CN102103690A - Method for automatically portioning hair area - Google Patents

Abstract

The invention provides a method for automatically portioning a hair area, relating to the technical field of a hair partitioning method and belonging to the field of computer vision. The method comprises the following steps of: step 1, detecting the face: using a face detection module for detecting the face position from one input face picture by using a trained cascade classifier; step 2, marking target background marks: using a target background marking module to line out the interested area, and finding out the most possible target mark and the background mark according to position and color characteristics; and step 3, image partitioning: according to the target background marks, using an image partitioning module to partition the hair area and output the hair area. The method provided by the invention uses the image processing technology to perform face detection, target background marking, image partitioning and the like to an input image, and output the hear area in the image. The method can accurately detect the hair area under complicated background conditions, which provides good basis to the research and application of identity recognition, gender and age estimation, image retrieval and the like.

Description

An Automatic Hair Region Segmentation Method

technical field

The invention relates to an automatic hair segmentation method, which belongs to the field of computer vision.

Background technique

Hair is an important feature of the human body. Using this feature, we can implement applications such as identification, age and gender estimation, and image retrieval. Studies have shown that hair is an important feature for us to distinguish similar faces. Changes in hairstyle may make us make wrong judgments about people's identities. Using hair features can provide auxiliary information for identification. Therefore, the application of hair features in the field of computer vision is of great significance.

The shape and color of hair will be different due to differences in human gender, age, and race. People can also change their hairstyle and color at will, which brings certain difficulties to the detection, description, analysis, and application of hair features. However, the color and shape of hair are always fixed in a certain period of time, which makes it possible to study the characteristics of hair.

Before describing, analyzing, and matching hair features, an important task is to find hair regions. Only after the hair region is determined, can the hair in the region be subjected to feature extraction, feature matching and other operations. At present, the main method of determining the hair area is generally based on the accurate detection of the face area, and uses the hair color, texture, shape, position and other features to perform area segmentation.

Yacoob and Davis were the first researchers to identify hair regions and describe hair characteristics. They proposed a method for skin color model matching. They first used a cascade classifier to accurately detect the position of the face and eyes, then drew three rectangular boxes under the forehead and eyes according to the relative positions, and extracted the color features in the rectangular boxes to build a skin color model. Then draw three small rectangular frames on the upper part, left part and right part of the face boundary according to the relative position. The relative positions of the three frames are determined through trial and error, so that the three rectangular frames are in the hair area as much as possible, and the pixels in the three frames are matched with the skin color model, and the skin color pixels are deleted. Based on the colors in these three boxes, a hair color feature model is established. Scan the entire image, and judge whether it is a hair pixel by comparing the distance between the RGB value of the pixel and the model. This method only uses the feature of color to judge hair pixels, and can only segment the hair in a simple and normalized background, and the accuracy rate is relatively low in the case of complex backgrounds.

Lee et al. proposed a Gaussian mixture model using hair color and location information for hair segmentation. The construction of the Gaussian mixture model is divided into two parts: offline training and online updating. For a picture, first find the most likely area according to the distribution probability of the background, face and hair, compare the points in the area with the GMM, and delete the points that do not satisfy the model. Then use the graph cut algorithm to minimize the energy function to determine some hair and face pixels. The GMM is updated according to the determined pixel points, and the updated model is used to segment until all pixels are segmented. This method uses an absolute position for the hair position information in the energy function constructed algorithmically, so it requires that the face in the test picture must be in the middle of the picture, otherwise a large number of false detections will occur. Secondly, as mentioned in the paper, the judgment of hair and dark background is prone to errors. Finally, because the multi-class graph cut algorithm based on the pixel unit has a complex calculation degree, it requires a long operation time.

