CN102426646B - Multi-angle human face detection device and method - Google Patents

Abstract

The invention discloses a multi-angle human face detection device and method, which mainly solve the problems in the prior art that color information in an image cannot be effectively utilized and a human face is mainly detected in a single angle. The device comprises an image acquisition module, an analog-digital conversion module, an image processing module, a data storage module and a communication module. The method comprises the steps that: (1) a human eye detector is established; (2) a human face profile template is established; (3) an image to be treated is acquired; (4) a mutual election human face area is acquired; (5) the front face is determined; (6) the profile face is determined; (7) the side face is determined; and (8) detection results are output. In the invention, the skin color information in the image is effectively utilized, so that computing resources are saved and the processing speed is improved; a two-value image establishing template can simply and effectively determine the side face; and the human face in a plurality of angles in a horizontal direction can be accurately detected by integrating skin color detection, human eye detection and template matching methods.

Description

Multi-angle human face detection device and method

Technical field

The invention belongs to a kind of Multi-angle human face detection device and method in technical field of image processing.The present invention uses Multi-angle face detection method to judge in the coloured image of camera collection whether have people's face, if there is people's face, and the position of people's face and shared region in positioning image.

Background technology

It is the position of everyone face (if existence) in judgement specify image and big or small process that people's face detects, it is a basic work, be widely used in man-machine interaction, recognition of face and tracking, figure image intensifying and retrieval, the content-based fields such as Video coding, have become research topic very active in pattern-recognition and computer vision.

The patented claim " method for detecting human face and system " (number of patent application CN200910077430.2, publication number CN101739549A) that Beijing ZANB Science & Technology Co., Ltd. proposes discloses a kind of method for detecting human face and system detecting for driver fatigue.The implementation step of the method is: step 1, and pretreatment image, coloured image gray processing is processed and reduction image resolution ratio; Step 2, processes image, comprises and obtains connected region and obtain integral image; Step 3, selected candidate face region, according to the connected region of obtaining and integral image, selected candidate face region; Step 4, checking candidate face region, by the human face region of judgment condition filtering falseness, and exports human face region.This patented claim also discloses a kind of device, comprising: pretreatment image module, processing image module, selected candidate face regions module and checking candidate face regions module.Although the simple and effective people's face that carried out of the method detects, people's face and human eye location in driver fatigue detection have been guaranteed preferably, but the deficiency still existing is: first the method is converted into coloured image gray level image and carries out the detection of people's face again, effectively do not utilize chromatic information, increased the processing time; The apparatus structure of this disclosure of the invention is complicated in addition, realizes difficulty.

The patented claim " a kind of method for detecting human face and device " (number of patent application 200810198047.8, publication number CN101344922A) that Huawei Tech Co., Ltd proposes discloses a kind of method for detecting human face.The method implementation step is: step 1, based on gray-scale statistical model, the people's face in video present frame is detected, and obtain candidate face region; Step 2, carries out color filter based on single channel complexion model to candidate face region, obtains people's face testing result.In addition, this patented claim also discloses a kind of device, and this device comprises: human face region detection module, for based on gray-scale statistical model, people's face of video present frame being detected, obtains candidate face region; Color filter module, carries out color filter for candidate face region human face region detection module being obtained based on single channel complexion model, obtains people's face testing result.Although the method utilizes the gray-level structure of people's face itself to exist the singularity of distinguishing mutually with other things to carry out the detection of people's face, utilize based on single channel complexion model simultaneously the most of false positive face of color filter eliminating is carried out in candidate face region, but the deficiency still existing is: first the method has adopted boosting algorithm to carry out the detection of people's face to carry out color filter again, boosting method is mainly comparatively effective to obverse face detection, but it is poor that other angles of horizontal direction people face (as side people's face) is detected to effect, can not realize multi-orientation Face and detect.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, propose the apparatus and method that in a kind of coloured image, multi-orientation Face detects, can effectively utilize Skin Color Information in coloured image, the multi-orientation Face of realizing horizontal direction detects.

