CN101751565B - Method for character identification through fusing binary image and gray level image - Google Patents

Abstract

The invention relates to a method for character identification through fusing a binary image and a gray level image, which comprises the following steps: processing the binary image of a character image and a fusion image of the gray level image to carry out character identification; fusing the binary image of the character image and the gray level image to obtain the fusion image; normalizing the dimension and the position of the fusion image; extracting the features of a gradient histogram of a normalized image; obtaining a feature dimension reduction transformation matrix by utilizing principle component analysis and linear discriminant analysis; and establishing a character feature template library to carry out the character identification. The invention overcomes the defect that the traditional character identification technology of the binary image based on characters or the gray level image based on characters can not simultaneously identify the degenerated character image or the character image containing complex background.

Description

Merge the method for the character recognition of bianry image and gray level image

Technical field

The invention belongs to character recognition field (being called for short OCR), relate to the method for the character recognition of consistent fusion bianry image and gray level image.

Background technology

Traditional character recognition technologies is based on the bianry image of character or based on the gray level image of character.When the recognition technology based on the bianry image of character is applied to various low-quality images, such as the low-resolution images such as character picture in the degraded character image in the video, ID Card Image, license plate, the natural scene, because the character picture quality after the binaryzation is low, recognition effect is poor.When the recognition technology based on the gray level image of character is applied to comprise the character picture of complex background, such as the character picture in the video, because character picture comprises incomparable inconsistent background, recognition effect is with variation.

Summary of the invention

In order to solve the problem of prior art, the object of the present invention is to provide a kind of bianry image of character and method that gray level image carries out character recognition of merging.

For reaching described purpose, the method for the character recognition of fusion bianry image provided by the invention and gray level image is processed the fused images of described bianry image and gray level image, carries out character recognition, and it may further comprise the steps:

Step 1: the bianry image of establishing the single character picture that obtains after the pre-service is B ₀=[b ₀(x, y)], the value that wherein is positioned at the pixel of the capable y row of x is b ₀(x, y), b ₀(x, y) is 0 or 1, and the size of image is W ₁* H ₁The gray level image of character is G _c=[g _c(x, y)], the value that is positioned at the pixel of the capable y row of x is g _c(x, y), 0≤g _c(x, y)≤255; Bianry image B with character picture ₀With gray level image G _cMerge, get the image G=[g (x, y) after the fusion], the value that is positioned at the pixel of the capable y row of x is g (x, y), 0≤g (x, y)≤255;

Step 2: extracting fused images G=[g (x, y)] feature before, carry out first fused images G=[g (x, y)] the position and the normalized of size; The input picture that image normalization is processed is G=[g (x, y)], the output image after the normalization is F=[f (x ', y ')], its size is respectively W ₁* H ₁And W ₂* H ₂Input picture G=[g (x, y)] pixel that is positioned at the capable y of x row will be mapped to F=[f (x ', y ')] being positioned at the pixel of x ' row y ' row, the coordinate by input picture and output image shines upon to realize image normalization:

$\left\{\begin{array}{c}x\′=x\′(x,y)\\ y\′=y\′(x,y)\end{array}\right.$

One dimension coordinate is mapped as:

$\left\{\begin{array}{c}x\′=x\′\left(x\right)\\ y\′=y\′\left(x\right)\end{array}\right.;$

Step 3: the feature of extracting the histogram of gradients of normalized image based on histogram of gradients;

Step 4: utilize principal component analysis and linear discriminant analysis that the feature of the histogram of gradients of normalized image is carried out dimension-reduction treatment, obtain the transformation matrix of Feature Dimension Reduction;

Step 5: set up character feature template library, read the transformation matrix of Feature Dimension Reduction and character is identified.

Beneficial effect of the present invention: the invention is characterized in the bianry image of character picture and the fused images of gray level image are processed that carry out character recognition, it may further comprise the steps: the fusion of (1) bianry image and gray level image; (2) normalization of image; (3) based on the feature extraction of histogram of gradients; (4) Feature Dimension Reduction; (5) classifier design and character recognition.The present invention has overcome based on the bianry image of character or the shortcoming that can not identify simultaneously the degraded character image and comprise the character picture of complex background based on traditional character recognition technologies of the gray level image of character.The technical field of application of the present invention comprises the character recognition in the video, the character recognition in ID Card Image, license plate, the natural scene image

Description of drawings

Fig. 1 is character recognition system process flow diagram of the present invention;

Fig. 2 is the framework synoptic diagram of the fusion of bianry image of the present invention and gray level image;

Fig. 3 is the normalized framework synoptic diagram of image of the present invention;

Fig. 4 is the framework synoptic diagram that the present invention is based on the feature extraction of histogram of gradients;

Fig. 5 is the configuration diagram that the present invention asks the transformation matrix of Feature Dimension Reduction;

Fig. 6 is classifier design of the present invention and character recognition configuration diagram;

Fig. 7 is Sobel gradient operator template;

Fig. 8 is L reference direction example, left side L=4, the right L=8;

Fig. 9 is that gradient is decomposed example;

Figure 10 be calculating pixel and center, rectangular area in the horizontal direction with vertical direction on distance examples.

