human face detection using geometric triangle relationship

7/27/2019 Human Face Detection Using Geometric Triangle Relationship

1/4


2/4

to a binary image. If this input image is not a grayscaleimage, transform the input image to grayscale format,and then convert this grayscale image to binary image bythresholding. Secondly, label all 4connectedcomponents in the image to form blocks and find out thecenter of each block. Thirdly, detect m y 3 centers of 3different blocks that form an isosceles triangle Fourthly,clip the blocks that satisfy the triangle criteria as thepotential face region.2 1 Preprocess the inputted image to a binaryimage

In this step, we first convert the input image to abinary image. For example, when we read in a RGBcolor image, we will preprocess the input image to graylevel image by eliminating the hue and saturationinformation while retaining the luminance. Then,binarize the gray level image to a binary image bysimple global thresholding with thresho ld T because theobjects of interest in our case are darker than thebackground. Before proceeding to the next step, weperform the opening operation (erosion first, thendilation) to remove noise, and then the closing operation(dilation first, then erosion) to eliminate holes. The detailof opening and closing operations can be found in [131.2 2 Label all 4xonnected components and findthe center of each block

Here, we use raster scanning (left-to-right and top-to-bottom) to get 4-connected com ponents, label them, andthen find the center of each block. The detail of rasterscanning can be found in [131.2 3 Find any 3 centers of 3 different blocks thatform an isosceles triangle

From careful observation, we discover that two eyesand one mouth in the frontal view will form an isoscelestriangle. This is the rationale on which the finding ofpotential face regions is based. We could search thepotential face regions that are gotten from the criteria ofthe combination of two eyes and o ne mouth .2.3.1.The matching rules for finding an isoscelestriangle

If the triangle i j k is an isosceles triangle as shown inFigure 2(a), then it should possess the characteristic ofthe distance of line i j = the distance of line j k. Fromobservation, we discover that the Euclidean distancebetween two eyes (line i k) is about 90% to 110 of theEuclidean distance between the center of the rightlleft

I

Figure 2. a)The isosceles triangle ij k . b) Three points i , j ,and k ) form an isosceles triangle.eye and the mouth. Due to the imaging effect andimperfect binarization result, a 25 deviation is given toabsore the tolerance. The first matching rule can therebybe s tated as (abs (D (i , j )-D i ,k))< 0.25*max (D(i,j), D(i,k)) , and the second matching rule is abs(D(i,j)-D( i, k))< 0.25*max(D(i, j ,DO, k)) 1Since the labeling processis operated from left to right then from top to bottom, wecan get the third matching rule as i < j < k. Here, absmeans the absolute value, D i, j) denotes theEuclidean distance between the centers of block i (righteye) and block j (mouth), D i,k) denotes the Euclideandistance between the center of block k (left eye) andblock j (mouth), D (i,k) represents the Euclideandistance between the centers of block i (right eye) andblock k (left eye). For example, as shown in Figure 2(b),if three points ij, nd k) satisfy the matching rules, thenwe think that they form an isosceles triangle.After we have found the isosceles triangle, it is easy toget the coordinates of the four com er points that form thepotential facial region. Since we think the real facialregion should cover the eyebrows, two eyes, mouth andsome area below mouth [9], the coordinates can becalcu lated as follows. Assume that (Xi, Yi), X j , Yj) and(Xk, Yk) are the three center points of blocks i, j , and kthat form an isosceles triangle. (Xl, Yl), (X2, Y2), (X3,Y3), and (X4, Y4) are the four corner points of the faceregion as shown in Figure 3. X1 and X4 locate at thesame coordinate of (= - 1/3* D (i, k)); X2 and X 3 locateat the same coordinate of( Xk + 1/3* D (i, k)); Y1 an d Y2locate at the same coordinate of (Yi + 1/3*D (i, k)); Y3and Y4 locate at the same coordina te of (Yj - 1/3*D i,k)); where D i, k) is the Euclidean distance betw een thecente rs of block i (right eye) and block k (left eye).