To achieve these goals, apparatus of the present invention comprise image capture module, analog-to-digital conversion module, and image processing module, five modules of data memory module and communication module, connect by bus between each module; Wherein, image capture module adopts charge-coupled image sensor camera collection external analog signal color image data, and sends data to analog-to-digital conversion module; The analog signal data that analog-to-digital conversion module adopts analog-digital chip that image capture module is obtained is converted to digital signal data, for image processing module provides data; Image processing module adopts digital signal processor to process the image digital signal obtaining, and by Multi-angle human face detection algorithm, realizes the detection of people's face in image; Data memory module completes the storage of program and view data, and intermediate data temporary in image processing process; Communication module adopts USB (universal serial bus) sending and receiving device, realizes and outside serial communication output people face testing result.

Apparatus of the present invention realize Multi-angle face detection method and comprise the steps:

(1) set up human eye sorter

1a) adopt manual markings method from the gray scale picture that comprises human eye, to cut out eyes and the left and right simple eye image of human eye; The gray scale picture that never comprises human eye cuts out non-eye image at random, and eye image and non-eye image are respectively as positive sample and the negative sample in when training;

1b) negative sample of the positive sample of the human eye eyes of collecting and simple eye image and non-eye image is scaled to the sample of same size;

1c) with histogram equalization method to convergent-divergent after positive negative sample carry out illumination compensation, obtain the positive negative sample after compensation;

1d) the positive negative sample after compensation is carried out to rectangular characteristic extraction, set up respectively eyes and simple eye cascade classifier.

(2) establishment side face face template

2a) collecting horizontal direction turns half side people's face between 20～50 degree and horizontal direction towards right avertence and turns full side facial image between 50～90 degree as sample towards right avertence;

2b) sample of collecting is scaled to the sample of same size size;

2c) by histogram equalization method, the sample after to convergent-divergent carries out illumination compensation, obtains the sample after compensation;

2d) sample after compensation is carried out to binarization of gray value processing, obtain binaryzation sample;

After 2e) respectively the gray-scale value of the binaryzation sample same position pixel of people's face is cumulative, average, using mean value image respectively as the half side face template turning towards right avertence and full side face template, the half side face template and the full side face template that utilize facial symmetry to obtain turning towards left avertence.

(3) obtain pending image

Image capture module is input to D/A converter module by the simulating signal coloured image obtaining, and analog-to-digital conversion module is using the digital color signal image after conversion as follow-up pending image.

(4) obtain candidate face region

4a) utilize oval complexion model method to detect the colour of skin point of pending image, obtain the area of skin color of image;

4b) area of skin color is carried out to the processing of colour of skin binary image, obtain bianry image;

4c) bianry image is carried out to integral projection, obtain length and the width of area of skin color, the area of skin color by Aspect Ratio between 0.5～2 is decided to be candidate face region.

(5) judge front face

5a) candidate face region is converted into gray level image;

5b) gray level image is dwindled continuously, the first half and step 1c in image after exhaustive dwindling) the onesize subwindow of sample of the eyes detecting device set up, if existed in these subwindows by the window of eyes detection of classifier, illustrate and eyes detected, judge and have front face, execution step (8); Otherwise, carry out next step.

(6) judge half side people's face

6a) by step 5a) in gray level image dwindle continuously, the first half and step 1c in image after exhaustive dwindling) the onesize subwindow of sample of the simple eye detecting device set up, if existed in these subwindows by the window of simple eye detection of classifier, illustrate detect simple eye, binarization of gray value processing is carried out in gray level image region, obtain bianry image, carry out next step; Otherwise, execution step (7);

6b) utilizing step 2e) the half side face template set up carries out template matches to bianry image, obtains the related coefficient of coupling, if related coefficient is more than or equal to 0.6, judges and have half side people's face, execution step (8); Otherwise, carry out next step.

