外文翻譯---基于半邊臉的人臉檢測（英文）

1、The Ninth International Conference on Electronic Measurement Step2 determine the size of face area and select face area; Step3 normalize selected face areas into the same size; Step4 compute the average value of ev

2、ery corresponding pixel of face area. Before constructing template, some images including observe face are chosen. At first, the size of face area is determined. Then, the face area selected manually in these fa

3、ce image. The number of the selected face area is n. The matrix vectors of face areas are distributed independently, so are the pixel values in same position of these face areas. The pixel value in the kth (k=1,2,?,n

4、) position of face area is fk(i,j) (k=1,2,…,n) and the normalized scale factor of these face image is wk (k=1,2,…,n), thus, the average face template can be described as follows. ? ? ? ? ? ? ?nk k k j i f w n j i T1 ,

5、 1 ,(1) According to statistics theory, if some pixel value fk(i,j) in the kth face area subjects to normal distribution N(?, ?2), where, ? is the mean of the pixel fk(i,j) and ? is the variance, T(i,j) subjects to N(

6、?, ?2/n) distribution. So the chanciness of local density in face template decreases in a large scale. If face image is sampled in the situation that the distance between the face and the camera is fixed, the size o

7、f face is unchanged correspondingly and the normalized scale factor wk can be 1. Then the average face template T(i,j) can be altered as follows. ? ? ? ? ? ? ?nk k j i f n j i T1 , 1 ,(2) 4-54__________________________

8、___978-1-4244-3864-8/09/$25.00 ©2009 IEEE The Ninth International Conference on Electronic Measurement & Instruments ICEMI’ 2009 (a)

9、 (b) Fig. 6 detected possible location of half face V. CRITERION FUNCTION In the experiment, the half faces in the images were detected with the template matching method. Its fundamental princip

10、le can be described as follows. The selected average half face-template is ransacked on the detecting image and the similarity between them is calculated. The half face image which is similar to the template is recko

11、ned among the detecting image, if the value of the similarity function in some position is greater than the threshold value given. The similarity is the statistical value of local areas in the image. The similarity v

12、alue of some different sub-image image might be equal to another, despite all that they are different sub-images. In the experiment, the similarity function value in the half face-template matched position should be

13、stood out and the value in the non-matched position should be bated. The method adopted is described as follows. Suppose the length of the half face-template T is I and the width is J, as is shown in figure 4. The le

14、ngth of the full face-template is 2I and the width is J. Of the detecting image, the length is L and the width is W. The sub-image corresponding to the position (m,n) in the image where the template is set on is Pm,n

15、 .Then the similarity between the template and the sub-image S(m,n) can be expressed with the formula (3) [9]. ? ? ? ? ? ? ? ? ?? ? ? ? ?IiJjn m j i T j i P n m S1 12 , , , ,(3) For the function above, the rule of dete

16、rmining whether there exists the half face is that given a threshold value th, if S(m,n)?th, the half face-template T is similar to the sub-image Pm,n, and if S(m,n)=0, the half face-template T is identical with the

17、sub-image Pm,n completely. Expanding the formula (3), it will be the following: ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?? ?? ?? ? ? ? ? ? ? ? ? ? ?IiJjIiJjn m IiJjn m j i T j i T j i P j i P n m S1 121 1,1 12 , , , , 2 , ,(4) Whe

18、re, ? ? ? ? ?? ? ?IiJjn m j i P1 12 , , is the energy of sub-image Pm,n which is covered by half face template T located in (i, j) of the images. During the ransacking of the images, its value is changing slowly. ? ?

19、? ? ?? ? ? ?IiJjn m j i T j i P1 1, , ,is the correlation between the template T and sub-image Pm,n, it gets the maximum value when template T match the sub-image Pm,n exactly. ? ? ? ? ?? ? ?IiJj j i T1 12 ,is the ener

20、gy of half face template T, its value can be determined when the template T is constructed. It has no relationship with the position of the sub-image Pm,n. Therefore, the ratio of the correlation of the template T and

21、 sub-image Pm,n to the energy of sub-image Pm,n can act ad the similarity, that is, ? ?? ? ? ?? ? ? ? ????? ?? ? ?? IiJjn mIiJjn mj i Pj i T j i Pn m s1 12 ,1 1,,, ,,(5) Normalizing formula (5), the following one could

22、 be deduced. ? ?? ? ? ?? ? ? ? ? ? ? ? ?? ????? ? ? ?? ? ?? IiJjIiJjn mIiJjn mj i T l i Pj i T j i Pn m s1 121 12 ,1 1,, ,, ,,(6) Where, s(m,n) is similarity, 0?s(m,n) ?1. As to similarity s(m,n), the rule for determin

23、ing whether the half face exists can be described as: if threshold is given as th, when s(m,n)?th, the conclusion that the half face template T is similar to the sub-image Pm,n is drawn, when s(m,n)=1, the half face

24、template T match the sub-image Pm,n exactly. Assuming that the O(T) represents time cost of face detection based on half face template and the O(F) represents that based on full face one, the calculating method of t

25、he O(T) and O(F) can be described as formulas (7) and (8). O(T) = I * J * ( L - I ) * ( W - J ) (7) O(F) = 2I * J * ( L - 2I ) * ( W - J ) (8) The ratio of O(T) to O(F) can

26、 be shown in formula (9) I LI LJ W I L J IJ W I L J IF OT O4 2 ) ( * ) 2 ( * * 2) ( * ) ( * *) () (?? ? ? ?? ? ?(9) When the value of L is much greater than that of I, the result of (9) approximates to1/2, that is to

27、say, the time cost of detecting faces in images with the half face template is about half of that with the full face one. Therefore, about half time of the processing is saved. VI. EXPERIMENTS RESULTS As to average h

28、alf face template shown in figure 5, the similarity of “ left face and left face” , “ right face and right face”and “ left face and right face”were calculated with the formula (6). The results are 1.0000?1.0000 and 0