of features,which improves the accuracy of face recognition and reduces the storage capacity of the database.Secondly,this thesis combines the relationship between different modes,and trains the new face feature extraction model by adding the correlation constraints between different modes in the loss function,which makes the face recognition system have good result in the public data set and the actual scene.The face recognition system is relatively less deployed on the embedded side because of the large amount of computation.This thesis combines the proposed liveness detection and 3D face recognition method to build a 3D face recognition system based on liveness detection on the NVIDIA TX2 embedded board,and the GPU module of the board card is used to accelerate and optimize the model.In this project,the depth camera is used to collect the actual data,and the data sets of face recognition and different face attack modes in the actual application scenarios are constructed.The accuracy of the proposed face activity detection algorithm in the data set can reach more than 95%,the accuracy of the fused face recognition algorithm can reach 99.82%,and the whole system takes about 86ms.Finally,this thesis summarizes the research work carried out,and looks forward to the future research and direction.