CN1975762A

CN1975762A - Skin detecting method

Info

Publication number: CN1975762A
Application number: CN 200610155132
Authority: CN
Inventors: 姚敏; 江志伟; 吴朝晖
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2006-12-11
Filing date: 2006-12-11
Publication date: 2007-06-06

Abstract

The invention discloses a skin detection method. On the basis of the histogram statistical method based on the skin pixel color, texture and space information are introduced, and the texture is used to filter out areas with colors close to human skin and obvious texture, thereby reducing false detection At the same time, the watershed segmentation algorithm controlled by markers is used to improve the positive detection rate, which overcomes the shortcomings of the existing statistical methods based on color histograms and improves the accuracy of skin detection.

Description

A kind of skin detection method

技术领域technical field

本发明涉及计算机图像处理技术领域，特别是指一种综合颜色、纹理和空间信息的皮肤检测方法。The invention relates to the technical field of computer image processing, in particular to a skin detection method integrating color, texture and spatial information.

背景技术Background technique

随着互连网发展和普及，人们可以在网络上获得各种信息，怎样防止青少年接触到色情内容是一个严肃的社会问题。由此产生一些网络过滤技术，色情图像的过滤是其中重要的一种。With the development and popularization of the Internet, people can obtain all kinds of information on the Internet, how to prevent teenagers from being exposed to pornographic content is a serious social problem. From this, some network filtering technologies are produced, and the filtering of pornographic images is one of the important ones.

基于图像内容的过滤技术通常有两种方法，第一种方法的代表是Forsyth等人的裸体检测系统，它首先检测出人体的皮肤区域，在此基础上识别出人的肢体部分，然后根据一定的规则将肢体进行组建，进而识别出裸体图像。第二种方法的代表是James Ze Wang等人的WIPE系统，它先利用皮肤和纹理过滤器将图像过滤，然后提取出形状和边缘等特征向量进行图像分类识别。There are usually two methods for filtering technology based on image content. The representative of the first method is the nude detection system of Forsyth et al., which first detects the skin area of the human body, and then identifies the body parts of the human body on this basis, and then according to a certain The rules of the body are assembled to identify nude images. The representative of the second method is the WIPE system of James Ze Wang et al. It uses skin and texture filters to filter the image, and then extracts feature vectors such as shape and edge for image classification and recognition.

在许多涉及皮肤检测的计算机视觉系统中，颜色是一个重要的特征，这些系统通常包含一个颜色过滤器，这个过滤器依靠颜色识别出皮肤区域。专利申请号为200410042877.3，发明名称为一种色情图像检测方法的专利申请，以及专利申请号为200510048577.0，发明名称为基于内容的网络色情图像和不良图像检测系统的专利申请，都使用了皮肤检测方法。在许多敏感图像检测方法中，皮肤检测是重要的先行步骤，它的准确性直接影响到后续处理的精度，研究如何提高皮肤检测精度具有重要的意义。Color is an important feature in many computer vision systems involving skin detection, and these systems usually include a color filter that identifies skin regions by color. The patent application number is 200410042877.3, the invention name is a patent application for a pornographic image detection method, and the patent application number is 200510048577.0, the invention name is a content-based network pornographic image and bad image detection system patent application, both use the skin detection method . In many sensitive image detection methods, skin detection is an important first step, and its accuracy directly affects the accuracy of subsequent processing. It is of great significance to study how to improve the accuracy of skin detection.