(7) judge full side people's face

7a) by step 5a) gray level image region carry out binarization of gray value processing, obtain bianry image;

7b) utilizing step 2e) the full side face template set up carries out template matches to bianry image, obtain the related coefficient of coupling, if related coefficient is more than or equal to 0.6, judges and have full side people's face, otherwise judge unmanned face in this image, execution step (8).

(8) output detections result.

The present invention compared with prior art has following advantage:

First, the present invention adopts complexion model to detect the relatively little area of skin color of ratio in coloured image, then in this region, carries out the detection of people's face, has overcome prior art entire image is carried out to the larger shortcoming of people's face detection computations amount, make the present invention save computational resource, improved processing speed.

The second, the method that the present invention adopts Face Detection, human eye detection and template matches to combine, has overcome the shortcoming that prior art can only detect front face, makes people's face that the present invention can a plurality of angles of detection level direction, comprises front face and people from side face.

The 3rd, the present invention adopts bianry image establishment side face face template to carry out profile face detection, take full advantage of people's face side feature simple feature that distributes, overcome prior art template and set up complicated shortcoming, side of the present invention face template is set up easily, judged that the face computing of people from side is simply effective.

Accompanying drawing explanation

Fig. 1 is the block scheme of apparatus of the present invention;

Fig. 2 is the process flow diagram of the inventive method;

Fig. 3 is simulated effect figure of the present invention.

Concrete implementing measure

Below in conjunction with accompanying drawing, invention is described further.

With reference to accompanying drawing 1, apparatus of the present invention comprise five modules: image capture module, and analog-to-digital conversion module, image processing module, data memory module and communication module, connect by bus between each module.Wherein, image capture module adopts charge-coupled image sensor camera collection external analog signal color image data, and sends data to analog-to-digital conversion module; The analog signal data that analog-to-digital conversion module adopts analog-digital chip that image capture module is obtained is converted to digital signal data, for image processing module provides data; Image processing module adopts digital signal processor to process the image digital signal obtaining, and by Multi-angle human face detection algorithm, realizes the detection of people's face in image; Data memory module completes the storage of program and view data, and intermediate data temporary in image processing process; Communication module adopts USB (universal serial bus) sending and receiving device, realizes and outside serial communication output people face testing result.

Apparatus of the present invention comprise flash memory and dynamic RAM, and flash memory completes the storage of program, and dynamic RAM completes the temporary of view data and provides Multi-angle human face detection algorithm required internal memory.

In embodiments of the present invention, image capture module adopts SONY HQ1 model charge-coupled image sensor camera, and its signaling mode is Phase Alternate Line, and resolution (horizontal center) is 540 television lines; Analog-to-digital conversion module adopts ADV7183 analog-digital chip, it is a enhancement mode Video Decoder that comprises 10 analog to digital converters, include two 10 accurate analog to digital converters and complete automatic gain control circuit, there are 6 analog video signal input channels, analog signal image can be converted to digital signal image; Image processing module adopts ADSPBF533 chip, powerful digital signal processing arithmetic capability and multiple interfaces can be provided, not only can complete the control function of conventional microprocessor, can also use two-dimentional direct memory access transmission data for feature of image, greatly accelerate transmission and the processing of view data; Flash memory adopts M25P64 chip, and dynamic RAM adopts MT48LC32M16A2TG-75 chip; Communication module adopts MAX232 chip to form the serial communication circuit that meets standard RS232 agreement.

The concrete steps of 2 pairs of the inventive method are described below by reference to the accompanying drawings:

Step 1, sets up human eye sorter

Adopt manual markings method from the gray scale picture that comprises human eye, to cut out eyes and the left and right simple eye image of human eye; The gray scale picture that never comprises human eye cuts out non-eye image at random, and eye image and non-eye image are respectively as positive sample and the negative sample in when training.

The positive negative sample of collecting is scaled to same size, in the embodiment of the present invention, eyes sample image is scaled to 22 * 5 pixels, simple eye sample image is normalized to 18 * 12 pixels, and non-human eye sample respectively corresponding human eye sample is scaled identical size.