Figure 3. Assume that Xi, Yi), Xj,YJ) and Xk, Yk) are thethree center points of blocks i, j , and k respectively. Thefour corner points of the face region will be X l, Yl) , X2,Y2), X3, Y3), and X4, Y4).X1 = X4 = Xi - 1/3* D(i, k);Y3 = Y4 = Yj - 1/3* D(i, k);

(1)

(4)X2 = X3 = k + 1/3*D(i, k);Y1 = Y2 = Yi + 1/3* D (i, k); (2)(3)

9 4 2


3/4

3 Face verificationThe second part of the designed system is to performthe task of face verification. In the previous section, wehave selected a set of potential face regions in an image.

In this section, we propose an efficient weighting maskfunction that is applied to decide whether a potential faceregion contains a face. There are three steps in this part.The first step is to normalize the size of all potentialfacial regions. The second step is to feed everynormalized potential facial region into the weightingmask function and calculate the weight. The third step isto perform the verification task by thresholding theweight obtained in the precious step.3 1 Normalization of potential facial regions

Normalization of a potential face region can reducethe effects of variation in the distance and location . Sinceall potential faces will be normalized to a standard size(e.g. 60 * 60 pixels) in this step, the potential faceregions that we have selected in the previous section areallowed to have different sizes. Here we resize thepotential facial region using bicubic interpolationtechnique.3 2 Weighting mask function and weightcalculation

If the normalized potential facial region is reallycontains a face, it should have high similarity to the maskthat is formed by 10 binary training faces. The idea issimilar to neural network, but does not need so manytraining samples. Every normalized potential facialregion is fed into the weighting mask function that isused to compute the similarity between the normalizedpotential facial region and the mask. T he computed valuecan be utilized in deciding whether a potential regioncontains a face or not.The method for generating a mask is to read in 10binary training masks that are cut manually from thefacial regions of images, then add the correspondingentries in the 10 training masks to form an added m ask.Next, binarize the added mask by thresholding ea ch entry.For example, we have 10 masks and the size of eachmask is 3 3. The first mask is formed by 9 zero(zero represents white pixel), so the first mask is awhite 33 block. The 10th mask is produced by 9 one(one represents black pixel), so the 10th mask is ablack 3 3 block. We added these 10 masks together andget an added mask with the value of each entry a s [7 8 94 5 6 3 2 13. If we select the threshold value 5 (if thevalue of each entry is larger than 4, then we assign itsvalue as 1; otherwise we assign its value as O. , we getthe final mask that has the values of [l 1 1 0 1 1 0 0 01.The illustrating procedure is depicted as shown below:

For all pixels of the potential facial region and themask: If both of the potential facial region and the maskcontain those pixels at the sam e location of the parts oftwo eyes, the nose, and the mouth (both of them areblackpixels.), then weight = weight + 6. If both of thepotential facial region and the mask contain those pixelsat the sam e location of the parts of the skin of the face(both of them are w hite pixels.), then weight = weight +2. If the pixel of potential facial region is black and thepixel of the mask is white, then weight = weight 4. Ifthe pixel of potential facial region is white and the pixelof the mask is black, then weight = weight -2. Theexperimental results show that the values of + 6 , + 2 , - 4,and 2 as given above can obtain the best match of thefacial region. In other words, the range of thresholdvalues is the narrowest.3 3 Verification

After we have calculated the weight of each potentialfacial region, then a threshold v alue is given for decisionmaking. Once a face region has been confirmed, the laststep is to eliminate those regions that overlap with thechosen face reg ion, and then ex hibit the result.Verification of the fron tal view - A set of experimentalresults demonstrates that the threshold values of thefrontal view should be set between 4,000 and 5,500.4 Experimental results and discussion

In this section, a set of experimental results isdemonstrated to verify the effectiveness and efficiency ofthe proposed system. There are 500 test images (include450 different persons) containing totally 600 faces whichare used to verify the validity of our system. Some testimages are taken from digital camera, some from scanner,and some from videotape. The sizes of the test imagesrange from 10*10to 640*480 pixels (actually ou r systemcould handle any size of image.) In these test images,human faces were presented in various environments.The execution time required to locate the preciselocations of the faces in the test image set s dependentupon the size, resolution, and complexity of images. Forexample, the gray level image which is 200*307 pixelsas shown in Figure 4 a) need less than 2 5 seconds tolocate the correct face position using PI1 233 PC. Thegray level image which is 200*145 pixels as shown inFigure 4(b) need about 28 seconds to locate the correctface position because the background is more complex

943


4/4

than Figure 4(a). Figure 4: Experimental results of graylevel/color images w ith simple/complex backgrounds.