C.Rousset and P.Y.Coulon proposed a method of combining frequency domain mask and color mask to segment hair. The method first uses the face box proposed by Viola and Jones and anthropometric data to define the head region. Since the hair has texture characteristics, a Gaussian band filter is used to scan the image to obtain a spectral map of the image. The hair pixels are segmented by thresholding, and the remaining pixels are background pixels. In the segmented hair area, draw a sample window above the face, extract the color in the window as hair color, establish a hair color model, and use the model to judge hair pixels. The hair pixels in the head area are used as the target markers, and the background pixels from the texture segmentation and the pixels from the texture segmentation as hair but not in the face area are used as the background markers. Then the image is segmented using the automatic segmentation method proposed by Levin et al. This method makes comprehensive use of color and texture features, but there are missed detections under complex background and lighting conditions.

In summary, some hair segmentation techniques that currently exist have certain limitations. Therefore, there is currently a need for a more efficient method for segmenting hair regions.

Contents of the invention

The purpose of the present invention is to provide a method for automatic segmentation of hair regions, using image processing technology to perform face detection, target background marking, image segmentation and other processing on an input image, and output the hair region in the image. The method can accurately detect hair regions under complex background conditions, and provides a good foundation for the research and application of identification, gender and age estimation, image retrieval and the like.

An automatic hair region segmentation method, comprising the following steps:

Step 1: Detect face: Use the face detection module to detect the position of the face through the trained cascade classifier from the input face picture;

Step 2: Mark the target background mark: Use the target background mark module to draw the area of interest on the face position, and find the most likely target mark and background mark according to the position and color characteristics;

Step 3: Image segmentation: According to the target background mark, use the image segmentation module to segment the hair area and output it.

The face detection method in step 1 includes the following steps: extracting Haar features, training a weak classifier, using the AdaBoost algorithm to select an optimized weak classifier iteratively, generating a strong classifier, and performing real-time detection to obtain the facial region R _f .

The specific process of the AdaBoost algorithm is as follows: respectively calculate the sample integral graph for the collected face sample set and non-face sample set, obtain the rectangular feature prototype, calculate the rectangular feature value, and obtain the feature set; determine the threshold value and generate it from the rectangular feature set Corresponding weak classifier, get the weak classifier set; select the optimal weak classifier, call the AdaBoost algorithm to train the strong classifier, and get the strong classifier set, then judge again whether there is a non-face picture set, if it is judged to be , add non-face samples to the non-face sample set and repeat the above steps, if the judgment is no, directly obtain the cascade classifier.

The target background marking method of step 2 comprises the following steps:

(1) Determination of the region of interest, according to the position of the face region and prior probability knowledge, determine the region of interest on the basis of the face region _Rf , the formula is as follows:

"Hair and face" area width = 3.6*face width,

"Hair and face" area height = 3.7*face height;

(2) face color feature extraction, make the probability distribution map of skin color in the face area R _f , find out the pixel point that probability distribution is greater than 1%, calculate mean value μ and covariance matrix C, obtain the Gaussian model of skin color;

(3) target pixel mark, draw a sample window area R _h that definitely includes hair on the basis of the face area R _f , wherein:

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$ ${\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$

${\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$

分别代表面部区域R_f的左右上下边界；in represent the left and right upper and lower boundaries of the hair sample window R _h respectively, and the same

Represent the left and right upper and lower boundaries of the facial region R _f respectively;

In this area, the skin color pixels are removed according to the face skin color model, the hair color features are extracted from the remaining pixels, and the hair pixels are marked as the target mark;

(4) Background pixel labeling, take an area at the center of the face area R _f and the left and right upper corners of the area of interest as the background area, and mark it as the background.

The image segmentation method of step 3 comprises:

Initial image segmentation, use the mean shift algorithm to perform initial segmentation on the image in the region of interest to obtain a small area;

Mark the target background area. If there is a background mark in an area, it will be marked as the background area. If there is a target mark in an area, it will be marked as the target area, and the remaining area will be marked as the area to be merged;

MSRM segmentation, iteratively merges the background area, the target area, and the area to be merged, and uses the principle of maximum similarity to obtain the final segmentation result after multiple iterations.

MSRM segmentation algorithm of the present invention comprises the following processes:

A. Iteratively merge the area in the background area and its adjacent area until no new area is formed, and the iteration stops;

B. Iteratively merge the area in the target area and its adjacent area until no new area is formed, and the iteration stops;

C. Iteratively merge the area in the area to be merged with its adjacent area until no new area is formed, and the iteration stops;

Repeat the above step A, step B, and step C until there is no iterative merge action in step C and stop.

