Facial Feature Research Based On Curve Representation

The development of society and great progress of technology provide human alot of benefits, but meanwhile, they also bring us many challenges. How to confirmthe identification of social members effectively? This is one of those problems whichwe have to face. In the last two decades, with the high speed development of com-puter and computer imaging technology, a large number of researchers focus on therange of face recognition technique. For this range, many methods try to find a way toextract discriminative information from facial images.Human can recognize and verify face images and other visual target precisely,even when we only have the edge information of original pictures. This phenomenonindicates that, edge pictures are rich of information which could be use for recognition.Based on this theory, one method called edge map was proposed, which use the edgeinformation of facial image to do recognition job. After that, a new approach namedLine Edge Map or LEM was proposed. By using a simple polygonal line fitting proc-ess, they divided facial edge map into a group of line segments. In this way, the LEMapproach not only has the advantages of edge map method, but also has the conven-ience of low data storage space and high efficiency of information coding. But insome circumstances, using line segment can’t describe some parts of facial edge mapwhich have the feature of curve segment, and can only destroy the integrity of infor-mation. For this reason, Parabola Edge Map or PEM approach which based on a fastcurve fitting method was proposed. PEM approach use parabola segments to extractfeatures from facial edge picture. They also use pixel value from original facial pic-ture as part of features. Inspired by these methods mentioned above, we propose several new algorithmsbased on curve representation:1. Facial image edge lines show a more complex feature in some specific part ofpictures. If only using linear or parabolic segment to divide the edge lines, it is easy todestroy the continuity of original curve characteristics and weaken the discriminativepower of features. Therefore, we propose a new facial feature representation methodbased on high order curve fitting. By using constructive fitting, we can easily extracthigh order curve feature from some certain parts of image. After we get the curvesegment sets, we locate some small regions beside these segments in original facialpictures and call them relevant region of curve. In the process of relevant regionanalysis, we first divided them into several sub blocks. Then we use the polar momentof inertia to represent invariant features inside these sub blocks. In this way, we offera better method to use the gray level distribution information of original image whichhas been reflected in the edge image. The experimental results based on FERET facialdatabase show that the performance of this new algorithm is better than the existingmethods mentioned above.2. The effect of expression change to facial recognition algorithm is obviously.In some cases, differences between image features of same person under different ex-pression condition may appear even more than differences between image features ofdifferent person under same expression condition. One solution for this problem is togather pictures of expression change of same person as many as possible. But in fact,with the limitation of objective conditions, sample face images collected by peoplecannot cover all possible type of expression. In order to more comprehensively andeffectively reflect the expression change of human face and reduce the complex ef-fects to recognition job which bring by it, we can analysis facial expression imagedata and find out the pattern behind it. By using the pattern and the image we alreadyhave, we can then have new synthetic expression image. We adopt the method ofcurve based face representation and turn the edge image into curve segments form. Inthe pairing of curves under different expression conditions, with the correspondingphysical model analysis, we can find out the internal factors caused the variation of curves and carry on the quantification analysis for it. At the same time, we measurethe change of the gray distribution on both sides of curves in original pictures. Aftercalculate and analysis a certain number of sample image of a particular expressionchange, we can get the internal rules of it. So, even only the neutral expression faceimage of someone can be obtained, by using the pattern of certain expression varia-tion and the curve representation of neutral face image which we already have, we canstill synthesize new face image with expression change in curve representation form.This synthesis image can be used to replace the real image with the same kind of ex-pression change. We use FERET and AR facial databases to test our method. In theexperiments, we selected some pictures which people have angry or scream expres-sions. After calculation and analysis, we use pictures of other people with neutral ex-pressions to synthesize new image with certain expression change. The experimentresults demonstrate that the synthesis image can effectively extended the range ofsample database, and improve the capability of recognition algorithm to adapt the ex-pression changes of test samples.In summary, this paper makes some research on facial feature extraction methodbased on curve representation, provides a new, simple and effective feature represen-tation method for face retrieval and recognition. At the same time, for the difficultybrought by facial expression changes, we proposed the corresponding solution, ex-pand research ideas in this field, and have achieved good results in the experiment.