皮肤检测方法有很多，比较典型的有：直方图统计方法、高斯混合模型方法和基于SOM的方法，根据Vladimir Vezhnevets等人的实验比较发现，由M.J.Jones和J.M.Rehg提出的基于皮肤像素颜色的直方图统计方法(SPM，Skin Probability Map)具有最好的性能。虽然不同的人种，不同环境，不同的光照，使得皮肤颜色有很大的多样性和变化性，根据M.J.Jones和J.M.Rehg的研究，给出大量的预先标记好皮肤和非皮肤区域的图像数据集，依然可以估计出皮肤和非皮肤颜色在颜色空间中的概率。由此可以计算出像素颜色属于皮肤的似然比，将似然比大于预先设定好的阈值的像素点归入皮肤区域。这便是皮肤检测的SPM方法，该方法具有性能好，速度快的特点。它是许多皮肤检测系统采用的方法。然而SPM方法的错检率在许多实际应用中是无法令人满意的，尤其是当自然图像中包含有许多在颜色上接近人类皮肤的内容(比如沙漠，火焰，花朵以及黄色的皮毛等)以及存在过度曝光和曝光不足情况时，SPM方法会错误地把这些颜色接近人类皮肤的像素识别成皮肤区域，出现错检率高的情况。这是因为SPM方法仅仅利用了像素级上的颜色信息。There are many skin detection methods, typical ones are: histogram statistical method, Gaussian mixture model method and SOM-based method. According to the experimental comparison of Vladimir Vezhnevets et al., the histogram based on skin pixel color proposed by M.J.Jones and J.M.Rehg The graph statistical method (SPM, Skin Probability Map) has the best performance. Although different races, different environments, and different lighting make the skin color have great diversity and variability, according to the research of M.J.Jones and J.M.Rehg, a large amount of image data of pre-marked skin and non-skin areas is given set, the probability of skin and non-skin colors in the color space can still be estimated. From this, the likelihood ratio of the pixel color belonging to the skin can be calculated, and the pixels whose likelihood ratio is greater than the preset threshold are classified into the skin area. This is the SPM method of skin detection, which has the characteristics of good performance and fast speed. It is the method employed by many skin detection systems. However, the false detection rate of the SPM method is unsatisfactory in many practical applications, especially when the natural image contains a lot of content close to human skin in color (such as deserts, flames, flowers and yellow fur, etc.) and When there are overexposure and underexposure, the SPM method will mistakenly identify these pixels with colors close to human skin as skin areas, resulting in a high false detection rate. This is because the SPM method only utilizes the color information on the pixel level.

分水岭分割算法(watershed)是一种借鉴了形态学理论的分割方法，在该方法中，将一幅图象看成为一个拓扑地形图，其中灰度值被认为是地形高度值，高灰度值对应着山峰，低灰度值处对应着山谷。将水从任一处流下，它会朝地势底的地方流动，直到某一局部低洼处才停下来，这个低洼处被称为聚水盆地(catchment basin)，最终所有的水会分聚在不同的聚水盆地，聚水盆地之间的山脊被称为分水岭(watershed)。将这种想法应用于图像分割，就是要在灰度图像中找出不同的聚水盆地和分水岭，由这些不同的聚水盆地和分水岭组成的区域即为我们要分割的目标。在分水岭分割算法实现中，我们可以认为水从各个聚水盆地由底往高上涨，随着水位的升高，各个聚水盆地的水平面区域不断扩展直到水位上升到分水岭的位置。它可以充分利用各种先验知识来设置局部最小。将没有通过颜色过滤器和没有通过纹理过滤器且具有较高纹理值的像素作为非皮肤区域极小标记，将通过颜色和纹理过滤器的像素点作为皮肤区域的局部极小标记，分割后得到许多封闭的区域，然后检查每个封闭的区域纹理值的平均值和标准方差，由此得出最后的皮肤区域。Watershed segmentation algorithm (watershed) is a segmentation method that draws lessons from morphological theory. In this method, an image is regarded as a topological topographic map, in which the gray value is considered as the terrain height value, and the high gray value Corresponds to the mountain peaks, and the low gray value corresponds to the valleys. Let the water flow down from any place, it will flow towards the bottom of the terrain, and it will not stop until a certain local low-lying place. This low-lying place is called a catchment basin, and finally all the water will gather in different places. The ridges between the watersheds are called watersheds. Applying this idea to image segmentation is to find different watershed basins and watersheds in the grayscale image. The area composed of these different watershed basins and watersheds is the target we want to segment. In the implementation of the watershed segmentation algorithm, we can think that the water rises from the bottom to the top of each watershed basin. As the water level rises, the water level area of each watershed basin continues to expand until the water level rises to the watershed position. It can make full use of various prior knowledge to set the local minimum. The pixels that do not pass the color filter and the texture filter and have a higher texture value are used as the minimum mark of the non-skin area, and the pixels that pass the color and texture filter are used as the local minimum mark of the skin area. After segmentation, we get A number of closed regions are then examined for the mean and standard deviation of the texture values for each closed region to derive the final skin region.

