CN105373798A - A calligraphy character extraction method based on K-nearest neighbor matting and mathematical morphology - Google Patents

Abstract

The invention discloses a method for extracting calligraphy characters based on K-nearest neighbor matting and mathematical morphology. A reference image is established through K-nearest neighbor matting. According to the characteristics of complex changes in the ink density of calligraphy images, an edge extraction method based on mathematical morphology is used. Guided filtering of different window sizes is performed on the inside and edge of the font, and then image fusion technology is used to find more information about shape, quality and spirit. The present invention can more accurately extract the calligrapher's emotional and personality information of the calligraphic works themselves, especially for the pasted images with complex ink shades and fuzzy font edges in ancient Chinese calligraphy works, which can more accurately extract Chinese character information, Especially for the extraction of the information of Chinese characters, the effect is remarkable.

Description

A calligraphy character extraction method based on K-nearest neighbor matting and mathematical morphology

technical field

The invention belongs to the technical field of image processing, and relates to a calligraphy art information extraction method based on K-nearest neighbor matting and digital morphology in the fields of Chinese calligraphy art research and historical and cultural heritage protection, which is used for extracting the spirit information of Chinese characters in calligraphy works images.

Background technique

In the field of Chinese calligraphy art research and historical and cultural heritage protection, image preprocessing and image segmentation methods are used to extract the shape, quality and spirit information of Chinese characters more completely and accurately from ancient Chinese calligraphy works. At present, the method of extracting Chinese character information mainly adopts the method of combining image denoising, edge detection and image segmentation to extract the shape and quality information of Chinese characters.

Peking University's patent application "a method for detecting text in pictures" (publication number: CN101122952, date of authorization: February 13, 2008, date of application: September 21, 2007) discloses a method for detecting text in pictures . This method first merges the edge maps of the original image on each color component to obtain the cumulative edge map; then sets the edge points in the cumulative edge map to their corresponding colors in the original image, and uses the clustering method according to the color of the edge points. The method decomposes the cumulative edge map into several sub-edge maps; finally, in each sub-edge map, multiple horizontal and vertical projections are performed, and the vertical and horizontal regions are segmented according to the projection map to locate the text area in the picture. This method can accurately obtain the area information of the text in the image and the shape and quality information of the text, but the key to detecting the text information in this method mainly depends on the edge information of the image, that is, it mainly focuses on the shape and quality information of the text, and does not consider the text information. As a result, the detected text information is not complete enough, which has a negative impact on the research work of later calligraphy works.

XiaoqingLu et al. In the document "XiaoqingLu, ZhiTang, YanLiu, LiangcaiGao, TingWang, ZhipengWang.'Stroke-basedCharacterSegmentationofLow-qualityImagesonAncientChineseTablet'[C], 201312thInternationalConferenceonDocumentAnalysisandRecognition", in 2013, a Chinese character extraction based on low-quality ancient Chinese tablet images was proposed. method. The specific steps of the method include: (1) denoising and preprocessing the original stele image; (2) applying a mapping-based segmentation method to the denoised image to obtain the initial segmentation result; (3) using the adaptive Otsu method to set the minimum Intensity threshold to obtain the Stroke filter mask, and use the mask to filter the denoised image to obtain the intensity information of the Stroke; (4) combine the segmentation result obtained in step 2 and the filter mask obtained in step 3 Model to select the combined component with higher Stroke strength as the initial seed; (5) Based on the guiding information in the seed window, use an iterative process to extract the Chinese character information contained in the stele image; (6) After the iteration is over, get The segmentation result is the extracted Chinese character information. Although this method can better deal with the influence of the cracks in the stele image on the extraction of Chinese characters, and can extract the Chinese character information in the stele image more completely. However, this method only focuses on the extraction of Chinese character shape and quality information, and it is difficult to apply to the extraction of Chinese character information in pasted images; and in the process of implementing the method, many parameters are obtained based on experience. Therefore, the method has great limitations.

Contents of the invention

In view of the problems existing in the above-mentioned prior art, the purpose of the present invention is to propose a method for extracting artistic information of calligraphy works based on K-nearest neighbor matting and mathematical morphology, so as to overcome the limitations of existing art information extraction techniques in calligraphy works. The adopted Chinese character extraction method has serious shortcomings of lack of detailed information, which improves the accuracy of artistic information extraction and artistic appreciation value of calligraphy works.