In order to adjust the brightness of positive negative sample, positive negative sample by histogram equalization method after to convergent-divergent carries out illumination compensation, first calculating the grey level histogram of positive negative sample, is then equally distributed form by histogrammic changes in distribution, obtains thus the rear positive negative sample of compensation.

Proper vector using rectangle as human eye detection is carried out feature extraction to the positive negative sample after compensating, adopt Adaboost algorithm to set up respectively eyes and simple eye cascade classifier, first utilize the features training Weak Classifier extracting, by Weak Classifier cascade, form strong classifier again, then by strong classifier cascade, form cascade classifier.The structure of cascade classifier adopts by heavily to gently, the method going from the simple to the complex, this method can be got rid of a large amount of non-human eyes when front layer detects human eye, and the less important feature layer of putting behind, to further get rid of, can complete fast and effectively thus the detection of human eye.

Step 2, establishment side face face template

Collecting horizontal direction turns half side people's face between 20～50 degree and horizontal direction towards right avertence and turns full side facial image between 50～90 degree as sample towards right avertence, selected sample all comprises complete hair and face, and in these images, hair does not all block ear.The sample of collecting is scaled to the sample of same size size, in the embodiment of the present invention, sample unification is scaled to 70 * 110 pixel sizes.

In order to adjust the brightness of sample, the sample by histogram equalization method after to convergent-divergent carries out illumination compensation, first calculates the grey level histogram of sample, by histogrammic changes in distribution, is then equally distributed form, obtains thus sample after compensation.

Sample after compensation is carried out to binarization of gray value processing, gray-scale value in sample is less than to an extreme value of the pixel tax gray-scale value of certain threshold value, the pixel that is greater than this threshold value is composed contrary extreme value, obtains thus two-value sample.In the embodiment of the present invention, the Threshold that binarization of gray value is processed is 90, and the pixel assignment 255 that gray-scale value is greater than to 90, is less than 90 pixel assignment 0.

After respectively the gray-scale value of half side and full side people's face two-value sample same position pixel being added up, average, obtain the average image, using the average image respectively as the half side face template turning towards right avertence and full side face template, the half side face template and the full side face template that utilize facial symmetry to obtain turning towards left avertence.

Step 3, obtains pending image

Image capture module is input to D/A converter module by the simulating signal coloured image obtaining, and analog-to-digital conversion module is using the digital color signal image after conversion as follow-up pending image.

Step 4, obtains candidate face region

Utilize oval complexion model method to detect the colour of skin point of pending image, in color space YCrCb, each pixel in pending image, through following formula computing, asked to this eigenwert in CrCb space:

$\left[\begin{array}{c}x\\ y\end{array}\right]=\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}{\mathrm{Cb}}^{\′}-c_x\\ {\mathrm{Cr}}^{\′}-c_y\end{array}\right]$

Wherein, x and y are the eigenwert of pixel in CrCb space, θ=2.53rad, c _x=109.38, c _y=152.02, Cr ' and Cb ' represent that respectively Cr and Cb are through the non-linear value obtaining after color transformed.

By the x of above formula gained, y value is brought in following formula, if acquired results is not more than 1, just judges that this point is as colour of skin point:

(x-ec _x) ²/a ²+(y-ec _y) ²/b ²

Wherein, x and y are the eigenwert of pixel in CrCb space, ec _x=1.60, a=25.39, ec _y=2.41, b=14.03;

The colour of skin point detecting is connected to form to area of skin color.In the embodiment of the present invention, in order to obtain smoother area of skin color, also this region has been carried out the operation of dilation and corrosion, what expand employing is that four directions is to the method for judgement, when the color that has up and down a point of current point is black, just current point is filled to black, can connects like this in area of skin color discontinuous; The same four directions that adopts of corrosion, to the method for judgement, checks current point four points up and down, if there is a point different with the color of current point, so just current some same color is filled, and can make so non-area of skin color diminish, and returns to the size before expansion.

