Face Recognition Of Specific Person In Open World Environment

As a natural, direct and easily-accepted biometric technology, face recognition hasbeen widely applied in the field of identification and intelligent monitoring. In real worldapplications, the majority of face recognition tasks are accomplished in an open worldenvironment, which means to identify a fraction of people specified by the system in alarge scale of probe faces. In this open environment, only a small portion of the faces arepreviously registered in the system and the system needs to effectively reject a largenumber of non-target persons’ faces. In addition, affected by illuminations, postures, facialexpressions and other factors, the same person has different face appearances. Thesefactors make face recognition task in an open world environment is quite difficult. In thispaper, we proposed an effective approach to conduct face recognition task under openworld environment. The main work of this paper can be described as follow:(1) Capture the mechanism of human recognizing patterns to build abinary-face-classifier for each target person to be recognized. We take borderline faces oftarget person as training data to build the person-specific-recognizer, which can identifywhether the input face belongs to the corresponding target person. Wherein the borderlinefaces are generated by the face morphing procedure between target person and a largeamount of non-target persons.(2) Apply the complementary feature descriptors of Local Binary Pattern and GaborWavelets to extractor training data of feature vectors from those borderline faces and useSupport Vector Machine to build each feature-based classifier. Then take advantage of thecomplementary attribute to build a parallel recognition network by combining therecognition result of the two feature-based classifiers. This parallel network can obtain agood recognition rate not only among target-faces, but also among non-target-faces.(3) Generate multi-view faces of borderline faces and improve the recognition rate of faces with different angles. For this purpose, we propose a multi-view face generationmethod based on face morphing mechinism, so that the system can generate a smooth facewith a specified angle by using faces of fixed angles.In this paper,10celebrities are set to be target person and we get a lot of their faceimages from the Internet. Along with the standard database FERET and CMU-PIE webuild a face data set similar to open environment. We conduct experiment on this data setand compare the proposed approach with the Robust Sparse Coding method. The resultsshow that in an open world environment, the recognition performance of the proposedapproach is better than RSC’s, with a recognition rate of more than92%in target person’sfaces, and a false alarm rate of less than0.5%. In addition, we adopt the68persons inCMU-PIE dataset as target figures to verify that the system have a good performanceenhancement to multi-view face’s recognition after the joining of multi-view borderlinefaces.