它有4个主要的特点：It has 4 main features:

(1)分割结果是一些标记好的封闭的区域，这样就避免了其它分割算法分割后额外需要的产生封闭区域的操作；(1) The segmentation result is some marked closed areas, which avoids the additional operation of generating closed areas after segmentation by other segmentation algorithms;

(2)分水岭总是对应于图像象中真实的边界；(2) The watershed always corresponds to the real boundary in the image;

(3)详尽分割；(3) detailed division;

(4)可以方便地利用各种先验知识来设置局部极小点标记(regionalminima)。(4) Various prior knowledge can be conveniently used to set the local minimum mark (regional minima).

纹理是图像的重要视觉信息，人眼可以轻松地捕获到纹理特征，然而，到目前为止还缺乏一种可以描述各种纹理的数学定义。Gabor滤波器可被看作是方向和尺度可调的边和线的检测器，这种微观统计属性使得它可以较好地刻画图像的纹理特征。Gabor函数是一个被复正弦函数调制的高斯函数，它是能够取得空域和频域联合测不准原理下限的唯一函数。此外，Gabor函数具有很好的方向选择性。以Gabor函数作为母函数，通过伸缩和旋转可获得Gabor小波。Gabor小波变换在分析数字图像中局部区域的频率和方向信息具有优异的性能，基于Gabor小波的纹理特征受方向和光照影响小，是一种有力的纹理特征提取工具，在计算机视觉和纹理分割中得到了广泛的应用。根据B.S.Manjunath等人的实验结果，利用纹理检索图象，Gabor小波变换比金字塔结构的小波变换(PWT)，树结构的小波变换(TWT)及多分辨率联立自回归模型(MRSAR)的效果要好。通常用Gabor小波变换系数的模的平均值和其标准方差来表示图像目标的纹理特征。我们在Gabor小波变换中，取三个尺度和四个方向，在每个像素点可以得到12个刻画纹理的变换系数，把各个变换系数的平方和的平方根作为该像素点的纹理特征值。我们把纹理特征值构成的图像称为纹理图，实验表明，纹理图可以非常好地表示图像的纹理特征。纹理是一种反映了图像中的一定区域的空间分布属性，从某个孤立的像素点来谈纹理是没有意义的。但是在我们的纹理特征计算方法中，每一像素点的特征值已经包含了它周围的空间结构信息，因此，纹理图的最大特点在于我们可以得到每个像素点的纹理特征值。Texture is an important visual information of an image, and human eyes can easily capture texture features. However, until now, there is still a lack of a mathematical definition that can describe various textures. The Gabor filter can be regarded as an edge and line detector with adjustable direction and scale. This microscopic statistical property makes it better describe the texture characteristics of the image. The Gabor function is a Gaussian function modulated by a complex sine function, and it is the only function that can obtain the lower limit of the joint uncertainty principle in the space domain and the frequency domain. In addition, the Gabor function has good direction selectivity. Taking the Gabor function as the parent function, the Gabor wavelet can be obtained by stretching and rotating. Gabor wavelet transform has excellent performance in analyzing the frequency and direction information of local areas in digital images. The texture features based on Gabor wavelet are less affected by direction and illumination, and it is a powerful texture feature extraction tool. It is used in computer vision and texture segmentation. Has been widely used. According to the experimental results of B.S.Manjunath et al., using texture to retrieve images, the effect of Gabor wavelet transform on pyramid structure wavelet transform (PWT), tree structure wavelet transform (TWT) and multi-resolution simultaneous autoregressive model (MRSAR) better. Usually, the mean value and standard deviation of the modulus of the Gabor wavelet transform coefficients are used to represent the texture features of the image target. In the Gabor wavelet transform, we take three scales and four directions, and we can get 12 transformation coefficients describing the texture at each pixel point, and take the square root of the square sum of each transformation coefficient as the texture feature value of the pixel point. We call the image composed of texture feature values as texture map. Experiments show that texture map can represent the texture features of the image very well. Texture is a spatial distribution attribute that reflects a certain area in an image. It is meaningless to talk about texture from an isolated pixel point. But in our texture feature calculation method, the feature value of each pixel already contains the spatial structure information around it, so the biggest feature of the texture map is that we can get the texture feature value of each pixel.

