The Research Of Multi-view Face Recognition Based On The Three Dimensional Face Modeling

Being a typical pattern recognition task, face recognition has a huge marketprospect and high practical value. Nowadays, It gives a promising effect when face toideal environment. However, when face to poor environment (pose change,exaggerated facial expressions, uneven illumination), the effect of most of the systemwill decrease. On the other hand, most of the present methods cannot recognizehuman’s face in real time. It also limits the application of face recognition.Under the patronage of the college students’ scientific andtechnological promotion project in Zhejiang province(2013R407063), we start fromthe background and significance of face recognition, considering the threedimensional structure of faces, then analyses methods of multi-view face recognitionbased on the three dimensional face modeling. In the paper, we have already mainlyachieved the following research work:(1) In the stage of face detection: in order to ensure the stability and real-timeperformance of the system, we choose face detection Algorithm based on AdaBoostand package the classifier into a XML file. These lay a foundation to build a facerecognition system with real-time performance and it can link to the face recognitionsystem seamlessly.(2) In the respect of three dimensional face modeling: After comparing somemethods of three dimensional face modeling based on face images, we choose themethod based on one front face image. Compared with conventional methods, theFeasibility and Practicality of this method is high. We purpose a method of threedimensional face modeling based on Candide-3model. The method iscomposed of two steps: first, it locates the landmarks on the front face using improvedASM algorithm and then locates the non-feature points by Nearestneighbor interpolation; second, in the stage of texture mapping, we create a modelwith more features Compared with Candide-3based on DMS spline function.Moreover, the reality and Continuity of our model is better. Finally, we show the faceimages with multi-view created by our three dimensional face model. Compared withclassic methods, our method can extract the deep features of face images, and itproves to be a beneficial attempt. (3) In the stage of feature extraction of face images: we introduce a new facerecognition method based deep neural network. Our method has two steps:unsupervised stage and supervised stage. In unsupervised stage, we train the deepneural network by recovering the face images in LFW using Sparse Autoencoder. Wetrain some kinds of networks with different layers to explore the influence on facerecognition of the numbers of layers; In supervised stage, we improve the ability ofunderstanding the faces of the network using NCA algorithm and the Hybrid facedataset with various face images(multi-poses, many expressions and A variety oflight). And this reduces the distance of same person and increase the distance ofdifferent persons.(4) During the actual operation of face recognition: we propose a new methodnamed "Search Radius". At the beginning, we represent each image as abinary number. In the retrieval process, the hamming distances between these binarynumbers of images were calculated, then we order the distances from smallestto largest. We choose the smallest one to figure out whether they belong to the sameperson based on EER. The results show that our method spends very little time onrecognition and it is suitable for real-time face recognition tasks.(5) At last, we integrated algorithms and apply some APIs, such as OpenCV,OpenGL and Armadillo, to build a real-time face recognition system. The result showsour system has good recognition efficiency and it spends little time on recognition.These make it a suitable system for face recognition.