The features of the present invention for implementing the above-mentioned technical method are as follows: (1) Use the position information of the human face to determine the region of interest, and find and mark the target pixel according to the color and position information. (2) When marking the target and background pixels, make sure that the marked pixels are located in the target and background. (3) Operate in the YC _r C _b color space, and use the two components of C _r C _b to calculate the similarity, which reduces the complexity of the system.

Description of drawings

Fig. 1 is four kinds of Haar feature schematic diagrams in the automatic hair region segmentation method of the present invention;

Fig. 2 is a schematic flow chart of the automatic hair region segmentation system of the present invention;

Fig. 3 is the schematic flow chart of AdaBoost training module in the automatic hair region segmentation system of the present invention;

Fig. 4 is a schematic flow chart of the target background marking module in the automatic hair region segmentation system of the present invention;

Fig. 5 is the schematic flow chart of the image segmentation module in the automatic hair region segmentation system of the present invention;

Detailed ways

As shown in Figure 2, an automatic hair region segmentation method includes the following steps:

Step 1: Detect the face, and use the trained cascade classifier to detect the position of the face from an input image.

Step 2: Mark the target background mark, use the face position to draw the region of interest, and find the most likely target mark and background mark according to the position, color and other characteristics.

Step 3: Image segmentation, according to the marks made in step 2, use image segmentation technology to segment out the hair region.

Wherein, the face detection described in step 1 comprises: the training of classifier, trains the same classifier (weak classifier) for different training sets, then the classifiers obtained on these different training sets are combined to form a The final strong classifier; the detection process, use the trained classifier to detect the face, and get the face area, which is represented by R _f .

Wherein, the classifier training process includes the following steps:

(1) Extract Haar features

Commonly used Haar features have 3 types and 4 forms, as shown in Figure 1A-Figure 1D. The three types are: 2-rectangle feature, 3-rectangle feature, 4-rectangle feature. Of course, more and more complex features can also be designed on the basis of these four features. Since there are usually nearly 10,000 training samples, and the number of rectangular features is very large, if the sum of all pixels in the rectangle is calculated every time the feature value is calculated, the speed of training and detection will be greatly reduced. Therefore, a new image representation method-integral image is introduced. The calculation of the eigenvalue of the rectangular feature is only related to the integral image of the endpoint of the feature rectangle, so regardless of the scale transformation of the feature rectangle, the calculation of the feature value consumes time is constant. In this way, the eigenvalues of all sub-windows can be obtained by traversing the image once.

(2) Generate a weak classifier

Each Haar feature corresponds to a weak classifier, and each weak classifier is defined according to the parameters of its corresponding Haar feature. Using the position information of the above-mentioned Haar features, the corresponding feature parameters can be obtained by performing statistics on the training samples. The weak classifier trained in the AdaBoost algorithm is any classifier, including decision trees, neural networks, and hidden Markov models. If the weak classifier is a linear neural network, then the AdaBoost algorithm will construct a node of the multilayer perceptron each time. .

(3) Use the AdaBoost algorithm to select an optimized weak classifier

Not any Haar feature can better describe a certain feature of the gray distribution of the face. The Adaboost algorithm is applied to face detection. The basic idea is to train the same classifier (weak classifier) for different training sets, and then combine the classifiers obtained on these different training sets to form a final strong classification. device. Different training sets in the Adaboost algorithm are realized by adjusting the weight corresponding to each sample. At the beginning, the corresponding weights of each sample are the same, for the misclassified samples of h1, increase their corresponding weights; and for the correctly classified samples, reduce their weights, so that the misclassified samples are highlighted, thus obtaining A new sample distribution U2. Under the new sample distribution, train the weak classifier again to obtain the weak classifier h2. By analogy, after T cycles, T weak classifiers are obtained, and these T weak classifiers are superimposed (boosted) according to a certain weight to obtain the final desired strong classifier.

As shown in Figure 4, wherein the target background marking method described in step 2 comprises the following steps:

(1) Determine the region of interest based on the face region.