发明内容Contents of the invention

本发明提供了一种皮肤检测方法，除颜色信息外，还引入了纹理和空间信息，克服了现有基于颜色直方图统计方法的缺点，提高皮肤检测的精度。The invention provides a skin detection method, which introduces texture and space information in addition to color information, overcomes the shortcomings of the existing statistical methods based on color histograms, and improves the precision of skin detection.

一种皮肤检测方法，依次包括以下步骤：A skin detection method, comprising the following steps in turn:

a)利用基于颜色直方图统计的方法，在图像中标记出皮肤和非皮肤像素，统计出皮肤和非皮肤颜色在颜色空间中的概率，计算出像素颜色属于皮肤的似然比，将颜色属于皮肤的似然比大于预先设定好的阈值的像素点归入皮肤区域；a) Use the method based on color histogram statistics to mark skin and non-skin pixels in the image, calculate the probability of skin and non-skin colors in the color space, calculate the likelihood ratio of the pixel color belonging to skin, and classify the color as belonging to skin The pixels whose likelihood ratio of the skin is greater than the preset threshold are classified into the skin area;

b)提取在步骤a)中被归入皮肤区域的像素点的纹理特征值，利用纹理屏蔽位图将颜色接近皮肤但纹理特征值大于预先设定好的阈值的区域过滤掉；b) extracting the texture feature value of the pixel points classified into the skin area in step a), and filtering out areas whose color is close to the skin but the texture feature value is greater than a preset threshold using the texture mask bitmap;

c)利用标记控制的分水岭分割算法扩展皮肤区域以提高正检率，对纹理图进行分割，将未通过颜色过滤的像素点和未通过纹理过滤且具有较高纹理特征值的像素点作为非皮肤区域极小标记，将通过颜色和纹理过滤的像素点作为皮肤区域的局部极小标记，分割得到若干封闭的区域，检查每个封闭区域的纹理平均值和标准方差，根据一致性原则从各封闭区域筛选出最终的皮肤区域。c) Use the marker-controlled watershed segmentation algorithm to expand the skin area to increase the positive detection rate, segment the texture map, and use the pixels that have not passed the color filter and the pixels that have not passed the texture filter and have a higher texture feature value as non-skin Regional minimum mark, use the pixels filtered by color and texture as the local minimum mark of the skin area, segment to obtain several closed areas, check the average value and standard deviation of the texture of each closed area, and select from each closed area according to the principle of consistency Regions filter out the final skin regions.

步骤b)中采用Gabor小波提取像素点对应的纹理特征值。In step b), the Gabor wavelet is used to extract the texture feature value corresponding to the pixel.

步骤b)中像素点的纹理特征值通过该像素点的Gabor变换各特征量的平方和的平方根来表征。The texture feature value of the pixel point in step b) is characterized by the square root of the sum of the squares of the Gabor transformed feature quantities of the pixel point.

本发明在基于皮肤像素颜色的直方图统计方法的基础上，还引入了纹理和空间信息，利用纹理将颜色接近人类皮肤而纹理明显的区域过滤掉，从而降低错检率；同时利用标记控制的分水岭分割算法来提高正检率，克服了现有基于颜色直方图统计方法的缺点，提高皮肤检测的精度。On the basis of the histogram statistical method based on skin pixel color, the present invention also introduces texture and space information, and uses texture to filter out areas with colors close to human skin and obvious texture, thereby reducing the false detection rate; The watershed segmentation algorithm is used to improve the positive detection rate, overcome the shortcomings of the existing statistical methods based on color histograms, and improve the accuracy of skin detection.

附图说明Description of drawings

图1为本发明皮肤检测的流程图。Fig. 1 is the flowchart of skin detection of the present invention.

