Face Detection for Automatic Avatar Creation by using Deformable Template and GA

Transcription

1 Face Detection or Automatic Avatar Creation by using Deormable Template and GA Tae-Young Park*, Ja-Yong Lee **, and Hoon Kang *** * School o lectrical and lectronics ngineering, Chung-Ang University, Seoul, Korea (Tel : ; -mail: tzeropak@sirius.cie.cau.ac.kr ** School o lectrical and lectronics ngineering, Chung-Ang University, Seoul, Korea (Tel : ; -mail: jalnans@sirius.cie.cau.ac.kr *** School o lectrical and lectronics ngineering, Chung-Ang University, Seoul, Korea (Tel : ; -mail: hkang@cau.ac.kr Abstract: In this paper, we propose a method to detect contours o a ace, eyes, and a mouth o a person in the color image in order to make an avatar automatically. First, we use the HSI color model to exclude the eect o various light conditions, and ind skin regions in the input image by using the skin color deined on HS-plane. And then, we use deormable templates and genetic algorithm (GA to detect contours o a ace, eyes, and a mouth. Deormable templates consist o B-spline curves and control point vectors. Those represent various shapes o a ace, eyes and a mouth. GA is a very useul search algorithm based on the principals o natural selection and genetics. Second, the avatar is automatically created by using GA-detected contours and Fuzzy C-Means clustering (FCM. FCM is used to reduce the number o ace colors. In result, we could create avatars which look like handmade caricatures representing user's identity. Our approach diers rom those generated by existing methods. Keywords: Face detection, Deormable Template, GA, B-spline curve, HIS color model, FCM, avatar. INTRODUCTION There are many methods or creating avatars in internet. Most o them are done by the method based on image templates. But, it is true that we cannot obtain the user s eature o appearance well with the template-based method. Also, it requires so many options or users to choose, and as a result, it makes users tired. This paper proposes an eective and interactive method that automatically recognizes user s eatures o appearance rom a 2D Image o the user ace, and creates an avatar by using these eatures. Thereore, the created avatars have the user appearance, interact with the user, so as to express the user s subjective sensitivity. To detect a eature rom an input image, the search or eatures in all regions is required. Here, reduction o search target regions as well as eicient usage o the computation time and resources are important issues. We use the HSI color model to reduce distortion o the tone o color caused by the light conditions. It deines the color o skin in the HS plane, and then applies it to detecting a ace region. In search o ace contours, it is diicult to search or such contours by using only the edge values in a complex background. So it is required that the search method may obtain an object contour even i serious variations present. We can increase accuracy by using a deormable template consisting o B-spline snakes in the eature extraction to which we apply the genetic algorithm or eiciency improvement. Basically, our method o creating avatars consists o two stages automatically and interactively. The irst stage is detecting acial eature points rom the user s input picture. In the second one, we create avatars by reerring to these points and ine-tune them in order to be enhanced during the interaction with the users. This work was supported by grant No (Development o the Key Technology or Intelligent and Interactive Module rom the project o Developing SIC(Super Intelligent Chip and its Applications under the program o Next generation technologies in 2000 o [Ministry o Commerce, Industry and nergy] The irst step in detecting user s acial eature points rom the picture is done by separating the background rom the ace area. The color inormation is used during the processing o the image. Actually, human s skin color has a uniorm color range. However, this method is not perect due to the intensity changes and the shading. Thereore, we perorm several preprocesses or the image. Once the acial region has deined, detection o acial eature points is perormed on the gray scale image because the color values o the image can be distorted by many causes such as illumination and input device s color characteristics. The detected edges have much inormation on the image. We rather perorm the modiied edge detection than the classical methods because the traditional edge detection methods have weakness under noises. By using these eature points, our application draws an original avatar and its variations to be shown or the user. And then, it interacts with the user to add the user s sensitivity to the generated avatars. In evolutionary computations, the important thing is how to deine a itness unction. It is proper that the user evaluate the itness unction subjectively. 2. FAC RGION DTCTION The input image o the RGB color model is converted to the HSI color model. It is robust against variance o light changes. 2. RGB model VS HSI model Generally, a stored image is represented with RGB color model. One o the problems in image recognition is that each component contains light's intensity, so the intensity values change a lot. There are requent occasions in such ways not only that light is distributed irregularly, but also that target o body is shaded by the light. In this case, i we use HSI color model, we may have some independence rom shadow or light eects. Intensity is used or measuring the degree o light's powerulness in the HSI model. It is regarded as one independent component, and thereore we can design robust