The possible hair regions are formed by many normalized hair training images, and the head region formula is defined by the hair regions. From the clustering results of hair, the definition formula of the head area containing hair and face can be obtained as follows:

"Hair and face" area width = 3.6*face width

"Hair and face" area height = 3.7*face height

In this way, an area containing all hair and faces can be obtained, and all subsequent operations are performed in this area.

(2) Extracting face color features

The present invention selects YC _r C _b space as the mapping space of color distribution statistics. This space is less affected by brightness changes, and is two-dimensional independent distribution, which can better limit the skin color distribution area. It has the brightness component in the HS1 format. The advantages of separation, and can be obtained from the linear change of the RGB format.

In the face region R _f obtained in step 1, there are other colors besides the skin color of the face, such as part of the hair, background color, etc., but the proportion of the skin color of the face is the highest. Divide the C _r C _b components into 64 equal parts, and make a Gaussian probability distribution for the C _r C _b components of all pixels in R _f . Select pixels with a probability greater than 1%, according to the following formula:

χ=(C _r , C _b ) ^T

μ=E(x)

C＝E[(χ-μ)(x-μ) ^T ]

Obtain the mean μ and covariance matrix C of the vector χ=(C _r , C _b ) ^T composed of the chromaticity information C _r C _b of the facial skin color, and fit the two-dimensional Gaussian model.

(3) mark target mark

On the upper part of the face region _Rf , draw a hair sample window _Rh , where:

${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$ ${\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$

${\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$

分别代表人脸区域R_f的左右上下边界。in represent the left and right upper and lower boundaries of the hair sample window R _h respectively, and the same

represent the left, right, upper and lower boundaries of the face region _Rf , respectively.

Calculate the similarity between the pixels in the hair sample window R _h and the Gaussian model obtained in the extracted face color feature, where the similarity calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; \{</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span>$

χ=(C _r , C _b ) ^T

Among them, μ is the mean value obtained by extracting face color features, and C is the obtained covariance matrix.

Remove the pixels similar to facial skin color in the hair sample window _Rh , divide the _Cr C _b components into 32 groups, and make a Gaussian probability distribution for the _Cr C _b components of the remaining pixels in the hair sample window _Rh . The range with the highest probability is selected as the hair color range space, and the pixels belonging to this range are marked as target marks in the hair sample window R _h to ensure that all the marked targets are hair.

(4) mark background mark

它的长宽都是R_f的1/2，将R_b2中的所有像素标记为背景。同样，在感兴趣区域的左上角和右上角各划出一个等腰直角三角形区域，顶点在感兴趣区域的左右顶点处，边长为

分别作为区域

将该区域内的像素点标记为背景。保证被标记为背景的区域中不能含有头发像素点。Draw a region at the center of the face region R _f

Its length and width are 1/2 of R _f , marking all pixels in R _b2 as background. Similarly, draw an isosceles right-angled triangle area at the upper left corner and upper right corner of the region of interest, the vertices are at the left and right vertices of the region of interest, and the side length is

respectively as regions

Mark the pixels in this area as the background. Make sure that the area marked as background does not contain hair pixels.

Wherein, the segmentation method described in step 3 includes the following steps: as shown in Figure 5, step 1, use mean shift for initial segmentation; step 2, mark out the background area M _B and target area M _o according to the marker; step 3, Region segmentation using the improved MSRM algorithm.

Among them, step 1 uses mean shift as the initial segmentation method, mainly because this method has less over-segmentation phenomenon and can better preserve boundary information compared with other methods. Using this method, a good segmentation effect can be obtained with only a small amount of label information, and it has better robustness.

Wherein, the method for determining the background area M _B and the target area M _O in step two is mainly judged according to whether the area segmented in step one contains a mark. The region containing the background mark is called the background mark region M _B , and the region containing the target mark is called the target mark region M _O . Those regions that contain neither target markers nor background markers are called regions to be merged, denoted by N. Most of the area in the image is the area to be merged N, our task is to judge the area N to be merged, the area N to be merged is either the target or the background.