In another width and the onesize gray level image of pending image, the position corresponding with area of skin color composed to an extreme value of gray-scale value, contrary extreme value is composed in the position corresponding with non-area of skin color, obtains bianry image.In the embodiment of the present invention, by the position assignment 0 corresponding with area of skin color, the position assignment 255 corresponding with non-area of skin color.

Bianry image is carried out to horizontal projection and vertical projection obtains perspective view, the length that in vertical projection diagram, the width of skin distribution concentrated area is area of skin color, in horizontal projection, the width of skin distribution concentrated area is area of skin color width, obtain thus area of skin color Aspect Ratio, the area of skin color by Aspect Ratio between 0.5～2 is decided to be candidate face region.

Step 5, judges front face

Distribution characteristics according to human eye in face, carries out human eye detection to the first half image in candidate face region.Candidate face region is converted into gray level image, gray level image is dwindled continuously, the onesize subwindow of sample of the eyes detecting device of the first half and foundation in the image after exhaustive dwindling, if existed in these subwindows by the window of eyes detection of classifier, illustrate and eyes detected, judge and have front face, perform step 8; Otherwise, carry out next step.

Step 6, judges half side people's face

Gray level image after candidate face is transformed dwindles continuously, the onesize subwindow of the sample of the simple eye detecting device of the first half and foundation in image after exhaustive dwindling, if existed in these subwindows by the window of simple eye detection of classifier, illustrate detect simple eye, binarization of gray value processing is carried out in gray level image region, an extreme value that gray-scale value in image is less than to the pixel tax gray-scale value of certain threshold value, the pixel that is greater than this threshold value is composed contrary extreme value, obtains thus bianry image; Otherwise, perform step 7.In the embodiment of the present invention, the Threshold that binarization of gray value is processed is 90, and the pixel assignment 255 that gray-scale value is greater than to 90, is less than 90 pixel assignment 0.

Utilize the half side face template of setting up to carry out template matches to bianry image, by following formula meter, mate related coefficient:

$R(i,j)=\frac{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{S}^{\mathrm{ij}}(m,n)\×T(m,n)}{\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[S^{\mathrm{ij}}(m,n)\right]}^2}\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[T(m,n)\right]}^2}}$

Wherein, R (i, j) represents the related coefficient of region of search and template, and i and j represent the coordinate range of region of search, and M and N represent length and the width of searched figure, and S represents searched figure, and m and n represent length and the width of template, and T represents template;

If related coefficient is more than or equal to 0.6, judge and have half side people's face, perform step 8; Otherwise, carry out next step.

Step 7, judges full side people's face

Binarization of gray value processing is carried out in gray level image region after candidate face region is transformed, gray-scale value in image is less than to an extreme value of the pixel tax gray-scale value of certain threshold value, and the pixel that is greater than this threshold value is composed contrary extreme value, obtains thus bianry image.In the embodiment of the present invention, the Threshold that binarization of gray value is processed is 90, and the pixel assignment 255 that gray-scale value is greater than to 90, is less than 90 pixel assignment 0.

Utilize the full side face template of setting up to carry out template matches to bianry image, by following formula meter, mate related coefficient:

$R(i,j)=\frac{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{S}^{\mathrm{ij}}(m,n)\×T(m,n)}{\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[S^{\mathrm{ij}}(m,n)\right]}^2}\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[T(m,n)\right]}^2}}$

Wherein, R (i, j) represents the related coefficient of region of search and template, and i and j represent the coordinate range of region of search, and M and N represent length and the width of searched figure, and S represents searched figure, and m and n represent length and the width of template, and T represents template;

If related coefficient is more than or equal to 0.6, judge and have full side people's face, otherwise judge unmanned face in this image, perform step 8.

Step 8, output detections result.

Below in conjunction with 3 pairs of effects of the present invention of accompanying drawing, be described further.