2 3 Rotation 4Y scale 5Y translation ICCAS2005 algorithm against variance o light i we neglect intensity. The picture below represents this concept. In case o the ace contour, chin line control points are added in order to enhance the accuracy. To extract the ace contour there 8 control point based on B-spline. There are two control point sets or chin line in them. Aine-transormed chin line control points have an inluence on a ace contour as shown in Fig Deormation parameter Fig. HSI color model 2.2 Skin color deinition in H-S plane Detecting acial area is launched by deining the region suitable or skin candidate area. However, it is almost impossible that skin area can be deined as one region in Hue and Saturation due to changing lights and environments. So, in this paper, we determined a standard ace color in H-S plane. We extract skin color rom various images in dierent conditions to deine a standard skin color cluster. The center point in H-S plane is deined as a standard skin color and the skin color is deined inside the ellipsoid o the color cluster. Using the skin color range, we could detect the acial skin. ach template is deormed independently. In the eature extraction, the processing step goes through ace, eyes and lips. Deinition o each parameter is as below. In case o the ace contour, there are 5 parameters; X or Y axis translation, X or Y axis scale, and rotation. In the eyes, there are 2 templates at let and right. But we change only one side because o symmetric shapes. Thereore a parameter is added or the distance between two eyes. So there are 7 parameters or the eyes; X or Y axis scale and rotation X or Y translation, rotations o the eye center points and the distance between two eyes. In the lips, there are 5 parameters; X or Y translation, X or Y scale and rotation. Fig. 5 shows these parameters. 5 Rotation 4 Y scale 2 Center Point Fig. 2 Detection o acial skin region 2.3 Deinition o deormable templates In this paper, object eatures are ace, eyes, and lip. There are basic template control points based on B-spline snake. Chin line 4 Rotation 3 Y translation 5Y scale Distance between eyes 3 X scale (a Face X translation 2 X scale 2 X translation 6 X scale (a Face (b yes (c Lips Fig. 3 Control points o the deormable template (b yes (c Lip Fig. 5 Deormable template parameters 3. FAC DTCTION USING GA Genetic Algorithm (GA is one o the most powerul methods to solve an optimization problem. In this paper, we use GA to search the best solution or an input image. 3. Chromosomes The chromosomes consist o parameters deined by the deormable template. Table ~3 shows these chromosomes. Fig. 4 Deormable o chin line with the templates

3 Table Chromosome o ace contour. Chin line - ~ + * 2Byte 0~65535 Center point ~ Number o skin pixel 2Byte 0~65535 X axis scale 0.5 ~.5 2Byte 0~65535 Rotation -0 ~+0 2Byte 0~65535 Table 2 Chromosome o eye contours. Distance between eyes - ~ + ** 2Byte 0~65535 X axis translation - ~ + ** 2Byte 0~65535 Y axis translation - ~ + ** 2Byte 0~65535 Rotation rom center point between -0 ~+0 2Byte 0~65535 eyes X axis scale 0.5 ~.5 2Byte 0~65535 Rotation rom center o eyes -0 ~+0 2Byte 0~65535 Table 3 Chromosome o lip contour. X axis translation - ~ + ** 2Byte 0~65535 Y axis translation - ~ + ** 2Byte 0~65535 X axis scale 0.5 ~.5 2Byte 0~65535 Rotation -0 ~+0 2Byte 0~65535 * Dependent on input image ** Dependent on skin region 3.2 Fitness unction We use a tone o color and edge value or searching. We use 3 itness unctions or ace contour detection. First, we estimate a itness by using the edge values. 57 ( i 0i i ( ( Second, we estimate a itness using H, S values at the inner place o the template. 57 ( H H + ( S S (57 D 2 2 i scolor i scolor 2 HS max (2 Third, we estimate a itness using H value at the outer place o the template. 40 H5i H3i (3 Finally, total itness o ace contour is calculated by weights. ω + ω + ω ω + ω + ω i : Fitness unction o ace contour 0i : dge value, i i : dge value at side control value H, S : H, S value i i H, S : Standard ace color D H scolor scolor (4 HS max : Distance between standard ace color and control point 5i, H 3i: H value (a (b 2 (c 3 Fig. 6 Fitness measurement o ace contour The itness measurement o the eye contour consists o 3 itness unctions. First, we use the edge values to ind an eye. 20 ri + li e (5 Second, the gray values relected inner place o the eyes. N linepixel DrGi + DlGi e2 NrLinePixel + NlLinePixel (6 Third, we estimate the eye points o lower part to remove an eyebrow. ( Gri + Gli e3 2 (7 Finally, a itness o the eye contour is calculated by the weights. ω e + ω2 e2 + ω3 e e ω+ ω2 + ω3 (8 ei : Fitness unction o eyes N, N : Control point o side llinepixel rlinepixel G i : Gray value o sample point Fig. 7 Sample point or itness measurement