Among them, the MRSM segmentation method described in step 3 is mainly based on the principle of maximum similarity to merge the area N to be merged with the background area M _B and the target area M _O , which can be specifically divided into three stages:

对于每一个A_i∈N，我们找到它的相邻区域

的集合

很明显B

计算A_i与

中每一个元素的相似度，如果满足如下公式Phase 1: For each area _B∈MB , we find the set of all its adjacent areas A _i

For each A _i ∈ N, we find its neighbors

collection of

Obviously B

Calculate A _i with

The similarity of each element in , if it satisfies the following formula

$\mathrm{<span\; class="notranslate">\; \&rho;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>}\mathrm{<span\; class="notranslate">\; B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}\mathrm{<span\; class="notranslate">\; \&rho;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; )</span>$

同理。Among them, ρ(A _i , B) represents the similarity between A _i and B area,

the same way.

Then, A _i is merged with B region, and the merged region is still marked as B. Otherwise, A _i and B areas remain unchanged.

The above process is carried out iteratively. After one iteration, _MB and N will be updated, and then continue to iterate. When no new area is merged, the iteration stops.

Phase 2: In the same way, iteratively merge each region O∈M _o until no new region is merged.

Stage 3: After the above two stages, there are still some unmarked regions, mainly because the similarity between these unmarked regions is too high, so any region in the background region M _B and the target region M _O Neither can be merged with them. Therefore, we need to merge this part of the area, the method is the same as the first stage.

After passing through stage three, return to stage one and continue until stage three ends when no merging process occurs.

Among them, the similarity between the two regions is represented by the Bhattachary coefficient, and the calculation method is as follows:

$\mathrm{<span\; class="notranslate">\; \&rho;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 256</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\sqrt{\mathrm{<span\; class="notranslate">\; His</span>}{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; His</span>}{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; Q</span>}}^{\mathrm{<span\; class="notranslate">\; u</span>}}}$

分别是R、Q区域的归一化直方图，上标u表示第u个Hist元素。Hist是将C_r、C_b颜色通道分别量化为16个等级，这样就产生了16×16＝256个颜色特征空间，对区域内的像素在这256个特征上做颜色直方图，统计落在每一个特征上的概率，用表示。in,

They are the normalized histograms of the R and Q regions, respectively, and the superscript u represents the uth Hist element. Hist quantizes the C _r and C _b color channels into 16 levels respectively, thus generating 16×16=256 color feature spaces, and making a color histogram on these 256 features for the pixels in the area, and the statistics fall in The probability of each feature, with express.

和向量之间夹角的余弦，P、Q区域间的巴氏系数越大，说明他们之间的相似性越高。如果两个区域具有相似的内容，那么之间的相似度就很高，他们之间的巴氏系数就会很大。The Bhattachary coefficient is actually a vector

and vector The cosine of the angle between them, the larger the Barthel coefficient between the P and Q regions, the higher the similarity between them. If two areas have similar content, then the similarity between them is high, and the Barthel coefficient between them will be large.

According to another aspect of the present invention, the present invention provides a system for automatic hair region segmentation, the system includes the following parts:

The face detection module is used to detect the face area from the input picture according to the Haar feature and the AdaBoost classifier, including the AdaBoost classifier training module and the real-time detection module.

The target background marking module is used to delineate the region of interest according to the face information, and mark the target and background pixels in the region of interest according to certain position and color information.

The image segmentation module uses the improved MSRM algorithm to separate images according to the marked marks.

According to the present invention, the target background labeling module includes: the region of interest area determination module, according to the position of the face area and prior probability knowledge, determines the area of interest on the basis of the face area _Rf ; the face color feature extraction module, in Make a probability distribution map of skin color in the face area R _f , find out the pixels whose probability distribution is greater than 1%, calculate the mean value μ and covariance matrix C, and obtain the Gaussian model of skin color; the target pixel marking module, in the face area R _f Draw out a sample window region R _h that must include hair on the basis of , remove the skin color pixels according to the face skin color model in this area, extract the hair color features from the remaining pixels, and mark the hair pixels; the background pixel marking module, in The center position of the face region R _f and the left and right upper corners of the region of interest are each taken as the background region and marked as the background.

