Face Recognition Under Unconstraint Environment Based On Metric Learning

Face recognition plays a significant role in promoting the development of image processing,pattern recognition,computer vision,computer graphics and other fields.With the developing needs of application areas in video surveillance,information security,access control,especially in internet information retrieval,Multi-pose Face Recognition under unconstrained environment has become one of the most hot research direction in the field of face recognition.The keys of the research focus on how to effectively measure the pixel information of face image,overcome difficulties on low face resolution,large range of scale change,illumination,dramatically pose change and occlusion.In addition,real-time face recognition has received more and more attention,increasingly higher compute speed is needed on how to find the target face in the massive human face database.To solve these problems,specific works based on metric learning,face preprocessing,facial recognition retrieval system are as described below.1.A new multi-supervised metric learning algorithm frame is proposed in this paper.This method is started by the alignment of the face image,which includes 2D face alignment and 3D face alignment.Then dense SIFT features are extracted,a Gauss mixture model is produced,and the Fisher Vector features are obtained.Finally the multi-supervised metric learning is used to reduce the dimension of face image,trying to reduce the distance between the two classes,and increase the distance within the class as far as possible.Experiments showed that this method proposed in this paper outperformed the state-of-the-art face verification performance on the challenging "LFW" benchmark on the same condition of 2D-alignment.2.For the face preprocessing,we considered a new field in this paper,which is face blood erasing.Firstly,according to the characteristics of human blood,decision tree is used to detect the location of the blood.Secondly,the face image restoration algorithm is used to recover the face.When the face region still contains blood after the restoration,the above process will be repeated until the blood is erased effectively.Also each facial organ is masked to prevent being modified by mistakes.Moreover,experiments showed that the blood erasing algorithm has positive effect on face detection and face recognition.3.For the design and implementation of face retrieval application system,we mainly consider the retrieval speed.The randomized k-d tree and priority k-means tree was used to build database index,which greatly improved the speed of real-time face retrieval,while the face recognition rate is almost unaffected.