CN102592141A - Method for shielding face in dynamic image - Google Patents

Abstract

The invention relates to the field of computer vision and image processing, and discloses a method for shielding a face in a dynamic image. The method comprises the following steps of: firstly, constructing a skin color division region and dividing the skin colors to obtain multiple probable image positions and sizes of a face region; secondly, carrying out template matching through region scanning and verifying the face region; and at last, carrying out fuzzy processing on the face region so as to realize an effect of mosaic. The method is quick in operation speed, is hardly influenced by changes of gesture, size, expression and the like, and is suitable for the application field having higher requirement about instantaneity.

Description

A kind of method that people's face in the dynamic image is blocked

Technical field

The present invention relates to computer vision and image processing field, the technology of especially people's face in the dynamic image being blocked.

Background technology

In some special application, need the people's face in the dynamic image be blocked (or adding mosaic).As in order to protect interviewer's in the news privacy; Need interviewer's face image be blocked processing (or Fuzzy Processing); Usually way is the zone of people's face set positions in image, and the image in this zone is asked for average according to certain block size, reaches fuzzy effect.Yet, because the interviewer is movable, so when carrying out the news post-production; Need dynamically adjust fuzzy region; If adopt manual operations, then producer's workload is bigger on the one hand, also has technical risk on the other hand; Promptly, window not exclusively exposed interviewer's privacy, the perhaps excessive news picture quality that influences of fuzzy region owing to covering.

In order automatically to accomplish the people's face in the dynamic image is blocked; At first need utilize automatic human face detection tech; Promptly adopt certain strategy that image is searched for, confirming wherein position, the size of people's face, and then this parts of images is carried out Fuzzy Processing.

Detect difficult point from moving face and show two broad aspect, on the one hand since the inherent variation of people's face cause: people's face has the quite variations in detail of complicacy, different appearance such as the shape of face, the colour of skin etc., the opening and close etc. of different expressions such as eye, mouth; Blocking of people's face is like glasses, hair and head jewelry and other exterior objects etc.On the other hand owing to the external condition variation causes: because the difference of imaging angle causes the colourful attitude of people's face, rotate and rotate up and down like rotation in the plane, the degree of depth, wherein degree of depth rotation influence is bigger; The influence of illumination is like the variation of the brightness in the image, contrast and shade etc.; The image-forming condition of image, like focal length, the image-forming range of picture pick-up device, approach that image obtains or the like.

Summary of the invention

The object of the invention is to provide a kind of automatic realization people's face in the dynamic image to be blocked the method for (or adding mosaic).

1. area of skin color is cut apart

The colour of skin is the important information of people's face, does not rely on facial minutia, can both be suitable for for situation such as rotation, expression shape change, has relative stable and distinguish mutually with the color of most of background objects.For coloured image, after having confirmed complexion model, can at first carry out Face Detection, after detecting skin pixel, be partitioned into possible human face region in similarity on the colourity and the correlativity on the space according to them.

For some simple situation, only can accomplish Region Segmentation according to the cluster characteristic of skin pixel.For than complicated situation, skin color segmentation then need consider two sides and problem: because the influence of illumination and face's organ, people's face possibly isolated is some mutual disconnected area of skin color; Human face region possibly link together with other type area of skin color.Cluster-merger-authentication policy is this type problem solution commonly used: at first skin pixel is the zone according to comparatively strict colourity consistance and geometrical constraint condition cluster; Carry out merger according to certain rule then, utilize further feature to verify again after the merger or in the merger process.

The advantage of area of skin color dividing method is that the travelling speed of system is fast; The influence that changed by attitude, size, expression etc. is little; Be suitable for application that real-time is had relatively high expectations, but responsive to the characteristic of illumination condition and image capture device, be subject to environmental factor and disturb.Therefore, the coarse positioning that is fit to human face region.

2. verify based on people's face of template matches

After area of skin color splits, need to utilize the geometric properties or the gray feature in zone, whether checking should the zone be people's face, to get rid of the zone of the similar colour of skin of color.