具体实施方式Detailed ways

如图1所示，一种皮肤检测方法，包括以下步骤：As shown in Figure 1, a kind of skin detection method comprises the following steps:

一、利用基于颜色直方图统计的方法(SPM)检测皮肤区域1. Using the color histogram statistical method (SPM) to detect skin regions

从互连网收集了600幅包括各种人种的图像，用手工标记出皮肤和非皮肤像素，然后在RGB空间中统计出颜色rgb属于皮肤和非皮肤的概率如下式所示：Collect 600 images including various races from the Internet, manually mark the skin and non-skin pixels, and then count the probability that the color rgb belongs to skin and non-skin in the RGB space as shown in the following formula:

$P P ((rgb rgb | | skin skin)) = = \frac{s the s [[rgb rgb]]}{{T T}_{s the s}}$

$P P ((rgb rgb | | &Not; &Not; skin skin)) = = \frac{n no [[rgb rgb]]}{{T T}_{n no}}$

其中，s[rgb]是皮肤颜色的直方图中rgb仓中的像素个数，n[rgb]是非皮肤颜色的直方图中rgb仓中的像素个数。T_s和T_n分别是皮肤和非皮肤直方图中的总像素个数。一般情况下，颜色rgb属于皮肤和非皮肤的概率如下式所示：Among them, s[rgb] is the number of pixels in the rgb bin in the histogram of the skin color, and n[rgb] is the number of pixels in the rgb bin in the histogram of the non-skin color. T _s and T _n are the total number of pixels in the skin and non-skin histograms, respectively. In general, the probability that the color rgb belongs to skin and non-skin is as follows:

$P P ((skin skin | | rgb rgb)) = = \frac{p p ((rgb rgb | | skin skin)) p p ((skin skin))}{p p ((rgb rgb))}$

$P P ((&Not; &Not; skin skin | | rgb rgb)) = = \frac{p p ((rgb rgb | | &Not; &Not; skin skin)) p p ((&Not; &Not; skin skin))}{p p ((rgb rgb))}$

颜色rgb属于皮肤的似然比L(rgb)为：The likelihood ratio L(rgb) that the color rgb belongs to the skin is:

$L L ((rgb rgb)) = = \frac{P P ((rgb rgb | | skin skin))}{P P ((rgb rgb | | &Not; &Not; skin skin))}$

当似然比L(rgb)满足L(rgb)≥θ时，具有颜色rgb的像素点被归类到皮肤区域，其中θ为阈值。When the likelihood ratio L(rgb) satisfies L(rgb)≥θ, pixels with color rgb are classified into skin regions, where θ is the threshold.

其中θ为阈值，当θ值取得小时，正检率和错检率都高；当θ值取得大时，错检率降低的同时正检率也会降低。由于我们并不单纯依赖SPM方法检测皮肤，因此一般把θ值取的比较小(θ≤0.3)使的绝大部分的皮肤像素都能通过这个阶段的检测。Among them, θ is the threshold value. When the value of θ is small, the rate of positive detection and false detection rate are both high; when the value of θ is large, the rate of false detection is reduced and the rate of correct detection is also reduced. Since we do not rely solely on the SPM method to detect skin, generally the value of θ is relatively small (θ≤0.3) so that most of the skin pixels can pass the detection at this stage.

二、利用纹理过滤降低错检率2. Use texture filtering to reduce false detection rate

Gabor滤波器可被看作是方向和尺度可调的边和线的检测器，这种微观统计属性使得它可以较好地刻画图像的纹理特征。基于Gabor小波的纹理特征受方向和光照影响小，是一种有力的纹理特征提取工具。Gabor函数是一个被复正弦函数调制的高斯函数。两维的Gabor函数如下式所示。The Gabor filter can be regarded as an edge and line detector with adjustable direction and scale. This microscopic statistical property makes it better describe the texture characteristics of the image. The texture feature based on Gabor wavelet is less affected by direction and light, and it is a powerful texture feature extraction tool. The Gabor function is a Gaussian function modulated by a complex sine function. The two-dimensional Gabor function is shown in the following formula.