According to the present invention, the image segmentation module includes: an initial segmentation module, which uses the mean shift algorithm to initially segment the image in the region of interest; a target background area labeling module, which uses the target and background labels to mark the segmented area as background area M _B , The target area M _o and the area to be merged N; the MSRM segmentation module, including the background area merging module, the target area merging module, and the to-be-merged area merging module, get the final segmentation result after multiple iterations.

Claims (6)

1. an automatic hair region segmentation method, is characterized in that comprising the steps: Step 1: Detect faces: Use the face detection module to detect the face position through the trained cascade classifier from an input face picture; Step 2: mark the target background mark: use the target background mark module to draw the area of interest on the face position, and find the most likely target mark and background mark according to the position and color characteristics; Step 3: Image segmentation: According to the target background mark, use the image segmentation module to segment the hair area and output it. 2. according to the described automatic hair region segmentation method of right 1, it is characterized in that, the detection people's face method of the first step comprises the following steps: extract Haar feature, train weak classifier, adopt the weak classifier of AdaBoost algorithm to select optimization The generator iterates, generates a strong classifier, and performs real-time detection to obtain the face region R _f . 3. The automatic hair region segmentation method according to claim 2, characterized in that the specific process of the AdaBoost algorithm is as follows: the sample integral graph is calculated for the face sample set and non-face sample set collected respectively to obtain the rectangular feature prototype , calculate the rectangular feature value to get the feature set; determine the threshold, generate the corresponding weak classifier from the rectangular feature set, and get the weak classifier set; select the optimal weak classifier, call the AdaBoost algorithm to train the strong classifier, and get the strong classifier At this time, it is judged again whether there is a non-face image set. If the judgment is yes, add non-face samples to the non-face sample set and repeat the above steps. If the judgment is no, directly obtain the cascade classifier. 4. the automatic hair region segmentation method according to right 1, is characterized in that, the target background labeling method of described second step comprises the steps: (1) Determination of the region of interest, according to the position of the face region and prior probability knowledge, determine the region of interest on the basis of the face region _Rf , the formula is as follows: "Hair and face" area width = 3.6*face width, "Hair and face" area height = 3.7*face height; (2) Face color feature extraction, make a probability distribution map of skin color in the facial region R _f , find out the pixels whose probability distribution is greater than 1%, calculate the mean value μ and covariance matrix C, and obtain the Gaussian model of skin color; (3) Mark the target pixel, draw a sample window area R _h that definitely includes hair on the basis of the facial area R _f , wherein: ${\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$ ${\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}$ ${\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; f</span>}}}<span\; class="notranslate">\; )</span>$ in

represent the left and right upper and lower boundaries of the hair sample window R _h respectively, and the same

Represent the left and right upper and lower boundaries of the facial region R _f respectively; In this area, the skin color pixels are removed according to the face skin color model, the hair color features are extracted from the remaining pixels, and the hair pixels are marked as the target mark; (4) Mark the background pixels. Take a region at the center of the face region R _f and the left and right upper corners of the region of interest as the background region, and mark it as the background. 5. method according to right 1, is characterized in that, the image segmentation method of described 3rd step comprises: initial image segmentation, uses mean shift algorithm to do initial segmentation to image in region of interest, obtains small area; Target background Area marking, if there is a background mark in an area, it will be marked as the background area, if there is a target mark in an area, it will be marked as the target area, and the remaining area will be marked as the area to be merged; MSRM segmentation, iteratively merges the background area, the target area, and the area to be merged, and uses the principle of maximum similarity to obtain the final segmentation result after multiple iterations. 6. method according to claim 5, is characterized in that, described MSRM segmentation algorithm comprises the following process: A. Iteratively merge the area in the background area and its adjacent area until no new area is formed, and the iteration stops; B. Iteratively merge the area in the target area and its adjacent area until no new area is formed, and the iteration stops; C. Iteratively merge the area in the area to be merged with its adjacent area until no new area is formed; repeat the above step A, step B, and step C until there is no iterative merge action in step C. Stop.

Images