Based on the method for template matches is in direct ground similarity between comparison object template and the candidate image area on the gradation of image characteristic; And based on the matching process of geometric properties the similarity of certain characteristic of relatively from image, extracting; People's face is detected, and characteristic comprises eyes, nose, face etc.The method of template matches has characteristics simply and intuitively, compares with the geometric properties matching process to have stronger adaptability.

Generating the template most principal work is the change of scale and the intensity profile standardization of image; Main mean value and the variance of considering two parameter-gray scales of token image intensity profile; They are adjusted to set-point, and factor such as illumination is to the influence of facial image when eliminate gathering.For template matches, this intensity profile standardized method is more effective than histogram equalization method commonly used, these two key figure characteristics of reason has been its unification average and variance.

Choose some secondary positive facial images of rectifying; Shear out human face region as people's face sample; After scale calibrationization (36 * 36) and intensity profile standardization, it is average that everyone face sample is got gray scale, obtains size and be the original average face template that rectify in 36 * 36 front.

Consider the importance of eyes in face characteristic, copy the eyes part of original average face template, shear out size and be 36 * 12 eye areas, it is carried out after the intensity profile standardization as the eyes template.

Respectively according to 1: 0.9, the length and width of 1: 1 and 1: 1.1 is than stretching with original face template, respectively they carried out after the intensity profile standardization as face template, to adapt to difform people's face.

Use each template to might yardstick and the image of shape (length breadth ratio) mate, the image that satisfies certain condition and reach the matching degree threshold value as people's face, is used the basis for estimation of related coefficient as template matches.

3. people's face occlusion method

(1) creates the skin color segmentation zone

Having obtained several through skin color segmentation possibly be picture position, the size of human face region, and then each maybe human face region be exactly that of template matches is by the region of search.

(2) carry out template matches through sector scanning

Scanned by the region of search for each; After intensity profile standardization pre-service; At first use the eyes template to mate, to the region of search of related coefficient above given threshold value, the face template that then re-uses various yardsticks matees; The degree of correlation is maximum and greater than the scanning area mark behaviour face position of people's face threshold value, and preserve position, the size that this moment should the zone.

(3) human face region is carried out Fuzzy Processing

To human face region is image known location, the size of template matches, and the image in this zone is asked for average according to certain block size, reaches Fuzzy Processing or adds the effect of mosaic.

Description of drawings

Fig. 1 is the inventive method process flow diagram.

Fig. 2 is cut apart process flow diagram for area of skin color.

Fig. 3 is an original color image.

Fig. 4 is the image after the skin color segmentation.

Fig. 5 is possible human face region.

Fig. 6 is the human face region of checking.

Embodiment

Embodiment 1:

The inventive method flow process is as shown in Figure 1, and it is as shown in Figure 2 that area of skin color is cut apart flow process, comprises the steps:

(1) read in original color image, as shown in Figure 3;

(2) scan image obtains each pixel R of this coloured image, G, B component value, calculates its gray-scale value according to formula (1); And the number of every kind of gray-scale value in the calculating entire image; Get preceding 5% pixel, try to achieve the light compensation coefficient, and with R, G, the B value of this each pixel of coefficient adjustment;

Gray＝R×0.3+G×0.59+B×0.11 (1)

(3) each pixel is transformed under the YCbCr color space according to formula (2), calculates its Y, Cb, Cr value;

$\left[\begin{array}{c}Y\\ \mathrm{Cb}\\ \mathrm{Cr}\\ 1\end{array}\right]=\left[\begin{array}{cccc}0.299& 0.587& 0.114& 0\\ -0.1687& -0.3313& 0.500& 128\\ 0.500& -0.4187& -0.0813& 128\\ 0& 0& 0& 1\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\\ 1\end{array}\right]---\left(2\right)$

(4) if Y between [125,188], then Cb, Cr remain unchanged, otherwise adjust its Cb, Cr value according to formula (3)～(6);

${\stackrel{\&OverBar;}{C}}_r\left(Y\right)=\left\{\begin{array}{c}154-\frac{(K_i-Y)(154-144)}{K_i-Y_{\min }},Y<K_i\\ 154-\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(3\right)$

${\stackrel{\&OverBar;}{C}}_b\left(Y\right)=\left\{\begin{array}{c}108+\frac{(K_i-Y)(118-108)}{K_l-Y_{\min }},Y<K_i\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(4\right)$

