CN105893925A - Human hand detection method based on complexion and device - Google Patents

Abstract

The invention provides a human hand detection method based on complexion. Acquired image to be detected is converted from an RGB color space to an HSV color space to acquire an HSV image. The image to be detected is converted from the RGB color space to an r-g color space to acquire an r-g image. The HSV image is converted into a first binary image. The r-g image is converted into a second binary image. Step-by-step operation is carried out on first and second binary images to acquire a comprehensive binary image. The comprehensive binary image is filtered to acquire an optimized binary image. The largest communication region in the optimized binary image is analyzed and is used as a skin region. A pre-trained K neighbor classifier is used to determine whether the largest communication region is hand-shaped to realize human hand recognition. The detection speed is fast. Human hand erroneous detection in gesture recognition is effectively solved.

Description

Staff detection method based on the colour of skin and device

Technical field

The present embodiments relate to computer vision field, particularly relate to the detection of a kind of staff based on the colour of skin Method and device.

Background technology

In the various Vision Builder for Automated Inspection relevant with people, gesture identification is more and more taken seriously, such as In man-machine interactive system based on gesture, need first image obtains the position of hands.And it is the most normal Method be through the colour of skin is detected thus obtain gesture information.Hands is split from image, The most the most frequently used dividing method is namely based on the segmentation of the colour of skin.

According to either with or without relating to the process of imaging, the method for Face Detection is divided into two kinds of fundamental types: based on The method of statistics and method based on physics.Skin color detection method based on statistics is mainly by setting up the colour of skin Statistical model carries out Face Detection, mainly includes two steps: color notation conversion space and skin color modeling；Base Method in physics then introduces the interaction between illumination and skin, by the research colour of skin in Face Detection Reflection model and spectral characteristic carry out Face Detection.

But existing based in the skin color detection method added up, the recognition efficiency of people's hand is low, false drop rate High and be highly susceptible to the impact of illumination, thus cause the accuracy of gesture identification to be restricted.

Therefore, a kind of quick and high-quality staff detection method urgently proposes.

Summary of the invention

The embodiment of the present invention provides a kind of staff detection method based on the colour of skin and device, existing in order to solve Face Detection based on statistics and staff recognition methods efficiency is low, false drop rate is high and very easy is subject in technology Defect to the impact of illumination, it is achieved that staff high efficiency based on Face Detection, the identification of high accuracy, Thus further increase the accuracy of gesture identification.

The embodiment of the present invention provides a kind of staff detection method based on the colour of skin, including:

The image to be detected got is transformed into hsv color space to obtain from RGB color HSV image, and described image to be detected is transformed into r-g color space to obtain from RGB color R-g image；

Traversal reads each pixel in described HSV image, and according to the HSV Nogata pre-build Described HSV image is converted into the first bianry image by graph model, and travels through in the described r-g image of reading Each pixel, is converted into the second binary map according to the mixed Gauss model pre-build by described r-g image Picture；

Described first bianry image and described second bianry image are carried out step-by-step and with computing thus obtains comprehensive Bianry image；

It is filtered described comprehensive bianry image obtaining the bianry image after optimizing；

Analyze connected region maximum in the bianry image after described optimization, by the connected region of described maximum As skin area；

The k nearest neighbor grader using training in advance judges whether the connected region of described maximum is hand, from And realize the identification of staff.

The embodiment of the present invention provides a kind of staff based on colour of skin detection device, including:

Image conversion module, for being transformed into HSV by the image to be detected got from RGB color Color space is to obtain HSV image, and from RGB color, described image to be detected is transformed into r-g Color space is to obtain r-g image；

Binary map acquisition module, for traveling through each pixel read in described HSV image, and according to Described HSV image is converted into the first bianry image by the HSV histogram model pre-build, and travels through Read each pixel in described r-g image, according to the mixed Gauss model pre-build by described r-g Image is converted into the second bianry image；

Digitwise operation module, for described first bianry image and described second bianry image carried out step-by-step with Computing thus obtain comprehensive bianry image；

Filtration module, for being filtered described comprehensive bianry image obtaining the bianry image after optimizing；

Connected region judge module, connected region maximum in the bianry image after analyzing described optimization, Using the connected region of described maximum as skin area；

Staff identification module, for using the k nearest neighbor grader of training in advance to judge the connection of described maximum Whether region is hand, thus realizes the identification of staff.

The embodiment of the present invention provide skin color detection method and device, by integrated use HSV rectangular histogram, The method that mixed Gauss model, filtering and noise reduction and connected domain are extracted, it is achieved that the high precision of skin area Degree detection, meanwhile, the staff being achieved fast accurate by k nearest neighbor grader is extracted.

Accompanying drawing explanation

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to reality Execute the required accompanying drawing used in example or description of the prior art to be briefly described, it should be apparent that below, Accompanying drawing in description is some embodiments of the present invention, for those of ordinary skill in the art, not On the premise of paying creative work, it is also possible to obtain other accompanying drawing according to these accompanying drawings.

Fig. 1 is the techniqueflow chart of the embodiment of the present invention one；

Fig. 2 is the techniqueflow chart of the embodiment of the present invention two；

Fig. 3 is the techniqueflow chart of the embodiment of the present invention three；

Fig. 4 is the apparatus structure schematic diagram of the embodiment of the present invention four.