In order to achieve the above tasks, the present invention adopts the following technical solutions:

A method for extracting calligraphy characters based on K-nearest neighbor matting and mathematical morphology, comprising the following steps:

Step 1, using software to read the color image to be processed in the computer;

Step 2, converting the color image to be processed from the RGB color space to the CTE-Lab color space to obtain the L channel;

Step 3, using the K-nearest neighbor matting algorithm to extract the Chinese character information in the image to obtain a grayscale image, which is used as a reference image for calligraphy character extraction;

Step 4, after binarizing the reference image, using mathematical morphology to corrode the binarized reference image, subtracting the corroded reference image from the reference image in step 3, to obtain the edge image of the calligraphy character;

Step 5, carry out the guided filtering process of the small window to the edge image, and select the window of suitable size to carry out the guided filtering process on the inner image of the calligraphy character;

Step 6, pixel-level image fusion is performed on the filtering result of the edge image and the filtering result of the internal image of the calligraphy character, and the process is completed.

Further, the specific process of the third step includes:

Step S30, using the color and position information of the image to extract a feature vector, the formula is as follows:

X(i)=(cos(h), sin(h), s, v, x, y) _i

In the above formula, X(i) is a feature vector, h, s, and v are three components of the HSV color space respectively, and x, y are the position coordinates of the pixel;

Step S31, define the kernel function:

k(i,j)=1-||X(i)-X(j)||/C

In the above formula, k(i,j) is the kernel function, X(i) and X(j) are different feature vectors, and C is the weight adjustment coefficient, which is used to ensure that k(i,j)∈(0, 1);

Get the Laplacian matrix from the kernel function:

L=D-A

In the above formula, D is a diagonal matrix, and the elements on the diagonal of D A is a similarity matrix;

Step S32, adding user constraint information to obtain a closed solution:

α＝(L+λM) ^-1 (λV)

In the above formula, M is a diagonal matrix, which represents the user's marking of known pixels, V is a vector, representing the user's marking of the foreground area, λ is the constraint coefficient, and L is the brightness of the color image in the Lab color space;

Step S33, put the value of the closed solution α into the following formula to obtain the reference image:

R=αf+(1-α)b

In the above formula, R is the reference image, f is the unknown foreground layer, and b is the unknown background layer.

Further, the specific process of the fourth step includes:

Step S40, binarize the reference image

Calculate the average pixel value of the reference image using the following formula:

$\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}$

In the above formula, u represents the average pixel value of the L channel, f(x, y) represents the pixel value of the pixel at the coordinates (x, y) in the image, and M and N represent the length and width of the image, respectively;

Step S41, assuming that the optimal threshold for binarization processing of the reference image is T, then counting the ratio w1 _of pixels with pixel values greater than T in the image in the L channel and the proportion of pixels with pixel values less than or equal to T in the image in the L channel. Scale w ₂ , and calculate the average pixel value u ₁ of the pixels whose pixel value is greater than T in the L channel and the average pixel value u ₂ of the pixels whose pixel value is less than or equal to T in the L channel;

Step S42, traversing through each possible value of T, using the following formula to calculate the inter-class difference value:

G＝w ₁ ×(u ₁ -u)×(u ₁ -u)+w ₂ ×(u ₂ -u)×(u ₂ -u)

In the above formula, G represents the difference between the target part and the background part in the binarization process. When G reaches the maximum, the optimal threshold T of binarization can be obtained, and then use the following formula to compare the reference The image is binarized:

Step S43, using a matrix of 3*3 structural elements to corrode the binarized reference image to reduce the edge of the calligraphy character by one pixel, and then subtracting the corroded reference image from the reference image to obtain the edge of the calligraphy character ;in:

The erosion operation is defined as:

RΘB ^s = {Z,B _z ∈ R}

The edge extraction operator of mathematical form is as follows:

ED(R)＝R-(RΘB)

In the above two formulas, B is a 3*3 structural element, B _Z is the result of the translation of the structural element by Z units, B ^S is the set of structural elements symmetrical about the origin, and ED(R) is the calligraphic character in the reference image R edge.

Compared with the prior art, the present invention has the following advantages:

1. The present invention uses K-nearest neighbor matting and a guided filter, which solves the problem that the traditional method lacks grayscale details and is difficult to extract blurred edges of calligraphy characters, and can better extract the shape, quality and spirit of Chinese characters in calligraphy works information, improving the integrity of Chinese character information extraction.

