The Research And Application Of Multi-view Facial Landmark Location

Perception and analysis of face image is a key front subject of pattern recognition and artificial intelligence includes face detection, facial features localization, facial pose estimation, expression analysis, face synthesis, face recognition and face animation and so on. Among them, face detection and facial features localization are the foundations and face recognition is one of applications of the most importance. Compared with the frontal face detection, multi-view face detection is more complicated but researched relatively weakly, they are the key problems and needed to resolve urgently for the facial perception and analysis technology becoming more practical.It is just in the background. This paper presents a Multi-view Tree Structured Model, namely M-TSM, and applies it into large databases` searching process to solve problems of face recognition under big data environments. This model will solve simultaneously face detection and landmark localization by establishing and maintaining facial templates under different viewpoints. By combining these two issues, we simplify the process of face recognition. At the same time, it is used to complete the registration of large face databases after its distributed process. This paper proposes adding two dimensions for each feature vector to store the viewpoint infromation captured by M-TSM. By taking this advantage of M-TSM in image analysis, the searching process will skip directly over inconsistent viewpoints registrated in system narrowing our search scope to accelerate the recognition speed. The main work and achievements are as follows:(1) We model each feature points from training samples to obtain all the local parts and archive parts shared under all viewpoints. Building multi-view model from local parts for different viewpoints and design its energy function which contains HOG[49] feature information of all the parts, first and second connection coefficients between parts.(2) Selecting samples marked with face region, feature points and viewpoints information. First of all, we obtain the maximum likelihood estimation from a given sample based on the principle of Maximum Spanning Tree which gives a crude estimation of the distribution of the given feature points; then we apply LSVM [23] on M-TSM to discriminatly train its parameters extracting identify characteristics for face detection; finally, analyzing the searching process of M-TSM, propose enumerating all viewpoints and searching facial landmarks through multiple dynamic programming principle in the kown face region;(3) M-TSM`s test and experiments compared with other algorithms will be finished here to evaluate our model respectively for face detection, pose estimation and landmard localization. Experimental results show that our model not only get a better performance for multi-view face analysis, but can also finish pictures of more than one person at the same time on face analysis.(4) Construting a lightweight distributed file system on Windows systems by Distributed File System [15] theory and Map Reduce[16] framework. Take the advantages of M-TSM in image analysis- output of facial pose and distribution of feature points- to register face images automaticly for large database. Follow-up experiments will demonstrate a significant role in building M-TSM distributed database and narrowing down the search scope.