${\mathrm{Wc}}_i\left(Y\right)=\left\{\begin{array}{c}{\mathrm{WLc}}_i+\frac{(Y-Y_{\min })\left({\mathrm{Wc}}_i\text{-}{\mathrm{WLc}}_i\right)}{K_l-Y_{\min }},Y<K_l\\ {\mathrm{WHc}}_i+\frac{(Y_{\max }-Y)({\mathrm{Wc}}_i-{\mathrm{WHc}}_i)}{Y_{\max }{K}_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(5\right)$

(5) if Cb between [77,127] and Cr between [133,173], then this pixel is made as white, otherwise is made as black;

(6) each pixel in the image is carried out the processing of step (3), (4), (5), obtain the image after the skin color segmentation, as shown in Figure 4, obtain possible human face region, as shown in Figure 5.

Embodiment 2:

If the gray-scale value matrix of image is D [W] [H]; Wherein W, H are respectively the width and the height of image, and then the average gray of this image

is expressed as

$\stackrel{\&OverBar;}{\mathrm{\&mu;}}=\frac{1}{W\&CenterDot;H}\underset{i=0}{\overset{W-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{H-1}{\mathrm{\&Sigma;}}}D\left[i\right]\left[j\right]---\left(7\right)$

The variance of intensity profile

is expressed as

${\stackrel{\&OverBar;}{\mathrm{\&sigma;}}}^2\frac{1}{W\&CenterDot;H}\underset{i=1}{\overset{W-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{H-1}{\text{\&Sigma;}}}{\left(D\right[i\left]\right[j]-\stackrel{\&OverBar;}{\mathrm{\&mu;}})}^2---\left(8\right)$

The intensity profile standardization is exactly that the average gray of image and gray variance are transformed to standard value μ ₀And σ ₀, each gray values of pixel points in the image is carried out conversion by formula (9), then obtain the image after the intensity profile standardization.

$\hat{D}\left[i\right]\left[j\right]=\frac{{\mathrm{\&sigma;}}_0}{\stackrel{\&OverBar;}{\mathrm{\&sigma;}}}\left(D\right[i\left]\right[j]-\stackrel{\&OverBar;}{\text{\&mu;}})+{\mathrm{\&mu;}}_0---\left(9\right)$

Wherein, 0≤i＜W, 0≤j＜H.

Embodiment 3:

The gray matrix of supposing face template is T [M] [N], and gray average and variance are respectively μ _TAnd σ _T, the gray matrix of image to be verified is R [M] [N], gray average and variance are respectively μ _RAnd σ _R, then (T R) does the correlation coefficient r between them

$r(T,R)=\frac{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\left(T\right[i\left]\right[j]-{\mathrm{\&mu;}}_T)\left(R\right[i\left]\right[j]-{\mathrm{\&mu;}}_R)}{M\&CenterDot;N\&CenterDot;{\mathrm{\&sigma;}}_T\&CenterDot;{\mathrm{\&sigma;}}_R}---\left(10\right)$

Wherein, ${\mathrm{\&mu;}}_T=\frac{1}{M\&CenterDot;N}\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}T\left[i\right]\left[j\right],$ ${{\mathrm{\&sigma;}}_T}^2=\frac{1}{M\&CenterDot;N}\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{\left(T\right[i\left]\right[j]-{\mathrm{\&mu;}}_T)}^2,$ ${\mathrm{\&mu;}}_R=\frac{1}{M\&CenterDot;N}\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}R\left[i\right]\left[j\right],$ ${{\mathrm{\&sigma;}}_R}^2=\frac{1}{M\&CenterDot;N}\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{\left(R\right[i\left]\right[j]-{\mathrm{\&mu;}}_R)}^2,$ (T R) has reflected the angle of image vector between input picture zone and template to r, and (T, R) matching degree in big more representation template and input picture zone is high more for r.