Detailed description of the invention

For making the purpose of the embodiment of the present invention, technical scheme and advantage clearer, below in conjunction with this Accompanying drawing in bright embodiment, is clearly and completely described the technical scheme in the embodiment of the present invention, Obviously, described embodiment is a part of embodiment of the present invention rather than whole embodiments.Based on Embodiment in the present invention, those of ordinary skill in the art are obtained under not making creative work premise The every other embodiment obtained, broadly falls into the scope of protection of the invention.

It should be noted that each embodiment of the present invention is not individually present, between several embodiments Can be complementary to one another or combine existence.

Embodiment one

Fig. 1 is the techniqueflow chart of the embodiment of the present invention one, in conjunction with Fig. 1, embodiment of the present invention one base Staff detection method in the colour of skin comprises the following steps that

Step 110: the image to be detected got is transformed into hsv color from RGB color empty Between to obtain HSV image, and described image to be detected is transformed into r-g color from RGB color Space is to obtain r-g image；

With lower part in order to make logical description clearer, step 110 is split as two steps: step 111 and step 112, it should be noted that there is no sequencing between step 111 and step 112, with Under description its execution sequence is not intended that restriction.

Step 111: the image to be detected got is transformed into hsv color from RGB color empty Between to obtain HSV image；

RGB color be by red (R), green (G), the change of blue (B) three Color Channels and it Superposition each other obtain color miscellaneous, RGB is i.e. to represent red, green, blue three The color value of passage, this standard almost include human eyesight can all colours of perception.

HSV (Hue Saturation Value: colourity intensity value) color space is straight according to color A kind of color space seeing characteristic and create, H, S and V represent tone, saturation and brightness respectively. Image to be detected is transformed into hsv color space from RGB color, overcomes to a certain extent The illumination variation impact on Face Detection.

In HSV color space, tone H represents color information, the position of i.e. residing spectral color. H angle is measured, and span is 0 °～360 °, starts to calculate counterclockwise from redness, red Being 0 °, green is 120 °, and blueness is 240 °.Their complementary color is: yellow is 60 °, and cyan is 180 °, Magenta is 300 °；Saturation S is expressed as the ratio between purity and the purity that this color is maximum of selected color Rate, the span of S is 0.0～1.0, is worth the biggest, and color is the most saturated, during S=0, only gray scale；Bright Degree V is generally with percentage measurement, from 0% (black) to 100% (in vain).RGB and CMY color Model is all towards hardware, and HSV (Hue Saturation Value) color model is user oriented 's.The three dimensional representation of HSV model develops from RGB Cube.Imagine from RGB along cube pair The White vertex of linea angulata is observed to Black vertices, it is possible to see cubical hexagonal external feature.Hexagon Boundary representation color, trunnion axis represents purity, and lightness is measured along vertical axis.

The embodiment of the present invention use formula below turned from RGB color by described image to be detected Change to hsv color space:

V=max (R, G, B)

$S=\left\{\begin{array}{cc}\frac{max(R,G,B)-min(R,G,B)}{max(R,G,B)}& max(R,G,B)\≠0\\ 0& max(R,G,B)=0\end{array}\right.$

$H=ar\cos \frac{(R-G)+(R-B)}{2\sqrt{(R-G)+(R-G)+(R-B)+(R-B)}}$

Wherein, R be the red value of described pixel, G be the green value of described pixel, B be described The blue valve of pixel；Max () represents and takes maximum operation, and min () represents and takes minimum operation, V be R, Maximum in G, B；H, S, V are respectively the color value that after converting, described pixel is corresponding.

Step 112: described image to be detected is transformed into r-g color space to obtain from RGB color Take r-g image；

In the embodiment of the present invention, equation below is used to be changed from by RGB color by described RGB image To r-g color space:

$r=\frac{R}{R+G+B}$

$g=\frac{G}{R+G+B}$

B=1-g-r

Wherein, R be the red value of described pixel, G be the green value of described pixel, B be described picture The blue valve of vegetarian refreshments；R, g, b are respectively the color value that after converting, described pixel is corresponding.

RGB color herein refers to by red (R), green (G), the change of blue (B) three Color Channels Change and they superpositions each other obtain color miscellaneous.Under normal circumstances, RGB is respectively arranged with 256 grades of brightness, with numeral be expressed as from 0,1,2... is until 255.One RGB color value specify red, Green, the relative luminance of primary colors, generates a particular color for display, i.e. any one color Can be recorded by one group of rgb value and be expressed.Such as, the rgb value that a certain pixel is corresponding be (149, 123,98), the color of this pixel is the superposition of different brightness of tri-kinds of colors of RGB.

In the embodiment of the present invention, use OpenCv can directly obtain the RGB that in picture, each pixel is corresponding Value, it is achieved code can be so that

CvScalar p；

P=cvGet2D (ImageIn, j, i)；

Double a=p.val [0]；

Double b=p.val [1]；

Double c=p.val [2]；

Wherein, i, j are pixel transverse and longitudinal coordinate on image respectively；Passage 0,1,2 correspondence respectively Be the brightness number of three kinds of colors blue, green, red；

After pixel value on duty is transformed into r-g space by rgb space, illumination can be overcome to a certain extent to become Change the impact on Face Detection.In the embodiment of the present invention, color space is converted into r-g by RGB, real It it is the normalization process to rgb color on border.During this is normalized, when certain pixel light According to or the impact of shade and produce Color Channel R, G, B value change time, the molecule in normalization formula divides Mother changes simultaneously, and the actual floating of normalized value obtained is the most little, and this mapping mode removes from image The information of illumination, therefore can weaken the impact of illumination.