2. The present invention uses an edge extraction method based on mathematical morphology, which well solves the problem that traditional methods cannot keep the edges of Chinese characters sharp. Keep the edges of calligraphy characters smooth and smooth, and improve the accuracy of Chinese character information extraction.

3. The present invention establishes a reference image through K-nearest-neighbor matting, and according to the characteristics of complex changes in the ink intensity of calligraphy images, uses an edge extraction method based on mathematical morphology to perform guided filtering of different window sizes on the inside and edge of the font, and then uses the image Fusion technology seeks more shape, quality and spirit information, solves the problem of serious lack of detail information in the Chinese character extraction method used in the existing calligraphy art information extraction technology, and improves the accuracy and artistic quality of calligraphy art information extraction. Appreciation value.

Description of drawings

Fig. 1 is the overall flowchart of the inventive method;

Fig. 2 is a partial image of the calligraphy image "Songfengge Poetry";

Figure 3 is the result of color space conversion and K-nearest neighbor matting, where Figure 3(a) is the L channel of the original image, and Figure 3(b) is the reference image;

Fig. 4 is the result after extracting the edge of the reference image, wherein Fig. 4 (a) is an internal image, and Fig. 4 (b) is an edge image;

Fig. 5 is the result of guiding filtering with different window sizes for the inner image and the edge image, where Fig. 5(a) is the inner image and Fig. 5(b) is the edge image;

Fig. 6 is the result figure of image fusion;

Fig. 7 is a comparison chart of the experimental results of the first group of two image works in the simulation experiment 2;

Fig. 8 is the comparison chart of the experimental results of the two image works of the second group in the simulation experiment 2;

Fig. 9 is a comparison chart of the experimental results of two image works of the third group in the simulation experiment 2;

detailed description

1. Detailed steps

Flow chart of the present invention is as shown in Figure 1, and concrete process is as follows:

A method for extracting calligraphy characters based on K-nearest neighbor matting and mathematical morphology, comprising the following steps:

Step 1, use software in the computer to read the color image to be processed; the software used in the computer can adopt Matlab software;

Step 2: Convert the color image to be processed from the RGB color space to the CTE-Lab color space to obtain the L channel. The specific conversion formula is as follows:

X＝0.412453×R+0.357580×G+0.180423×B

Y＝0.212671×R+0.715160×G+0.072169×B

Z＝0.019334×R+0.119193×G+0.950227×B

X ₁ =X/(255×0.950456)

Y ₁ =Y/255

Z ₁ =Z/(255×1.088754)

Among them, X, Y, Z represent CIE1931 standard chromaticity observer spectral tristimulus value respectively, R, G, B represent the red, green, blue three channels of color image in RGB color space respectively, X ₁ , Y ₁ , Z ₁ represents the values of linearly normalized X, Y, and Z respectively;

If Y>0.008856, then:

f(X ₁ )=X ₁ ^(1/3)

f(Y ₁ )=Y ₁ ^(1/3)

f(Z ₁ )=Z ₁ ^(1/3)

L=116×f(Y ₁ )-16

If Y<0.008856, then:

f(X ₁ )＝7.787×X ₁ +16/116

f(Y ₁ )=7.787×Y ₁ +16/116

f(Z ₁ )=7.787×Z ₁ +16/116

L=903.3×Y ₁

a=500×(f(X ₁ )-f(Y ₁ ))+128

b=200×(f(Y ₁ )-f(Z ₁ ))+128

Among them, f( ) is a correction function, L represents the brightness of the color image in the Lab color space, a, b represent the color of the color image in the Lab color space, the positive half axis of a represents red, and the negative half axis represents green, The positive half-axis of b represents yellow, and the negative half-axis represents blue.

Step 3, use the K-nearest neighbor matting algorithm to extract the Chinese character information in the image, and obtain a grayscale image as a reference image for calligraphy character extraction; the specific process is as follows:

Step S30, using the color and position information of the image to extract a feature vector, the formula is as follows:

X(i)=(cos(h), sin(h), s, v, x, y) _i

In the above formula, X(i) is a feature vector, h, s, and v are three components of the HSV color space respectively, and x, y are the position coordinates of the pixel;

Step S31, define the kernel function:

k(i,j)=1-||X(i)-X(j)||/C

In the above formula, k(i,j) is the kernel function, X(i) and X(j) are different feature vectors, and C is the weight adjustment coefficient, which is used to ensure that k(i,j)∈(0, 1);

Get the Laplacian matrix from the kernel function:

L=D-A

In the above formula, D is a diagonal matrix, and the elements on the diagonal of D A is a similarity matrix;

Step S32, adding user constraint information to obtain a closed solution:

α＝(L+λM) ^-1 (λV)

In the above formula, M is a diagonal matrix, which represents the user's marking of known pixels, V is a vector, representing the user's marking of the foreground area, λ is the constraint coefficient, and L is the brightness of the color image in the Lab color space;

Step S33, put the value of the closed solution α into the following formula to obtain the reference image:

R=αf+(1-α)b

In the above formula, R is the reference image, f is the unknown foreground layer, and b is the unknown background layer.