4 In case o the lips, there are 3 itness unctions. First, we use an edge value to measure a itness. mi m ( 255 (9 Second, the tones o color near the mouth dierent rom the skin color. m2 ( H H + ( S S ( D 2 2 i scolor i scolor HS max Third, the lip is more red than the skin color. Riin Riout m Finally, we get a total itness o the lip contour. ω m + ω2 m2 + ω3 m m ω + ω + ω 2 3 (0 ( (2 Fig. 2 represents the simpliied image according to the proposed method. Fig. 0 Result image 5. SIMULATION Fig. is low chart o proposed method. START mi : Fitness unction o lip R R mi : dge value iin : Red value at inner point iout : Red value at outer point Input image Face region detection -HSI color model A color tone -Standard ace color Feature extraction using GA -xtract ace contour -xtract eye contour -xtract lip contour Shape ino. -Deormable Template Original ace contour Original eye contour Original lip contour Fig. 8 Sample point or itness measurement Create avatar FCM 4. AVATAR CRATION 4. Facial color simpliication using FCM It needs to simpliy the color levels o the detected acial image or an avatar. Only the skin region must be extracted in the acial region determined by the preceding procedure. It can be clustered with FCM on the basis o pixel intensity at this region. The pixels which belong to each cluster are simpliied to the predeined color levels based on the degree o intensity. ND Fig. Flow chart o proposed method Fig. 2 is showing a result image made by proposed method. Fig. 9 Simpliied image by FCM 4.2 Creation o ace image We make an image like a cartoon by using simpliied image rom FCM. In case o eyes, we use the contours obtained by the above procedure to make an eye image. With the similar approach, we get a simpliied lip image. Fig. 2 Simulation result In ig.2, the itness and the ace contour are shown. The result image shows that the ace contour is reined well comparatively and the itness graph shows that the total itness value gets higher as the generation evolves. I the itness value meets the condition, the searching process is terminated. In case o the eyes and the lips, we know that the accuracy is limited because the templates are rigidly ixed in shape.

5 6. CONCULUTION In the past, image recognition system used gray image and edge value but it is hard to ind a good result in the uncertain environmental conditions, especially, light changes. Thereore we are concerned about hue and saturation to ind the acial eature sets. The deormable template is connected with the B-spline snakes. Then we ind the acial eatures using genetic algorithm. As we use the deormable template to relect various shapes o aces, the proposed genetic algorithm reduces the searching time. The genetic algorithm is the most powerul method to ind the optimized eature in a global search region. As a result, we ind eicient methods o extraction in spite o the noise distortion or the lack o inormation. We hope to improve the proposed method and then apply it to many other uncertain or diicult situations. RFRNCS [] Andrew Blake, Michalel Isard, Active Contours, Springer, 2000 [2] Raael C. Gonzalez, Richard. Woods, Digital Image Processing, second edition, Prentice Hall, 2002 [3] William H. Press, Saul A. Teukolsky, William T. Vetterling, Brian P. Flannery, Numerical Recipes in C++, CAMBRIDG UNIVRSITY PRSS, 2002 [4] James D. Foley, Andries van Dam, Steven K. Feiner, John F. Hughes, Computer Graphics Principles and Practice, Addison-Wesley, 997 [5] B.R.Moon, Genetic Algorithm, Dasung, seoul, 200 [6] David. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley Publishing Company, Inc., 989 [7] Hoon Kang, Fuzzy System, Daeyoung, seoul, 2000 [8] J.S.R Jang, C.T. Sun,. Mizutani, Neuro-Fuzzy & Sot Computing, Prentice Hall, Inc, 997 [9] Saeed B. Niku, Introduction to Robotics Analysis, Systems, Applications, Prentice Hall, 200 [0] Perez.P., Gidas.B. Motion detection and tracking using deormable templates Image Processing, 994. Proceedings. ICIP-94., I international Conerence Volume 2, 2-6 Nov. 994 pp vol.2 [] Silsbee.P.L Motion in deormable templates, Image Processing, 994. Proceedings. ICIP-94., I international Conerence Vol., pp , 994