Claims (4)

1. the method that people's face in the dynamic image is blocked is characterized in that comprising the steps:

(1) obtain picture position, the size that several possibly be human face regions through skin color segmentation, then each maybe human face region be exactly that of template matches is by the region of search;

(2) scanned by the region of search for each; After intensity profile standardization pre-service; Use the eyes template to mate, to the region of search of related coefficient above given threshold value, the face template that then re-uses various yardsticks matees; The degree of correlation is maximum and greater than the scanning area mark behaviour face position of people's face threshold value, and preserve position, the size that this moment should the zone;

(3) be image known location, the size of template matches to human face region, the image in this zone is asked for average according to certain block size, reach Fuzzy Processing or add the effect of mosaic.

2. the method that people's face in the dynamic image is blocked according to claim 1 is characterized in that skin color segmentation comprises the steps:

(1) reads in original color image;

(2) scan image obtains each pixel R of this coloured image, G, B component value, calculates its gray-scale value; And the number of every kind of gray-scale value in the calculating entire image; Get preceding 5% pixel, try to achieve the light compensation coefficient, and with R, G, the B value of this each pixel of coefficient adjustment;

(3) each pixel transitions is arrived under the YCbCr color space, calculate its Y, Cb, Cr value;

(4) if Y between [125,188], then Cb, Cr remain unchanged, otherwise adjust its Cb, Cr value according to formula (3)～(6);

${\stackrel{\&OverBar;}{C}}_r\left(Y\right)=\left\{\begin{array}{c}154-\frac{(K_i-Y)(154-144),}{K_i-Y_{\min }}Y<K_i\\ 154-\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(3\right)$

${\stackrel{\&OverBar;}{C}}_b\left(Y\right)=\left\{\begin{array}{c}108+\frac{(K_i-Y)(118-108)}{K_l-Y_{\min }},Y<K_i\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(4\right)$

${\mathrm{Wc}}_i\left(Y\right)=\left\{\begin{array}{c}{\mathrm{WLc}}_i+\frac{(Y-Y_{\min })({\mathrm{Wc}}_i-{\mathrm{WLc}}_i)}{K_l-Y_{\min }},Y<K_l\\ {\mathrm{WHc}}_i+\frac{(Y_{\max }-Y)({\mathrm{Wc}}_i-{\mathrm{WHc}}_i)}{Y_{\max }-K_h},Y\&GreaterEqual;K_h\end{array}\right.---\left(5\right)$

(5) if Cb between [77,127] and Cr between [133,173], then this pixel is made as white, otherwise is made as black;

(6) each pixel in the image is carried out the processing of step (3), (4), (5), obtain the image after the skin color segmentation, obtain possible human face region.

3. the method that people's face in the dynamic image is blocked according to claim 1 is characterized in that generating face template and comprises the steps:

(1) chooses some secondary positive facial images of rectifying; Shear out human face region as people's face sample; After scale calibrationization (36 * 36) and intensity profile standardization, it is average that everyone face sample is got gray scale, obtains size and be the original average face template that rectify in 36 * 36 front;

(2) consider the importance of eyes in face characteristic, copy the eyes part of original average face template, shear out size and be 36 * 12 eye areas, it is carried out after the intensity profile standardization as the eyes template;

(3) with original face template respectively according to 1: 0.9, the length and width of 1: 1 and 1: 1.1 is than stretching, and respectively they carried out after the intensity profile standardization as face template, to adapt to difform people's face.

4. the method that people's face in the dynamic image is blocked according to claim 1 is characterized in that verifying foundation as follows based on people's face of template matches:

If the gray matrix of face template is T [M] [N], gray average and variance are respectively μ _TAnd σ _T, be R [M] [N] by the gray matrix of authentication image, gray average and variance are respectively μ _RAnd σ _R, then the related coefficient between them does

$r(T,R)=\frac{\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\left(T\right[i\left]\right[j]-{\mathrm{\&mu;}}_T)\left(R\right[i\left]\right[j]-{\mathrm{\&mu;}}_R)}{M\&CenterDot;N\&CenterDot;{\mathrm{\&sigma;}}_T\&CenterDot;{\mathrm{\&sigma;}}_R}---\left(10\right)$

Wherein, (T R) has reflected the angle of image vector between input picture zone and template to r, and (T, R) matching degree in big more representation template and input picture zone is high more for r.

Images