Step 4, after binarizing the reference image, use mathematical morphology to corrode the binarized reference image, subtract the corroded reference image from the reference image in step 3, and obtain the edge image of the calligraphy character; the specific process is as follows:

Step S40, using OTSU to binarize the reference image

Calculate the average pixel value of the reference image using the following formula:

$\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}$

In the above formula, u represents the average pixel value of the L channel, f(x, y) represents the pixel value of the pixel at the coordinates (x, y) in the image, and M and N represent the length and width of the image, respectively;

Step S41, assuming that the optimal threshold for binarization processing of the reference image is T, then counting the ratio w1 _of pixels with pixel values greater than T in the image in the L channel and the proportion of pixels with pixel values less than or equal to T in the image in the L channel. Ratio w ₂ , and calculate the average pixel value u ₁ of the pixels whose pixel value is greater than T in the L channel and the average pixel value u ₂ of the pixels whose pixel value is less than or equal to T in the L channel; the related formula is:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}$

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}$

$\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; \&Sigma;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}& \mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; T</span>}\end{array}$

$\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; \&Sigma;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}& \mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; T</span>}\end{array}$

Among them, W ₁ and W ₂ respectively represent the number of pixels whose pixel value is greater than T and the number of pixels whose pixel value is less than or equal to T in the L channel, i represents the pixel value of the pixel in the image, and n(i) represents the number of pixels whose pixel value is equal to i ;

Step S42, traversing through each possible value of T, using the following formula to calculate the inter-class difference value:

G＝w ₁ ×(u ₁ -u)×(u ₁ -u)+w ₂ ×(u ₂ -u)×(u ₂ -u)

In the above formula, G represents the difference between the target part and the background part in the binarization process. When G reaches the maximum, the optimal threshold T of binarization can be obtained, and then use the following formula to compare the reference The image is binarized:

Step S43, in the mathematical morphology operation, erosion has the function of eliminating the boundary of the object. Use a matrix of 3*3 structural elements to corrode the binarized reference image to reduce the edge of the calligraphy character by one pixel, and then subtract the corroded reference image from the reference image to obtain the edge of the calligraphy character; among them:

The erosion operation is defined as:

RΘB ^s = {Z,B _z ∈ R}

The edge extraction operator of mathematical form is as follows:

ED(R)＝R-(RΘB)

In the above two formulas, B is a 3*3 structural element, B _Z is the result of the translation of the structural element by Z units, B ^S is the set of structural elements symmetrical about the origin, and ED(R) is the calligraphic character in the reference image R edge.

Step 5, carry out the guided filtering process of the small window to the edge image, and select the window of suitable size to carry out the guided filtering process on the inner image of the calligraphy character;

Step S50, firstly carry out the guided filtering process of the small window on the edge image ED:

Taking the edge image ED as the input image I, and the image L channel as the guide image Ig, the guide filter is a linear transformation of the guide image, namely:

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; g</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&ForAll;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}$

Among them, I _o (x, y) is the pixel value at the coordinate position in the filtered output image, a _k and b _k are linear coefficients, and I _g (x, y) is the coordinate position (x, y) in the guide image The pixel value of , ω _k is a local window with the pixel as the center and radius r. When guiding the filtering process to the edge image, in order to keep the edge sharp, use a small window to guide the filter, that is, use the local window ω _{2 of r=2} ;

In order to minimize the difference between the input image I and the output image _Io , the following function needs to be minimized in the window _ω2 :

E＝∑((I _o (x,y)-I(x,y)) ² +εa _k ² )

＝∑((a _k I _g (x,y)+b _k -I(x,y)) ² +εa _k ² )

where I(x,y) is the pixel value at the coordinate position (x,y) in the input image, E is the difference between _Io (x,y) and I(x,y), and ε is a A regularization parameter that prevents the value from being too large. When E reaches the minimum, a _k and b _k are respectively:

${\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&omega;</span>}<span\; class="notranslate">\; |</span>}\underset{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{<span\; class="notranslate">\; \&Sigma;</span>}{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; g</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; I</span>}}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; I</span>}}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}$

Among them, σ _k ² and μ _k are the mean and variance of I _g (x, y) in the window ω ₂ respectively, is the mean value of I(x,y) in the window ω ₂ , |ω ₂ | is the number of pixels in the window ω ₂ .