Such as: before normalization, the pixel value of the pixel A in T1 moment is: RGB (30,60,90), the T2 moment, Owing to by illumination effect, the color value of tri-Color Channels of RGB creates change, the pixel value of pixel A Become RGB (60,120,180).

After normalization formula is converted into r-g space, the pixel value of the pixel A in T1 moment is: rgb (1/6, 1/3,2/3), the pixel value of the pixel A in T2 moment is: rgb (1/6,1/3,2/3).As can be seen here, T1 and The value of the normalization RGB in T2 moment does not change.

Step 120: traversal reads each pixel in described HSV image, and according to pre-building Described HSV image is converted into the first bianry image by HSV histogram model, and travels through the described r-g of reading Each pixel in image, is converted into according to the mixed Gauss model pre-build by described r-g image Two bianry images；

Clearer in order to describe with lower part, step 120 is split as four steps: step 121～step Rapid 125.Wherein, the entirety that step 122～step 125 are constituted there is no in reality performs with step 121 Fixing sequencing, the embodiment of the present invention does not limits.

Step 121: read the HSV value of described pixel, calculate described HSV value respectively with described skin The coupling of the HSV histogram model of skin pixel and the HSV histogram model of described non-skin pixel is general Rate value, and judge whether described pixel belongs to skin area according to described matching degree value；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel It is not belonging to skin area, is then described pixel assignment with y, thus obtains described first bianry image. Wherein, x typically takes 255, and y typically takes 0.

The described HSV histogram model of training in advance is preserved skin pixels and non-skin pixel The histogram distribution of HSV value, using this distribution as judging a new pixel in the embodiment of the present invention It it is whether the one reference of skin pixels.It is implemented as: the described pixel reading in image to be detected HSV value, calculate described HSV value respectively with HSV histogram model and the institute of described skin pixels State the matching probability value of the HSV histogram model of non-skin pixel, and judge according to described matching degree value Whether described pixel belongs to skin area.

In the present embodiment, by RGB image is converted to hsv color space so that carry out colour of skin inspection During survey, testing result has certain stability to the change of illumination.

S122: calculate the described pixel the first probability density under skin mixed Gauss model and described The pixel the second probability density under non-skin mixed Gauss model；

Mixed Gauss model GMM, also referred to as MOG, be the extension of single Gauss model, and it uses K (base Originally it being 3 to 10) individual Gauss model carrys out the feature of each pixel in phenogram picture.

The formulae express of single Gauss model is as follows:

Wherein, x belongs to d and ties up Euclidean space, and a is the mean vector of single Gauss model, and S is single Gauss The covariance matrix of model, ()^TThe transposition computing of representing matrix, ()^-1The inverse operation of representing matrix.

The formula of mixed Gauss model is formed according to weight is cumulative by K single Gauss model, uses following formula Embody:

$p(x;a_k,S_k,{\mathrm{\π}}_k)=\underset{k=1}{\overset{m}{\Σ}}{\mathrm{\π}}_k{p}_k\left(x\right),{\mathrm{\π}}_k\≥0,\underset{k=1}{\overset{m}{\Σ}}{\mathrm{\π}}_k=1$

Wherein, π_kBeing the weight of kth Gauss model, m is the number of default Gauss model, p_kX () is Kth list Gauss model.Wherein, for kth list Gauss model, its formulae express is as follows:

As it has been described above, x belongs to d ties up Euclidean space, m is the number of default Gauss model, p_kX () is The probability density of kth Gauss model, a_kIt is the mean vector of kth Gauss model, S_kIt is kth Gauss The covariance matrix of model, π_kIt it is the weight of kth Gauss model；

It should be noted that p (x；a_k, S_k, π_k) and p_kWhat x () results of calculation characterized is that x is in respective mode Probability density under type.

In the embodiment of the present invention, skin pixels and non-skin pixel are set up respectively mixed Gauss model, two kinds The formulae express of model is identical, and difference is the parameter in model, i.e. mean vector a_kAnd covariance Matrix S_kDifferent.

For each pixel in image to be detected, the embodiment of the present invention is under skin mixed Gauss model Calculate its first probability density, under non-skin mixed Gauss model, calculate its second probability density, until Travel through all of pixel.

In the embodiment of the present invention, the process of described traversal can be by row by row travel through one by one, it is also possible to be with A pixel chosen by machine, it is judged that whether it is the pixel of skin area, first to one Pixel in sizing neighborhood travels through, and the present invention is not limiting as.

When the mean vector of skin mixed Gauss model is a_k1, covariance matrix be S_k1And it is multiple single high The weight that this model is corresponding respectively is π_k1Time,

$p_{skin}(x;a_{k1},S_{k1},{\mathrm{\π}}_{k1})=\underset{k1=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k1}{p}_{k1}\left(x\right),{\mathrm{\π}}_{k1}\≥0,\underset{k1=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k1}=1$

When the mean vector of non-skin mixed Gauss model is a_k2, covariance matrix be S_k2And multiple list The weight that Gauss model is corresponding respectively is π_k2Time,

$p_{non-skin}(x;a_{k2},S_{k2},{\mathrm{\π}}_{k2})=\underset{k2=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k2}{p}_{k2}\left(x\right),{\mathrm{\π}}_{k2}\≥0,\underset{k2=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k2}=1$

S123: described first probability density and described second probability density according to described pixel calculate institute State pixel and belong to the posterior probability of skin area；

In the embodiment of the present invention, the computing formula of posterior probability is as follows:

$P=\frac{p_{skin}}{p_{skin}+p_{non-skin}}$

Wherein, P is the value of described posterior probability, p_skinFor described first probability density；p_non-skinFor institute State the second probability density.