$g g ((x x,, y the y)) = = \frac{11}{22 π π {σ σ}_{x x} {σ σ}_{y the y}} exp exp ((- - \frac{11}{22} ((\frac{{x x}^{22}}{{σ σ}_{x x}^{22}} + + \frac{{y the y}^{22}}{{σ σ}_{y the y}^{22}})) + + 22 πjwx πjwx))$

以g(x，y)为母函数，通过伸缩和旋转可获得Gabor小波如下式所示。Taking g(x, y) as the parent function, the Gabor wavelet can be obtained by stretching and rotating as shown in the following formula.

g_mn(x，y)＝a^-mg(x′，y′)，a＞1，m，n＝integerg _mn (x, y) = a ^-m g (x', y'), a>1, m, n=integer

x′＝a^-m(xcosθ+ysinθ)，x'=a ^-m (xcosθ+ysinθ),

y′＝a^-m(-xsinθ+ycosθ)，y'=a ^-m (-xsinθ+ycosθ),

θ＝nπ/k，k＝integerθ=nπ/k, k=integer

其中k是方向的总数，m为尺度因子。图像I(x，y)的Gabor小波变换为where k is the total number of directions and m is the scaling factor. The Gabor wavelet transform of image I(x, y) is

${W W}_{mn mn} ((x x,, y the y)) = = &Integral; &Integral; I I ((x x 11,, y the y 11)) {g g}_{mn mn}^{* *} ((x x - - x x 11,, y the y - - y the y 11)) {dx dx}_{11} {dy dy}_{11}$

其中g^*为g的复共轭。我们取三个尺度(m＝0，1，2)和四个方向(n＝0，1，2，3)，在每个像素点可以得到12个刻画纹理的特征量，把每个像素点各个特征量的平方和的平方根作为该像素点的纹理特征值，得到最后的纹理图如下式所示。where g ^* is the complex conjugate of g. We take three scales (m=0, 1, 2) and four directions (n=0, 1, 2, 3), and we can get 12 feature quantities describing the texture at each pixel point. The square root of the sum of squares of each feature quantity is used as the texture feature value of the pixel point, and the final texture map is obtained as shown in the following formula.

$T T ((x x,, y the y)) = = {(({Σ Σ}_{m m = = 00}^{22} {Σ Σ}_{n no = = 00}^{33} {W W}_{mn mn}^{22} ((x x,, y the y))))}^{11 / / 22}$

首先用公式I(x，y)＝0.3R(x，y)+0.59G(x，y)+0.11B(x，y)把彩色图像转换为灰度图像，然后计算出纹理图，按下式将纹理图二值化得到纹理屏蔽位图。First use the formula I (x, y) = 0.3R (x, y) + 0.59G (x, y) + 0.11B (x, y) to convert the color image into a grayscale image, and then calculate the texture map, press Binarize the texture map to obtain the texture mask bitmap.

$M m ((x x,, y the y)) = = \{\begin{matrix} 11,, & T T ((x x,, y the y)) \leq \leq {θ θ}_{T T} \\ 00,, & T T ((x x,, y the y)) > > {θ θ}_{T T} \end{matrix}$

其中θ_T为纹理阈值。由于人的皮肤是比较光滑的，其纹理特征值也相对较小，因此利用纹理屏蔽位图可以将颜色接近人类皮肤而纹理明显的区域过滤掉，从而降低错检率。where θ _T is the texture threshold. Since human skin is relatively smooth, its texture feature value is relatively small, so using the texture mask bitmap can filter out areas with colors close to human skin and obvious texture, thereby reducing the false detection rate.

三、利用标记控制的分水岭分割算法提高正检率。3. Using the marker-controlled watershed segmentation algorithm to improve the positive detection rate.