Since a pixel may be covered by multiple windows, the filtered output I _o (x,y) can be calculated by the following formula according to the calculated parameters a _k and b _k :

${\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}}_{\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}}{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; g</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; x</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}}$

in, and is the mean of all window coefficients covering pixel (x,y).

The guided filtering result ED' of the edge image ED is obtained from I _o (x, y).

Step S51, performing guided filtering processing on the inner image of the calligraphy character:

Taking the internal image IE as the input image I, taking the image L channel as the guide image Ig, in order to extract more grayscale information, use a larger filter window, that is, use the local window ω ₈ of r=8; other processes are all the same as the steps S50 is the same, and the guided filtering result IE' of the internal image is obtained.

Step 6, pixel-level image fusion is performed on the filtering result of the edge image and the filtering result of the internal image of the calligraphy character, and the process is completed.

In order to make the final extraction result contain both the rich grayscale information inside the calligraphy characters and the sharp edges of the strokes, the images of ED′ and IE′ are fused to obtain the complete and accurate shape, quality and charm of the calligraphy characters.

Take the minimum pixel value of ED' and IE' at (x, y), that is:

g(x,y)=min(ED'(x,y),IE'(x,y))

Among them, ED′(x, y) is the pixel value of ED′ at coordinates (x, y), IE′(x, y) is the pixel value of IE′ at coordinates (x, y), g(x, y ) is the pixel value of the image fusion result at coordinates (x, y), and min(·) is the minimum value.

2. Simulation experiment

Simulation experiment 1: Simulation of the method for extracting the shape, quality and spirit information of Chinese characters in calligraphy images in the present invention.

The simulation conditions of simulation 1 are carried out under MATLABR2013a software, the parameters of guided filtering ε=0.1 ¹⁰ , the parameters of K-nearest neighbor matting algorithm lambda=100, level=0.5, l=1.

Referring to Fig. 2 to Fig. 6, a simulation experiment is performed on part of the calligraphy image "Songfengge Poem" by Huang Tingjian. The calligraphy image is relatively well preserved, and the reference image contains the main shape and quality information of Chinese characters. Then, edge extraction is performed on the reference image to obtain the internal image and the edge image, which are filtered by guiding filters with different window sizes respectively to obtain For grayscale details and sharp edges, image fusion can be performed on the filtering results to extract complete information of the spirit, and can truly reflect the changes in ink density and the trend of strokes, and accurately restore the edge of calligraphy characters. In the resulting image, the calligrapher's strokes can be well displayed when writing Chinese characters, as well as the sudden change and gradual change of the stroke.

Simulation experiment 2 is a simulation for comparative analysis of the method of the present invention.

The simulation conditions of the simulation experiment 2 are carried out under the software MATLABR2013a, the parameters of the guided filter ε=0.1 ¹⁰ , the parameters of the K-nearest neighbor matting algorithm lambda=100, level=0.5, l=1. The method of the present invention is mainly compared and analyzed with OTSU, FastFuzzyC-means (FFCM) and MultiChannels and Guided Filters (MCGF) to prove that the method of the present invention has significant advantages in extracting the information of Chinese characters in calligraphy works. The comparison and analysis of the experimental results are described as follows:

Referring to Fig. 7, Fig. 8 and Fig. 9 (two calligraphy image works are selected for each group), for calligraphy images, it is necessary to extract complete and accurate shape, quality and expression information at the same time. First of all, for the extraction of the shape and quality information of Chinese characters, all the methods can extract the shape and quality information of Chinese characters relatively accurately. However, the shape and quality information extracted by the method of the present invention is more complete, as shown in FIG. 7 and FIG. 9 . In the extraction of the spirit information of Chinese characters, for the areas with blurred edges in calligraphy characters, the extraction results of Otsu, FFCM and MCGF have a large loss of details, as shown in Figure 8 and Figure 9. It is very difficult to extract the strokes written with a dry brush, and OTSU, FFCM and MCGF lose most of the information, as shown in Figure 7, Figure 8 and Figure 9. However, in contrast, the method of the present invention has high accuracy in the extraction of shape information and spirit information, and can retain rich grayscale details for both the white area and the dry strokes, and the edges of the strokes Clear and complete.