S124: when judging described posterior probability more than the posterior probability threshold value preset, by described pixel Belong to skin area；

Preferably, described posterior probability threshold value is set to 0.5 by the embodiment of the present invention, i.e. when described posterior probability Value more than 0.5 time, it is judged that pixel corresponding to described posterior probability belongs to skin area.Posterior probability Threshold value 0.5 is an empirical value, draws through substantial amounts of experimental judgment, if a pixel belongs to skin picture When the posterior probability of element is more than 0.5, this pixel belongs to the skin area of image.Certainly, according to not Same picture sample, described posterior probability threshold value can also be dynamically to adjust, and the present invention is not limited to this.

S125: if described pixel belongs to skin area, then be described pixel assignment with x, if described Pixel is not belonging to skin area, then be described pixel assignment with y, thus obtain described first two-value Image and described second bianry image.

In the step 120 of the embodiment of the present invention, make that (x, y)=(255,0), i.e. with 255 as skin pixels Point assignment, with 0 for non-skin pixel assignment, then respectively obtains described the under HSV histogram model Described second bianry image under one bianry image and mixed Gauss model.

Step 130: described first bianry image and described second bianry image are carried out step-by-step and computing from And obtain comprehensive bianry image；

Concrete, step-by-step with the operating principle of computing is, two numerals of same position are all 1, then transport Calculating result is 1；If having one is not 1, then operation result is 0.In embodiments of the present invention, for same One pixel, if by with described HSV histogram model and the matching result of described mixed Gauss model All judge that described pixel belongs to skin area, then the result of step-by-step operation is that described pixel belongs to skin Pixel；If the matching result of described HSV histogram model and described mixed Gauss model is inconsistent, then press The result of bit manipulation is that described pixel belongs to non-skin pixel.

Use step-by-step and comprehensive two testing results of computing, obtain the result detected more accurately, reduce by mistake The probability of detection.

Step 140: be filtered described comprehensive bianry image obtaining the bianry image after optimizing；

In the embodiment of the present invention, medium filtering is used described comprehensive bianry image to be carried out denoising, in order to go Except in the image of binaryzation some scattered pixels thus improve the efficiency of follow-up searching connected region.

Medium filtering is the most ripe algorithm, and it can eliminate the noise of image, and its ultimate principle is target In image, the pixel value of certain position depends on original image same location and neighbouring pixel value thereof, the most former The pixel of certain position of image and near have 9 pixels, then to these 9 rank-ordered pixels after, fetch bit The pixel value pixel value as target image pixel is obtained in centre.

Step 150: analyze connected region maximum in the bianry image after described optimization, by described maximum Connected region as skin area.

Connected region (Connected Component) generally refers to have same pixel value and position in image Put the image-region (Region, Blob) of adjacent foreground pixel point composition.Connected component analysis (Connected Component Analysis, Connected Component Labeling) refers to scheme Each connected region in Xiang is found out and labelling.Generally connected component analysis process to as if binaryzation after Image.

From the definition of connected region it is recognised that a connected region is by having the adjacent of same pixel value Pixel groups pixel set, therefore, it can find connected region in the picture by the two condition, right In each connected region found, give one unique mark (Label), to distinguish other connections Region.

The algorithms most in use of connected component analysis has Two-Pass (two-pass scan) method and Seed-Filling (to plant Sub-completion method).

Two-pass scan method, as its name, being through of referring to scans twice image, it is possible to by image The all connected regions existed are found out and labelling.Its main realization approach is: give every during first pass One label of individual location of pixels, may be by the collection of pixels in same connected region in scanning process Give one or more different label, it is therefore desirable to these are belonged to same connected region but there is difference The label of value merges, and i.e. records the relation of equality between them；Second time scanning will have equal pass exactly The pixel of equal_labels institute labelling of system is classified as a connected region and to give an identical label (logical Often this label is the minima in equal_labels).

Seed filling method derives from computer graphics, is usually used in being filled with certain figure.It is main The realization approach is wanted to be: to choose a foreground pixel point as seed, then according to two bases of connected region The foreground pixel adjacent with seed is merged into same picture by this condition (pixel value is identical, position is adjacent) In element set, this collection of pixels finally obtained is then a connected region.

In connected region, pixel neighbouring relations mainly have 4 neighborhoods, 8 neighborhoods, use 4 in the embodiment of the present invention Connected region maximum in bianry image after optimizing described in neighbor analysis.

Step 160: use whether the k nearest neighbor grader of training in advance judges the connected region of described maximum For hand, thus realize the identification of gesture.

K nearest neighbor grader is the most ripe a kind of grader, and its principle is, if nearest with certain data M data in, the data amount check of i-th class occupies the majority, then these data belong to i-th class, wherein Data be usually a vector, the feature of class can be represented.