1. experiment condition and content

The simulated environment that accompanying drawing 3 is realized is: with SONY HQ1 model charge-coupled image sensor camera collection external image, ADV7183 analog-digital chip is converted to digital signal image by analog signal image, ADSPBF533 chip operation Multi-angle human face detection algorithm, M25P64 chip-stored program, MT48LC32M 16A2TG-75 provides algorithm operation required memory, MAX232 chipset becomes serial communication circuit output detections result, and obtains the image in the detection of people's face by the jtag interface of ADSPBF533.

Concrete emulation content of the present invention is: under indoor environment, gather respectively the side image of people's face front and horizontal direction deflection, the Multi-angle face detection method that adopts the present invention to propose carries out the detection of people's face in image.

2. experimental result

Accompanying drawing 3 is simulation result figure of the present invention, wherein, in Fig. 3 (a), is front face, and in figure, square frame has marked eyes and the front face detecting; In Fig. 3 (b), be half side people's face, in figure, square frame has marked the simple eye and half side people's face detecting; In Fig. 3 (c), be full side people's face, in figure, square frame has marked the full side people's face detecting.As seen from Figure 3, use apparatus and method of the present invention people's face of horizontal direction multi-angle can be detected accurately.

Claims (8)

1. a Multi-angle face detection method, the hardware system forming based on ADSPBF533 digital signal processor, ADV7183 analog-digital chip, M25P64 flash memory and MT48LC32M16A2TG-75 dynamic RAM, realize multi-orientation Face and detect, the concrete steps of the method are as follows:

(1) set up human eye sorter

1a) adopt manual markings method from the gray scale picture that comprises human eye, to cut out eyes and the left and right simple eye image of human eye; The gray scale picture that never comprises human eye cuts out non-eye image at random, and eye image and non-eye image are respectively as positive sample and the negative sample in when training;

1b) negative sample of the positive sample of the human eye eyes of collecting and simple eye image and non-eye image is scaled to the sample of same size;

1c) with histogram equalization method to convergent-divergent after positive negative sample carry out illumination compensation, obtain the positive negative sample after compensation;

1d) the positive negative sample after compensation is carried out to rectangular characteristic extraction, set up respectively eyes and simple eye cascade classifier;

(2) establishment side face face template

2a) collecting horizontal direction turns half side people's face between 20～50 degree and horizontal direction towards right avertence and turns full side facial image between 50～90 degree as sample towards right avertence;

2b) sample of collecting is scaled to the sample of same size size;

2c) by histogram equalization method, the sample after to convergent-divergent carries out illumination compensation, obtains the sample after compensation;

2d) sample after compensation is carried out to binarization of gray value processing, obtain binaryzation sample;

After 2e) respectively the gray-scale value of the binaryzation sample same position pixel of people's face is cumulative, average, using mean value image respectively as the half side face template turning towards right avertence and full side face template, the half side face template and the full side face template that utilize facial symmetry to obtain turning towards left avertence;

(3) obtain pending image

Image capture module is input to ADV7183 analog-digital chip by the simulating signal coloured image obtaining, and ADV7183 analog-digital chip is using the digital color signal image after conversion as follow-up pending image;

(4) obtain candidate face region

4a) utilize oval complexion model method to detect the colour of skin point of pending image, obtain the area of skin color of image;

4b) area of skin color is carried out to the processing of colour of skin binary image, obtain bianry image;

4c) bianry image is carried out to integral projection, obtain length and the width of area of skin color, the area of skin color by Aspect Ratio between 0.5～2 is decided to be candidate face region;

(5) judge front face

5a) candidate face region is converted into gray level image;

5b) gray level image is dwindled continuously, the first half and step 1c in image after exhaustive dwindling) the onesize subwindow of sample of the eyes set up, if existed in these subwindows by the window of eyes detection of classifier, illustrate and eyes detected, judge and have front face, execution step (8); Otherwise, carry out next step;