Figure : Experimental results of gray leveVcolor imageswith simplelcomplex backgroun ds.

5 ConclusionsIn this paper, a rob ust and effective face identif icationsystem is presented to ext ract face in var ious kinds of

face images. Moreover, i t can manage dist inct size,changed l ighting condit ion, varying pose and expression,and the noise and defocus problem. In addit ion, thepresented sys tem can a l so cope wi th the problem ofpartial occlusion of mouth and wearing sunglasses. Theexperimental results reveal that the proposed method isbetter than tradit ional methods in terms of al teredcircumstance. efficiency, an d accuracy.6 References[ I ] M. Turk and A Pentland, Eigenfaces for recognition,Joumal of Cognitive Neuroscience, vol. 3, no. I , pp. 71-36,1991.[2] David J. Beymer, Fac e recognition under varying pose,Technical Report, MIT AI Lab, AI memo 1461, December1993.[3] D. Valentin, H. Abdi, A. Otoole, and G. Cottrell,Connectionist models of face processing: A survey , PattemRecog nition, vol. 27, no. 9, pp. 1209-1230, 1994.[4] A. Pentland, B. Moghaddam, and T Stamer, View-basedand modular eigenspaces for face recognition, in IEEEConference on Computer Vision and Pattern Recognition,Seattle, Washington, Jun. 1994, pp. 84-9 1.[ 5 ] K K. Sung and T. Poggio, Example-based learning forview-based human face detection, in Proc. ImageUnderstanding Workshop, Monterey, Calif., Nov. 1994, pp.[6] T. K. Leung, M. C. Burl, and P. Perona, Finding faces inclustered scenes using random labeled graph matching, inProc. Computer Vision and Pattem Recognition,Cam bridge, Mass., Jun. 199 5, pp. 637-644.[7] H. A. Rowley, S. Baluja, and T. Kanade, Hum an facedetection in visual scenes, Tech. Rep. CMU-CS-95-I58R,Cameg ie Mellon University, 1995. (The paper is gotten fromhttp://www.cmu.edu/-harlfaces.htm1[SI S. Y. Lee, Y. K. Ham, and R. H Park, Recognition ofhuman front faces using knowledge-based feature extractionand neuro-fuzzy algorithm, Pattem R ecogn ition, vol. 29, no.[9] P. Juell and R. Marsh, A hierarchical neural network forhuman face detection,Pattern Recogn ition, vol. 29, no. 5 , pp. 781-787, 1996.[ I O ] K. Sobottka and I Pitas, Extraction of facial regions andfeatures using color and shape information , in Proc. 13thIntemational Conference on Pattem Recognition, Vienna,Austria, Aug. 1 996, pp. 421-425.[1I ] H. Wu, Q. Chen, and M. Y achida, A fuzzy-theory-basedface detector, in Proc. 13th Intemational Conference onPattem Recognition, Vienna, Austria, Aug. 1996,[121 C. C. Han, H. Y. Mark Liao, G. J. Yu, and L. H. Chen,Fast face detection via morphology-based pre-processing , nProc. 9th Intemational Conference on Image[ 131 Rafael C. Gonzalez and Richard E. Woods, Digital ImageProcessing , copy ngh t 992 by Addison-Wesley PublishingCompany, Inc.

843-850.

11, pp. 1863-1876, 1996.

944
http://www.cmu.edu/-harlfaces.htm1http://www.cmu.edu/-harlfaces.htm1

human face detection using geometric triangle relationship

Documents