Training in advance k nearest neighbor grader it is crucial that extract the feature of samples pictures, and according to these features Samples pictures is divided into different classes.The embodiment of the present invention have selected following four features:

Feature 1: connected region girth square and area ratio；

Feature 2: the area of connected region；

Feature 3: the connected region pixel obtained by GMM (mixed Gauss model) belongs to skin area Mathematical expectation of probability；

Feature 4: it is equal that the connected region pixel obtained by HSV histogram model belongs to the probability that skin goes Value；

Wherein, the HSV Nogata that feature 3 and feature 4 are good by calling embodiment of the present invention training in advance Graph model and GMM mixed Gauss model calculate, and do not repeat.

In the embodiment of the present invention, the k nearest neighbor grader of training in advance, by using a number of hand With the picture sample of non-hand the feature 1 of the connected region calculating its maximum～feature 4, obtain region in one's hands Sample with non-hands region.For a connected graph to be detected, extract features described above 1～feature 4, base Statistical result in these samples can determine whether whether to comprise in described connected graph staff region.

It implements and can be so that and judge in connected graph to be detected feature 1～4 and k nearest neighbor one by one The likelihood of the feature 1～4 in grader, and one rational threshold value is set for described likelihood, described When likelihood is more than described threshold value, it is judged that described connected graph to be detected comprises staff region.

In this enforcement, by image to be detected being carried out based on HSV rectangular histogram and GMM model inspection Survey, the skin pixels in the image described to be detected of identification；Further, examined by two kinds of different models The comprehensive computing of survey mode and filtering have obtained the bianry image after the optimization that described image to be detected is corresponding； By analysis and the judgement of K neighborhood classification device in largest connected region, achieve the knowledge of hand accurately Not, speed is fast and efficiently solves the error detection of hand in prior art, thus indirectly improves man-machine friendship The efficiency of gesture identification in mutually.

Embodiment two

Fig. 2 is the techniqueflow chart of the embodiment of the present invention two, in conjunction with Fig. 2, embodiment of the present invention one base In the staff detection method of the colour of skin, the training of HSV histogram model is mainly realized by following step:

Step 210: sample image is carried out skin area and the labelling in non-skin region, obtains skin picture Element sample and non-skin pixel samples；

The mark mode of sample can be by being accomplished manually to ensure the high degree of accuracy of sample.

Step 220: by described skin pixels sample and described non-skin pixel samples from RGB color It is transformed into hsv color space to obtain skin HSV pixel samples and non-skin HSV pixel samples；

Formula and technique effect thereof is implemented such as from what RGB color was transformed into hsv color space Shown in the step 110 of embodiment one, here is omitted.

Step 230: add up the HSV value of described skin HSV pixel samples, and according to described skin HSV The HSV histogram model of skin pixels is set up in the distribution of the HSV value of pixel samples；

In this step, the pixel to dermatological specimens, add up its H-number (tone), S value respectively (full And degree), the frequency distribution of V-value (brightness), thus set up the HSV histogram model of skin pixels, Meanwhile the pixel for non-skin sample performs same operation.

It should be noted that the core of the present invention is, the gray level of described HSV histogram model is pressed The statistics with histogram effect obtaining optimizing it is compressed according to default proportionate relationship.

H, S and V passage is respectively arranged with 256 gray levels, if using all of gray level, histogrammic A length of 2²⁴, about 16,000,000, the statistics effect that this cannot obtain when sample size is insufficient to big Really.Therefore, rectangular histogram length is compressed by the embodiment of the present invention, and the ratio of its compression can basis Experience selects.In the present embodiment, according to the ratio of 4:2:1 by 64 gray levels of H channel compressions, By 32 gray levels of channel S boil down to, it is 16 gray levels by V channel compressions, straight after compression Side's figure a length of 2¹⁵, i.e. 65536.Tri-passages of HSV employ the gray level of varying number, because Tri-passages of HSV are different by the influence degree of intensity of illumination, and H (colourity) passage is not by illumination variation shadow Ring, V passage be proportional to intensity of illumination change, channel S by illumination influence degree therebetween.

By the compression to Gray Histogram level, even if high precision also can be carried out in the case of a small amount of sample The Face Detection of rate.

Step 240: add up the HSV value of described non-skin HSV pixel samples, and according to described non-skin The HSV histogram model of non-skin pixel is set up in the distribution of the HSV value of skin HSV pixel samples.

Non-skin pixel samples is set up the execution process of HSV histogram model and technique effect with above-mentioned Step 230, does not repeats.It should be noted that step 230 and step 240 is actual there is no elder generation Rear order, the embodiment of the present invention is not intended to.

In the present embodiment, by dermatological specimens and the training of non-skin sample and HSV Gray Histogram The compression of level establishes the HSV histogram model of skin pixels and non-skin pixel respectively, though training sample This negligible amounts, also can greatly reduce the false drop rate of skin pixels.

Embodiment three

Fig. 3 is the techniqueflow chart of the embodiment of the present invention three, in conjunction with Fig. 2, embodiment of the present invention one base In the staff detection method of the colour of skin, the foundation of mixed Gauss model (GMM) mainly includes following step Rapid:

Step 310: skin pixel regions and non-skin pixel region to RGB samples pictures are marked, Obtain skin pixels sample and non-skin pixel samples；

In the embodiment of the present invention, first RGB samples pictures is marked, can be artificial, in order to Distinguish the skin area in picture and non-skin region, i.e. obtain skin pixels sample and non-skin pixel Sample.In advance sample is classified, be favorably improved follow-up EM algorithm and calculating mixed Gauss model The degree of closeness of the efficiency of parameter and parameter and realistic model.