(6) judge half side people's face

6a) by step 5a) in gray level image dwindle continuously, the first half and step 1c in image after exhaustive dwindling) the onesize subwindow of simple eye sample set up, if existed in these subwindows by the window of simple eye detection of classifier, illustrate detect simple eye, binarization of gray value processing is carried out in gray level image region, obtain bianry image, carry out next step; Otherwise, execution step (7);

6b) utilizing step 2e) the half side face template set up carries out template matches to bianry image, obtains the related coefficient of coupling, if related coefficient is more than or equal to 0.6, judges and have half side people's face, execution step (8); Otherwise, carry out next step;

(7) judge full side people's face

7a) by step 5a) gray level image region carry out binarization of gray value processing, obtain bianry image;

7b) utilizing step 2e) the full side face template set up carries out template matches to bianry image, obtain the related coefficient of coupling, if related coefficient is more than or equal to 0.6, judges and have full side people's face, otherwise judge unmanned face in this image, execution step (8);

(8) output detections result.

2. Multi-angle face detection method according to claim 1, is characterized in that, step 1d) described rectangular characteristic refers to the proper vector using rectangle as human eye detection.

3. Multi-angle face detection method according to claim 1, it is characterized in that step 1d) described cascade classifier employing Adaboost algorithm foundation, the method is first trained Weak Classifier, by Weak Classifier cascade, form strong classifier, then the sorter being formed by strong classifier cascade.

4. Multi-angle face detection method according to claim 1, it is characterized in that, step 2d), step 6a) and step 7a) described binarization of gray value is processed is the extreme value that pixel that gray-scale value in gray level image is less than to certain threshold value is composed gray-scale value, the pixel that is greater than this threshold value is composed contrary extreme value.

5. Multi-angle face detection method according to claim 1, it is characterized in that, step 4a) described oval complexion model method is: in color space YCrCb, a pixel in image, through following formula computing, is asked to this eigenwert in CrCb space:

$\left[\begin{array}{c}x\\ y\end{array}\right]=\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}{\mathrm{Cb}}^{\′}-c_x\\ {\mathrm{Cr}}^{\′}-c_y\end{array}\right]$

Wherein, x and y are the eigenwert of pixel in CrCb space, θ=2.53rad, c _x=109.38, c _y=152.02, Cr ' and Cb ' represent that respectively Cr and Cb are through the non-linear value obtaining after color transformed;

By the x of above formula gained, y value is brought in following formula, if acquired results is not more than 1, just judges that this point is as colour of skin point;

(x-ec _x) ²/a ²+(y-ec _y) ²/b ²

Wherein, x and y are the eigenwert of pixel in CrCb space, ec _x=1.60, a=25.39, ec _y=2.41, b=14.03.

6. Multi-angle face detection method according to claim 1, it is characterized in that, step 4b) described colour of skin binary conversion treatment is the extreme value that the correspondence position in another onesize gray level image is composed gray-scale value by area of skin color in image, and the correspondence position of non-area of skin color is composed contrary extreme value.

7. Multi-angle face detection method according to claim 1, it is characterized in that, step 4c) described integral projection method is bianry image to be carried out to horizontal projection and vertical projection obtains perspective view, the length that in vertical projection diagram, the width of skin distribution concentrated area is area of skin color, in horizontal projection, the width of skin distribution concentrated area is area of skin color width.

8. Multi-angle face detection method according to claim 1, is characterized in that, step 6b) and step 7b) described related coefficient calculates acquisition by following formula:

$R(i,j)=\frac{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{S}^{\mathrm{ij}}(m,n)\×T(m,n)}{\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[S^{\mathrm{ij}}(m,n)\right]}^2}\sqrt{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\left[T(m,n)\right]}^2}}$

Wherein, R (i, j) represents the related coefficient of region of search and template, and i and j represent the coordinate range of region of search, and M and N represent length and the width of searched figure, and S represents searched figure, and m and n represent length and the width of template, and T represents template.

Images