The Algorithm Research Of Frontal And Upright Human Face Detection

Recently face recognition has already been a researchful hotspot of pattern recognition and artificial intelligence because of its enormous application foreground. However, face detection is the precondition and linchpin of face recognition. Face detection affects the precision and speed of latter face recognition and it is an important subject to turn face recognition technique into application. Because of face pattern's anfractuosity and destructibility, algorithms of face detection have a lot of disadvantages such as plentiful calculation,extraordinarily slow operation speed and very high error detection ratio under common circs. So that, face detection is very challenging.In this paper, three human face detection methods are proposed in allusion to the status quo and difficulties of face detection research.The first algorithm, of which subjects investigated are dynamic video frequency grey images shooted by our lab and grey images of CMU single face database, combines coarse segmentation of difference image with multi-template matching to detect faces. The algorithm reduces the range where faces are to be searched, therefore its detection speed is three times of that of other detection methods used singly by multi-template matching.The second algorithm is proposed because the method of difference image is not effective when the different information coming from two frames of difference image is too much or too little. Dynamic video frequency color images is the subject investigated of this algotithm.Firstly, by making the best of tinct information, the face region, of which complexion is concentrative and steady, is segmented through color space conversion. In this way, the range where faces are to be searched is reduced. Secondly, multi-template matching,mosaic image validation and geometric rule affirmance are combined to detect faces. Compared with the only one method, the false detection ratio of this multiple method is reduced further more.The third algorithm is presented in allusion that the methods of skin color segmentation and the method of difference image are of no effect in case that the similar face regions are too many and the different information coming from two frames of difference image is too much or too little. The improved BP algorithm in which both momentum coefficient and learning rate can beadjusted is used in this algorithm. Therefore, the rapidity of convergence is increased ten times and face detection ratio is improved a lot.The three algorithms proposed are easily realized. Compared with other traditional algorithms, detection speed is much improved and false detection ratio is much reduced in different degree. Test result shows that the three algorithms are effective and feasible.