Step 320: described skin pixels sample and non-skin pixel samples are changed by RGB color To r-g color space；

Identical with described in embodiment one of conversion regime in this step, uses equation below:

$r=\frac{R}{R+G+B}$

$g=\frac{G}{R+G+B}$

B=1-g-r

Wherein, R be the red value of described pixel, G be the green value of described pixel, B be described The blue valve of pixel；R, g, b are respectively the color value that after converting, described pixel is corresponding.

Step 330: use expectation-maximization algorithm, according to the described skin pixels after color space conversion Sample and non-skin pixel samples calculate described skin pixels mixed Gauss model and described non-skin respectively The parameter of pixel mixed Gauss model, wherein, described parameter includes a_k、S_kAnd π_k。

Mixed Gauss model is the superposition of multiple single Gauss model, in mixed Gauss model, each single high The weight of this model differs, i.e. data in mixed Gauss model are to generate from several single Gauss models 's.Number K of single Gauss model needs to pre-set, π_kIt it is i.e. the weight of each single Gauss model.

In statistical computation, it is desirable to maximizing (EM) algorithm is in probability (probabilistic) model Finding parameter maximal possibility estimation or the algorithm of MAP estimation, wherein depend on cannot for probabilistic model The hidden variable (Latent Variable) of observation.When having part shortage of data or cannot observe, EM algorithm provides an efficient iterative program for calculating the maximal possibility estimation of these data.? Every single-step iteration is divided into two steps: expects (Expectation) step and maximizes (Maximization) Step, because of referred to herein as EM algorithm.EM algorithm is highly developed algorithm and derivation complexity, this Inventive embodiments is not described further.

Step 340: set up mixed Gauss model according to mixed Gauss model formula.

According to the skin pixels sample after labelling, in conjunction with EM algorithm, skin mixed Gaussian can be calculated The mean vector a of model_k1, covariance matrix S_k1And weight π that multiple single Gauss model is the most corresponding_k1, Substituting into mixed Gauss model formula, can obtain skin mixed Gauss model is:

$p_{skin}(x;a_{k1},S_{k1},{\mathrm{\π}}_{k1})=\underset{k1=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k1}{p}_{k1}\left(x\right),{\mathrm{\π}}_{k1}\≥0,\underset{k1=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k1}=1$

According to the non-skin pixel samples after labelling, in conjunction with EM algorithm, non-skin mixing can be calculated The mean vector a of Gauss model_k2, covariance matrix S_k2And the power that multiple single Gauss model is the most corresponding Weight π_k2, the non-skin mixed Gauss model obtained is:

$p_{non-skin}(x;a_{k2},S_{k2},{\mathrm{\π}}_{k2})=\underset{k2=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k2}{p}_{k2}\left(x\right),{\mathrm{\π}}_{k2}\≥0,\underset{k2=1}{\overset{m}{\Σ}}{\mathrm{\π}}_{k2}=1$

When reading the new picture to be detected of a width, after color notation conversion space, read described mapping to be checked Described pixel is also substituted into above-mentioned two model by each pixel of sheet, calculates described pixel respectively p_skinAnd p_non-skin。

In the present embodiment, by the skin area of a small amount of samples pictures is marked with non-skin region, It is aided with EM algorithm and sets up the mixed Gauss model of skin pixels and non-skin pixel, with base in prior art Carry out Face Detection in rectangular histogram to compare, be not required to a large amount of training sample, save disappearing of various resource Consumption, improves the efficiency of Face Detection.

It should be noted that in the embodiment of the present invention, the foundation of HSV histogram model and mixed Gaussian mould The foundation of type there is no between any one model in sequencing, picture to be detected and above-mentioned two model Matching process is also without sequencing.The layout of each embodiment of the present invention is only for illustrating each of two models From setting up process, the order of the use of the order that it is set up is not done any restriction.

Embodiment four

Fig. 4 is the techniqueflow chart of the embodiment of the present invention 4, and in conjunction with Fig. 4, the present invention is a kind of based on the colour of skin Staff detection method mainly include following big module: image conversion module 410, binary map obtain Delivery block 420, digitwise operation module 430, filtration module 440, connected region judge module 450, mould Type training module 460.

Described image conversion module 410, for turning the image to be detected got from RGB color Change to hsv color space to obtain HSV image, and by described image to be detected from RGB color It is transformed into r-g color space to obtain r-g image；

Described binary map acquisition module 420, for traveling through each pixel read in described HSV image, And call HSV histogram model that described model training module 460 pre-builds by described HSV image It is converted into the first bianry image, and travels through each pixel read in described r-g image, call described mould Described r-g image is converted into the second bianry image by the mixed Gauss model that type training module pre-builds；

Described digitwise operation module 430, for described first bianry image and described second bianry image Carry out step-by-step and with computing thus obtain comprehensive bianry image；

Described filtration module 440, after being filtered described comprehensive bianry image to obtain optimization Bianry image；

Described connected region judge module 450, maximum in the bianry image after analyzing described optimization Connected region, using the connected region of described maximum as skin area.

