1. Proceedings of the 2008 IEEE International Conference on Robotics and Biomimetics Bangkok, Thailand, February 21 - 26, 2009 Prefessor : 謝銘原 Student : 謝琮閔 ID : M9820113 PPT 100 ％原創

2. Abstract (1/4) In human face detection applications, face region usually form an inconsequential part of images. Preliminary segmentation of images into regions that contain "non-face" objects and regions that may contain "face" candidates.

3. Abstract (2/4) Color information based methods take a great attention, because colors have obviously character and robust visual cue for detection. This paper proposed a new method based on RGB color centroids segmentation (CCS) for face detection.

4. Abstract (3/4) Include two parts, first part is color image thresholding based on CCS. Second part is face detection based on region growing and facial features structure character combined method.

5. Abstract (4/4) The experimental results show the ideal thresholding result and better than the result of other color space analysis based thresholding methods. Proposed method can conquer the influence of different background conditions, position, scale instance and orientation in images from several photo collections and database; the effect is also better than existing skin color segmentation based methods.

6. INTRODUCTION Nowadays, many application technologies are developed to secure access control based on biometrics recognition such as fingerprints, iris pattern and face recognition. Many applications such as financial transactions, monitoring system, credit card verification, ATM access, personal PC access, video surveillance etc.

7. Recent surveys on face detection (1/3) Principal component analysis (PCA), Neural networks (NN), Support vector machines (SVM), Hough transform (HT), Geometrical template matching (GTM), Color analysis etc.

8. Recent surveys on face detection (2/3) PCA need create Eigen face by many dimension data and training sample data. The NN require a large number "face" and "non-face" images to train respectively for getting the network model. SVM are a linear classifier and can classify goal region in hyper-plane. HT and GTM were incorporated to detect gray faces in real time applications.

9. Recent surveys on face detection (3/3) A combination of holistic and feature-based approaches is a promising approach to face detection as well as face recognition. This paper proposed a color image thresholding algorithm based on color centroids segmentation (CCS) for detection and tracking is able to handle a wide range of variations in color images sequence, various backgrounds can detect face region effetely.

10. COLOR IMAGE THRESHOLDING BASED ON CCS This section introduce how to thresholding the color image by transform RGB components of RGB 3-D color space to 2-D polar coordinate system. Use multi-threshold to segment the centroids region. By analyzing and processing, it can cluster the color of image to 2~7 colors by 2~7 thresholds for require and the effect better than traditional methods.

11. A. Color Triangle (1/3) To create the color triangle, a standard 2-D Cartesian coordinate system is used to describe R, G and B values and then transform it to polar coordinate system as (1) show:

12. A. Color Triangle (2/3) By the following steps can create the color triangle: Step 1: create a standard 2-D polar coordinate system; Step 2: create three color vectors to reflect R, G and B colors; every vector’s value range is 0, 255 and alternation 120° reciprocally. Step 3: connect the three apexes.

13. A. Color Triangle (3/3) After above processes, the color triangle can be created as Fig. 1. For different R, G and B values, the shape of triangle is changeable. No matter the R, G and B value change the main structure are fixed.

14. B. Color Centroids Hexagon Region Distributing (1/3) R, G, B vectors direction is fixed and the value is change from 0 to 255, so different combination of R, G, B value will create different color, and the shape of color triangle is changed too. The different shape triangle has different centroid, and the centroids distributing region of color triangle is show hexagon as Fig. 2.

15. B. Color Centroids Hexagon Region Distributing (2/3) In this hexagon region, it divided to 7 regions: R (Red), G (Green), B (Blue), C (Cyan), M (Magenta), Y (Yellow) and L (Luminance, achromatic) regions. In Fig. 2 we use seven threshold curves as the dividing line for thresholding.

16. B. Color Centroids Hexagon Region Distributing (3/3) The R, G and B values are closely, no matter small or large it only reflect the luminance information (weak color information). The centroids of corresponding color triangles will in a circular region (L region). And other six color regions reflect the color character of R, G and B combination.

17. C. Color Centroids Segmentation Thresholds Acquisition (1/7) Considering the L region usually is not the goal region and existing method cannot effective to divide white and black region usually. This region is noise region, so clustering the value of this kind to one region wills effective to ignore the influence of white, black and other achromatic region. Here let as the threshold of L region, as the angle, the function of threshold curve is:

18. C. Color Centroids Segmentation Thresholds Acquisition (2/7) The other six regions which around the L region as follows formulas show:

19. C. Color Centroids Segmentation Thresholds Acquisition (3/7) In formula (2) and (3), and are the thresholds, and the initial value of them is 5, 60°, 120°, 180°, 240°, 300° and 360°. Considering the advantages and disadvantages, here proposed an automatic thresholds selection method to get the thresholding for different scene.