Claims (3)

1. A calligraphy word extraction method based on K nearest neighbor matting and mathematical morphology, is characterized in that, comprises the following steps: Step 1, using software to read the color image to be processed in the computer; Step 2, converting the color image to be processed from the RGB color space to the CTE-Lab color space to obtain the L channel; Step 3, using the K-nearest neighbor matting algorithm to extract the Chinese character information in the image to obtain a grayscale image, which is used as a reference image for calligraphy character extraction; Step 4, after binarizing the reference image, using mathematical morphology to corrode the binarized reference image, subtracting the corroded reference image from the reference image in step 3, to obtain the edge image of the calligraphy character; Step 5, carry out the guided filtering process of the small window to the edge image, and select the window of suitable size to carry out the guided filtering process on the inner image of the calligraphy character; Step 6, pixel-level image fusion is performed on the filtering result of the edge image and the filtering result of the internal image of the calligraphy character, and the process is completed. 2. the calligraphy word extraction method based on K nearest neighbor matting and mathematical morphology as claimed in claim 1, is characterized in that, the concrete process of described step 3 comprises: Step S30, using the color and position information of the image to extract a feature vector, the formula is as follows: X(i)=(cos(h), sin(h), s, v, x, y) _i In the above formula, X(i) is a feature vector, h, s, and v are three components of the HSV color space respectively, and x, y are the position coordinates of the pixel; Step S31, define the kernel function: k(i,j)=1-||X(i)-X(j)||/C In the above formula, k(i,j) is the kernel function, X(i) and X(j) are different feature vectors, and C is the weight adjustment coefficient, which is used to ensure that k(i,j)∈(0, 1); Get the Laplacian matrix from the kernel function: L=D-A In the above formula, D is a diagonal matrix, and the elements on the diagonal of D A is a similarity matrix; Step S32, adding user constraint information to obtain a closed solution: α＝(L+λM) ^-1 (λV) In the above formula, M is a diagonal matrix, which represents the user's marking of known pixels, V is a vector, representing the user's marking of the foreground area, λ is the constraint coefficient, and L is the brightness of the color image in the Lab color space; Step S33, put the value of the closed solution α into the following formula to obtain the reference image: R=αf+(1-α)b In the above formula, R is the reference image, f is the unknown foreground layer, and b is the unknown background layer. . 3. the calligraphy word extraction method based on K nearest neighbor matting and mathematical morphology as claimed in claim 1, is characterized in that, the concrete process of described step 4 comprises: Step S40, binarize the reference image Calculate the average pixel value of the reference image using the following formula: $\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; N</span>}}$ In the above formula, u represents the average pixel value of the L channel, f(x, y) represents the pixel value of the pixel at the coordinates (x, y) in the image, and M and N represent the length and width of the image, respectively; Step S41, assuming that the optimal threshold for binarization processing of the reference image is T, then counting the ratio w1 _of pixels with pixel values greater than T in the image in the L channel and the proportion of pixels with pixel values less than or equal to T in the image in the L channel. Scale w ₂ , and calculate the average pixel value u ₁ of the pixels whose pixel value is greater than T in the L channel and the average pixel value u ₂ of the pixels whose pixel value is less than or equal to T in the L channel; Step S42, traversing through each possible value of T, using the following formula to calculate the inter-class difference value: G＝w ₁ ×(u ₁ -u)×(u ₁ -u)+w ₂ ×(u ₂ -u)×(u ₂ -u) In the above formula, G represents the difference between the target part and the background part in the binarization process. When G reaches the maximum, the optimal threshold T of binarization can be obtained, and then use the following formula to compare the reference The image is binarized: Step S43, using a matrix of 3*3 structural elements to corrode the binarized reference image to reduce the edge of the calligraphy character by one pixel, and then subtracting the corroded reference image from the reference image to obtain the edge of the calligraphy character ;in: The erosion operation is defined as: RΘB ^s = {Z,B _z ∈ R} The edge extraction operator of mathematical form is as follows: ED(R)＝R-(RΘB) In the above two formulas, B is a 3*3 structural element, B _Z is the result of the translation of the structural element by Z units, B ^S is the set of structural elements symmetrical about the origin, and ED(R) is the calligraphic character in the reference image R edge.