Specifically, described model training module 460 is used for:

Sample image is carried out skin area and the labelling in non-skin region, obtains skin pixels sample and non- Skin pixels sample；

Call described image conversion module 410 by described skin pixels sample and described non-skin pixel samples It is transformed into hsv color space to obtain skin HSV pixel samples and non-skin from RGB color HSV pixel samples；

Add up the HSV value of described skin HSV pixel samples, and according to described skin HSV pixel samples The distribution of HSV value set up the HSV histogram model of skin pixels；

Add up the HSV value of described non-skin HSV pixel samples, and according to described non-skin HSV pixel The HSV histogram model of non-skin pixel is set up in the distribution of the HSV value of sample；

Specifically, described model training module 460 is additionally operable to:

Call described image conversion module 410 by described skin pixels sample and non-skin pixel samples by RGB color is transformed into r-g color space and obtains r-g skin pixels sample and r-g non-skin pixel sample This；

Use expectation-maximization algorithm, according to described r-g skin pixels sample and described r-g non-skin pixel Sample calculates described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model respectively Parameter thus set up described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model, Wherein, described parameter includes the mean vector of each Gauss model, covariance matrix in mixed Gauss model And weight.

Specifically, described binary map acquisition module 420, it is further used for:

Read the HSV value of described pixel, calculate described HSV value respectively with the HSV of described skin pixels The matching probability value of the HSV histogram model of histogram model and described non-skin pixel, and according to institute State matching degree value and judge whether described pixel belongs to skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel Belong to skin area, be then described pixel assignment with y, thus obtain described first bianry image；

Described binary map acquisition module 420, is additionally operable to further:

Calculate the described pixel the first probability density under skin mixed Gauss model and described pixel The second probability density under non-skin mixed Gauss model；

Described first probability density and described second probability density according to described pixel calculate described pixel Point belongs to the posterior probability of skin area；

When judging described posterior probability more than the posterior probability threshold value preset, described pixel is belonged to Skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel Be not belonging to skin area, be then described pixel assignment with y, thus obtain described first bianry image and Described second bianry image.

Specifically, connected region judge module 450, it is further used for:

The k nearest neighbor grader using training in advance judges whether the connected region of described maximum is hand, from And realize the identification of gesture.

The execution process of the embodiment that Fig. 4 is corresponding and the technique effect enforcement corresponding with Fig. 1, Fig. 2, Fig. 3 Example is identical, and here is omitted.

Device embodiment described above is only schematically, wherein said illustrates as separating component Unit can be or may not be physically separate, the parts shown as unit can be or Person may not be physical location, i.e. may be located at a place, or can also be distributed to multiple network On unit.Some or all of module therein can be selected according to the actual needs to realize the present embodiment The purpose of scheme.Those of ordinary skill in the art are not in the case of paying performing creative labour, the most permissible Understand and implement.

Through the above description of the embodiments, those skilled in the art is it can be understood that arrive each reality The mode of executing can add the mode of required general hardware platform by software and realize, naturally it is also possible to by firmly Part.Based on such understanding, the portion that prior art is contributed by technique scheme the most in other words Dividing and can embody with the form of software product, this computer software product can be stored in computer can Read in storage medium, such as ROM/RAM, magnetic disc, CD etc., including some instructions with so that one Computer installation (can be personal computer, server, or network equipment etc.) performs each to be implemented The method described in some part of example or embodiment.

Last it is noted that above example is only in order to illustrate technical scheme, rather than to it Limit；Although the present invention being described in detail with reference to previous embodiment, the ordinary skill of this area Personnel it is understood that the technical scheme described in foregoing embodiments still can be modified by it, or Person carries out equivalent to wherein portion of techniques feature；And these amendments or replacement, do not make corresponding skill The essence of art scheme departs from the spirit and scope of various embodiments of the present invention technical scheme.

Claims (10)

1. a staff detection method based on the colour of skin, it is characterised in that comprise the following steps that

The image to be detected got is transformed into hsv color space to obtain from RGB color HSV image, and described image to be detected is transformed into r-g color space to obtain from RGB color R-g image；

Traversal reads each pixel in described HSV image, and according to the HSV Nogata pre-build Described HSV image is converted into the first bianry image by graph model, and travels through in the described r-g image of reading Each pixel, is converted into the second binary map according to the mixed Gauss model pre-build by described r-g image Picture；

Described first bianry image and described second bianry image are carried out step-by-step and with computing thus obtains comprehensive Bianry image；

It is filtered described comprehensive bianry image obtaining the bianry image after optimizing；

Analyze connected region maximum in the bianry image after described optimization, by the connected region of described maximum As skin area；

The k nearest neighbor grader using training in advance judges whether the connected region of described maximum is hand, from And realize the identification of staff.

Method the most according to claim 1, it is characterised in that straight according to the HSV pre-build Described HSV image is converted into the first bianry image by side's graph model, farther includes:

Sample image is carried out skin area and the labelling in non-skin region, obtains skin pixels sample and non- Skin pixels sample；

Described skin pixels sample and described non-skin pixel samples are transformed into from RGB color Hsv color space is to obtain skin HSV pixel samples and non-skin HSV pixel samples；

Add up the HSV value of described skin HSV pixel samples, and according to described skin HSV pixel samples The distribution of HSV value set up the HSV histogram model of skin pixels；

Add up the HSV value of described non-skin HSV pixel samples, and according to described non-skin HSV pixel The HSV histogram model of non-skin pixel is set up in the distribution of the HSV value of sample.