20. C. Color Centroids Segmentation Thresholds Acquisition (4/7) By analyzing largely face region distributing, we can see that the color of face usually included in R region and lean to Y region. To display distributing character more clearly, we transform the Polar coordinate system to Cartesian coordinate system as Fig. 3(d) to reflect the distributing of centroids.

21. C. Color Centroids Segmentation Thresholds Acquisition (5/7) In the Fig. 3(d) horizontal axis is vertical axis is and other six vertical color-line are color threshold curves The face region is belonging to Red and Yellow region

22. C. Color Centroids Segmentation Thresholds Acquisition (6/7) By observing many face included image in different condition, we let the threshold curve can move to left or right 20° for find best value. Then find the left and right valley bottom respectively as in the fix rang by histogram analysis method.

23. C. Color Centroids Segmentation Thresholds Acquisition (7/7) In Fig. 3, (a) is original image and (b) showing the distributing of color centroids. By transforming Fig. 3(b) to (d) and calculate the we can get the pre-face region, and the binary image show in Fig. 3(c):

25. FACE DETECTION AND TRACKING A. Thresholding 1. Thresholding Based on CCS The binary image can be got as Fig. 4(b). From the result we can see that the white background region (wall), pale color clothing region ①, ④ and ⑤ and dark color clothing region ⑦ are clustering to black and only the goal region clustering to white.

26. 2. Correction Using Nonlinear Thresholding For denoise the incorrect region, this paper adopt the nonlinear thresholding method to correct the binary image which thresholdinged by CCS.

27. The hair region ② and ⑧, so use (5) processed image to correct the CCS based method processed the image with and operation can get idea result as follows:

28. Fig. 4(d) is the corrected binary image, it correct the region ②, ⑥ and ⑧ effetely.

29. B. Pre-face Region Decision After get the idea binary image, the white region is the wait-decision region. Here all wait-decision regions are analysed in a selection process and some of them accepted by aspect ratio and size.

30. Accepted by aspect ratio: here C is aspect ratio,L is the length of boundary. S is the area of wait-decision region. If, it will be accepted.

31. Accepted by size: After accepted by aspect ratio, then calculate the average area of all wait-decision regions without the largest and smallest regions. If it will be accepted.

32. C. Face Region Tracking After face region fixed, use a circle to draw it by follows: Step1: divide the face region to 9 blocks and thresholding respectively as Fig. 6. Step2: wipe off noise region by median filter. Step3: fix eyes and mouth region and then calculate the area centroids of eyes and mouth respectively.

33. Step4: draw a circumcircle (blue circle of Fig. 6) of triangle which created by the three centroids. Then use its 1.5 times’ concentric circle (green circle of Fig. 6) to mark face.

34. EXPERIMENTAL RESULTS ANALYZE AND COMPARISON A. The Result of Proposed Method 1. Thresholding result and compare with other methods:

36. 2. Face detection result Fig. 8(a) is a white background image with multi-face; usually the white clothing region will influence the thresholding result which method based on luminance or histogram analysis and the pink clothing region will influence the thresholding result which method based on color analysis.

37. Fig. 9 shows a multi-person image under indoor situation and everyone ware different color clothing. From Fig. 9(b) we can see that proposed method can conquer the influence of the light, ground region and different color of clothing effectively. Fig. 10 is one frame of video under outdoor situation. In this image, some region of background can influence thresholding result, because the color is near face color.

39. B. Compare with reference[22] Fig. 11(a) is [22] used image, Passing analysis the method based on skin information of [22] and experiment result, it only adopt to propose simple background and little color influence images. [22] Y. Wu, and X. Ai, "Face Detection in Color Images Using AdaBoost Algorithm Based on Skin Color Information," 2008 Int’l Workshop on Knowledge Discovery and Data Mining, pp. 33-342, Jan. 2008.

40. C. Compare with reference [23] Fig. 12 shows the sample image of reference [23], and Fig. 12(b) and Fig. 12(d)~(f) is from experiment result of [23]. [23] L. Sabeti, and Q. M. J. Wu, "High-speed Skin Color Segmentation for Real-time Human Tracking," 2007 IEEE Int’l Conf. on Systems, Man and Cybernetics, pp. 2378-2382, Oct. 2007

41. CONCLUSIONS All the experiment results show that the proposed method can get ideal detection and tracking result under complex background, multi-face and color influence. The future works are how to make the CCS method more quickly and have better thredholded effect for detecting and tracking.