Multi-pose Face Recognition Based On The Single View

Face recognition is a technology which analyzes the image and then extracts theinformation of effective recognition from it by computer. Face recognition has significantlycommercial value in the field of public security, digital identity and multimedia, so it causesextensive attention. Over the past decades, great progress and developments have beenmade in the field of face recognition，but these technologies still can not fully meet theevolutuinary practical application. At present, face recognition systems perform very wellunder the condition of control and cooperative, While, the rate of face recognition willdegrade sharply without control and cooperative, such as face recongnition under thecondition of multi-pose.Now, researchers have put forward the following four kinds of solutions of multi-poseface recognition: multi-perspective face recognition technology, face recognition technologybased invariant features, face recognition technology based3D model and multi-pose facegeneration technology based single view. The first three kinds of technology have to haveenough face data before being applied. Researchers use single view to generate multi-poseface to extend training smaples, when training samples are single view. The accuracy ofgeneration of these technologies is not high.To improve the accuracy of generating multi-pose faces in multi-pose face recognition,this paper proposed a multi-pose face generation algorithm based on local weighted meanmethod. Then this paper use principal component analysis to extract face feature vectors anduse support vector machine to recognize the multi-pose faces. Specific research as follow:(1) This paper found the mapping function set of local feature points between front faceand multi-pose face by applying local weighted mean method, and then designed thetransformation function of each pixel for generating the multi-pose face by using weightedmean of mapping functions of neighboring feature points. The multi-pose faces aregenerated by these transformation functions, and these multi-faces constitute tranningsamples library. The experimental results show that the peak signal-to-noise ratio betweenmulti-pose faces genertated by this algorithm and photos is high, so the accuracy ofgeneration of multi-pose faces is inproved.(2) This paper chose principal component analysis to extract face feature vectors, at last,utilize support vector machine to recognize the multi-pose faces. Our approach overcomesthe difficulty of obtaining multi-pose face images in multi-pose face recognition and solves the problem of the rapid fall of recognition rate due to the face pose changing. Experimentalresults have shown that the multi-pose faces produced by local weighted mean algorithmpreserve the local features of face and have a high similarity with original faces in libraryORL, consequently effectively improve multi-pose face recognition rate.