Method the most according to claim 1 and 2, it is characterised in that according to the HSV pre-build Described HSV image is converted into the first bianry image by histogram model, farther includes:

Read the HSV value of described pixel, calculate described HSV value respectively with the HSV of described skin pixels The matching probability value of the HSV histogram model of histogram model and described non-skin pixel, and according to institute State matching degree value and judge whether described pixel belongs to skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel It is not belonging to skin area, is then described pixel assignment with y, thus obtains described first bianry image.

Method the most according to claim 1, it is characterised in that according to the mixed Gaussian pre-build Described r-g image is converted into the second bianry image by model, farther includes:

Skin pixel regions and non-skin pixel region to RGB samples pictures are marked, and obtain skin Pixel samples and non-skin pixel samples；

Described skin pixels sample and non-skin pixel samples are transformed into r-g color by RGB color Space obtains r-g skin pixels sample and r-g non-skin pixel samples；

Use expectation-maximization algorithm, according to described r-g skin pixels sample and described r-g non-skin pixel Sample calculates described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model respectively Parameter thus set up described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model, Wherein, described parameter includes the mean vector of each Gauss model, covariance matrix in mixed Gauss model And weight.

5. according to the method described in claim 1 or 4, it is characterised in that according to the mixing pre-build Described HSV image is converted into the second bianry image by Gauss model, farther includes:

Calculate the described pixel the first probability density under skin mixed Gauss model and described pixel The second probability density under non-skin mixed Gauss model；

Described first probability density and described second probability density according to described pixel calculate described pixel Point belongs to the posterior probability of skin area；

When judging described posterior probability more than the posterior probability threshold value preset, described pixel is belonged to Skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel Be not belonging to skin area, be then described pixel assignment with y, thus obtain described first bianry image and Described second bianry image.

6. a staff based on colour of skin detection device, it is characterised in that include following module:

Image conversion module, for being transformed into HSV by the image to be detected got from RGB color Color space is to obtain HSV image, and from RGB color, described image to be detected is transformed into r-g Color space is to obtain r-g image；

Binary map acquisition module, for traveling through each pixel read in described HSV image, and according to Described HSV image is converted into the first bianry image by the HSV histogram model pre-build, and travels through Read each pixel in described r-g image, according to the mixed Gauss model pre-build by described r-g Image is converted into the second bianry image；

Digitwise operation module, for carrying out step-by-step to described first bianry image and described second bianry image With computing thus obtain comprehensive bianry image；

Filtration module, for being filtered described comprehensive bianry image obtaining the bianry image after optimizing；

Connected region judge module, connected region maximum in the bianry image after analyzing described optimization, Using the connected region of described maximum as skin area；

Staff identification module, for using the k nearest neighbor grader of training in advance to judge the connection of described maximum Whether region is hand, thus realizes the identification of staff.

Device the most according to claim 6, it is characterised in that described device farther includes model Training module, described model training module is used for:

Sample image is carried out skin area and the labelling in non-skin region, obtains skin pixels sample and non- Skin pixels sample；

Described skin pixels sample and described non-skin pixel samples are transformed into from RGB color Hsv color space is to obtain skin HSV pixel samples and non-skin HSV pixel samples；

Add up the HSV value of described skin HSV pixel samples, and according to described skin HSV pixel samples The distribution of HSV value set up the HSV histogram model of skin pixels；

Add up the HSV value of described non-skin HSV pixel samples, and according to described non-skin HSV pixel The HSV histogram model of non-skin pixel is set up in the distribution of the HSV value of sample.

Device the most according to claim 6, it is characterised in that described device farther includes model Training module, described model training module is additionally operable to:

Described skin pixels sample and non-skin pixel samples are transformed into r-g color by RGB color Space obtains r-g skin pixels sample and r-g non-skin pixel samples；

Use expectation-maximization algorithm, according to described r-g skin pixels sample and described r-g non-skin pixel Sample calculates described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model respectively Parameter thus set up described skin pixels mixed Gauss model and described non-skin pixel mixed Gauss model, Wherein, described parameter includes the mean vector of each Gauss model, covariance matrix in mixed Gauss model And weight.

9. according to the device described in claim 6 or 7, it is characterised in that described binary map acquisition module, It is further used for:

Read the HSV value of described pixel, calculate described HSV value respectively with the HSV of described skin pixels The matching probability value of the HSV histogram model of histogram model and described non-skin pixel, and according to institute State matching degree value and judge whether described pixel belongs to skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel It is not belonging to skin area, is then described pixel assignment with y, thus obtains described first bianry image.

10. according to the device described in claim 6 or 8, it is characterised in that described binary map obtains mould Block, described binary map acquisition module, it is additionally operable to further:

Calculate the described pixel the first probability density under skin mixed Gauss model and described pixel The second probability density under non-skin mixed Gauss model；

Described first probability density and described second probability density according to described pixel calculate described pixel Point belongs to the posterior probability of skin area；

When judging described posterior probability more than the posterior probability threshold value preset, described pixel is belonged to Skin area；

If described pixel belongs to skin area, then it is described pixel assignment with x, if described pixel Be not belonging to skin area, be then described pixel assignment with y, thus obtain described first bianry image and Described second bianry image.