经过颜色和纹理过滤后，可以将已有的皮肤区域作为聚水盆地的最低点，利用分水岭算法来扩展皮肤区域。它就是标记控制的分水岭分割算法。可以利用在颜色过滤和纹理过滤阶段的知识来设置标记，具体操作如下：After color and texture filtering, the existing skin area can be used as the lowest point of the water basin, and the watershed algorithm can be used to expand the skin area. It is the marker-controlled watershed segmentation algorithm. Markers can be set using knowledge from the color filtering and texture filtering stages as follows:

(1)未通过颜色过滤器的像素作为非皮肤区域极小标记(1) Pixels that do not pass the color filter are regarded as extremely small non-skin area marks

(2)未通过纹理过滤器的且具有较高纹理值的像素作为非皮肤区域极小标记(2) Pixels that do not pass the texture filter and have a higher texture value are used as the minimum mark of the non-skin area

(3)通过颜色和纹理过滤器的像素作为皮肤区域极小标记(3) Pixels passed through the color and texture filters serve as minimal markers for skin regions

分割是在纹理图上进行的，分割完成以后，我们计算出每一个皮肤扩展区域的纹理平均值μ_i和标准方差σ_i。一致性原则用下式表示：The segmentation is carried out on the texture map. After the segmentation is completed, we calculate the average value μ _i and standard deviation σ _i of the texture of each skin extension area. The consistency principle is expressed by the following formula:

0.9μ＜μ_i＜1.1μ0.9μ<μ _i <1.1μ

0.9σ＜σ_i＜1.1σ0.9σ<σ _i <1.1σ

满足一致性原则的区域判断为皮肤区域。其中μ和σ是在作颜色直方图统计的同时在人工标记为皮肤像素的区域上计算出来的。Areas satisfying the principle of consistency are judged as skin areas. Among them, μ and σ are calculated on the area artificially marked as skin pixels while making color histogram statistics.

需要提出的是，分割也可以在灰度图或三个颜色通道上进行，判断皮肤区域的一致性原则也可以采用灰度或颜色的均值与方差，这些变化都在本发明的专利要求范围内。What needs to be pointed out is that the segmentation can also be carried out on the grayscale or three color channels, and the principle of judging the consistency of the skin area can also use the mean and variance of grayscale or color, and these changes are all within the scope of the patent requirements of the present invention .

从互联网上收集了300幅各种类型的图像用于测试，这300幅图像中，100幅包含有人；100幅包含有在颜色上接近人的皮肤颜色的内容；另外100幅是任意的图像，保证了测试图像具有很好的多样性和代表性。实验结果表明，和SPM方法比较，本发明的方法可以将正检率由92.7％提高到95.2％，错检率由20.1％下降至4.3％。300 images of various types were collected from the Internet for testing. Of these 300 images, 100 contained people; 100 contained content that was close to human skin color in color; the other 100 were arbitrary images. It is guaranteed that the test images have good diversity and representativeness. Experimental results show that, compared with the SPM method, the method of the present invention can increase the correct detection rate from 92.7% to 95.2%, and the false detection rate can be reduced from 20.1% to 4.3%.

Claims

1. A skin detection method, characterized in that: comprises the following steps successively:

a) Use the method based on color histogram statistics to mark skin and non-skin pixels in the image, calculate the probability of skin and non-skin colors in the color space, calculate the likelihood ratio of the pixel color belonging to skin, and classify the color as belonging to skin The pixels whose likelihood ratio of the skin is greater than the preset threshold are classified into the skin area;

b) extracting the texture feature value of the pixel points classified into the skin area in step a), and filtering out areas whose color is close to the skin but the texture feature value is greater than a preset threshold using the texture mask bitmap;

c) Use the marker-controlled watershed segmentation algorithm to expand the skin area to increase the positive detection rate, segment the texture map, and use the pixels that have not passed the color filter and the pixels that have not passed the texture filter and have a higher texture feature value as non-skin Regional minimum mark, use the pixels filtered by color and texture as the local minimum mark of the skin area, segment to obtain several closed areas, check the average value and standard deviation of the texture of each closed area, and select from each closed area according to the principle of consistency Regions filter out the final skin regions.

2. The skin detection method according to claim 1, further characterized in that: in the step b), Gabor wavelet is used to extract the texture feature value corresponding to the pixel point.

3. skin detection method as claimed in claim 1, is characterized in that: the texture feature value of pixel point is characterized by the square root of the square sum of each feature quantity of the Gabor wavelet transform of this pixel point in the step b).