Embodiment

Describe each related detailed problem in the technical solution of the present invention in detail below in conjunction with accompanying drawing.Be to be noted that described embodiment only is intended to be convenient to the understanding of the present invention, and it is not played any restriction effect.

As shown in Figure 1, character recognition system process flow diagram of the present invention, recognizer can be divided into two parts: training system and recognition system.Training system merges its bianry image and gray level image to each character training sample, and fused images is carried out the normalization of size and position, extracts the feature of histogram of gradients; The feature that utilization is extracted from training sample is found the solution the transformation matrix that carries out Feature Dimension Reduction, obtains the character recognition storehouse.In recognition system, merge bianry image and the gray level image of character to be identified, fused images is carried out the normalization of size and position, extract the feature of histogram of gradients, the transformation matrix that utilizes training system to obtain carries out dimensionality reduction to feature, then sends into recognizer, obtains recognition result.

Fusion character bianry image and gray level image carry out the realization of character recognition system need to consider following several aspect:

1) realization of training system;

2) realization of recognition system.

The below describes in detail to these two aspects respectively.

The realization of 1 training system

1.1 see also the framework of the fusion of the bianry image shown in Fig. 2 and gray level image.

If the bianry image of the single character picture that obtains after the pre-service is B ₀=[b ₀(x, y)], the value that wherein is positioned at the pixel of the capable y row of x is b ₀(x, y), b ₀(x, y) is 0 or 1.The gray level image of character is G _c=[g _c(x, y)], the value that is positioned at the pixel of the capable y row of x is g _c(x, y), 0≤g _c(x, y)≤255.Bianry image is B ₀With gray level image G _cSize be W ₁* H ₁Carry out the bianry image B of character by following flow process ₀Gray level image G with character _cFusion, the image G=[g (x, y) after obtaining merging], the value that is positioned at the pixel of the capable y of x row is g (x, y), 0≤g (x, y)≤255:

Ask a total head threshold value th, for example can be to gray level image g _c(x, y) utilizes traditional maximum variance between clusters (OSTU method) to try to achieve this total head threshold value th _Ostu, make th=a x th _Ostu, a is a constant.Utilize global threshold th that gray level image g (x, y) is carried out threshold processing and obtain bianry image B _g=[b _g(x, y)], b _g(x, y) is defined as:

$b_g(x,y)=\left\{\begin{array}{cc}1& g_c(x,y)>\mathrm{th}\\ 0& g_c(x,y)\≤\mathrm{th}\end{array}\right.,x=0,\·\·\·,W_1-1,y=0,\·\·\·,H_1-1;$

Utilize two-value morphology to bianry image B _oCarry out condition expansion.If D is the image of 3 x 3, the pixel value of its each position is 1.To bianry image B _oCarrying out condition expansion is:

$B_{i+1}=(B_i\&CirclePlus;D)\&cap;B_g,i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N,{\mathrm{<figure-callout\; id="1"\; label="B"\; filenames="H2008102393315E00012.png,H2008102393315E00041.png"\; state="\{\{state\}\}">B</figure-callout>}}_1={\mathrm{<figure-callout\; id="0"\; label="B"\; filenames="H2008102393315E00012.png,H2008102393315E00041.png"\; state="\{\{state\}\}">B</figure-callout>}}_0$

According to following formula to bianry image B _iRepeatedly carry out condition expansion, until B _I+1=B _iOr reach maximum iterations, establishing the bianry image that obtains at last is B=[b (x, y)].

Obtain the fused images G=[g (x, y) for single character recognition], g (x, y) is defined as:

$g(x,y)=\left\{\begin{array}{cc}{g}_c(x,y)& b(x,y)=1\\ 0& b(x,y)=0\end{array}\right.,x=0,\&CenterDot;\&CenterDot;\&CenterDot;,W_1-1,y=0,\&CenterDot;\&CenterDot;\&CenterDot;,H_1-1.$

1.2 see also the normalized framework of the image shown in Fig. 3.

Before the feature of extracting fused images, the normalized of the position of advanced line character image and size.The input picture of image normalization is G=[g (x, y)], the output image after the normalization is F=[f (x ', y ')], its size is respectively W ₁* H ₁And W ₂* H ₂Input picture G=[g (x, y)] pixel that is positioned at the capable y of x row will be mapped to F=[f (x ', y ')] being positioned at the pixel of x ' row y ' row, the coordinate by input picture and output image shines upon to realize image normalization:

$\left\{\begin{array}{c}x\&prime;=x\&prime;(x,y)\\ y\&prime;=y\&prime;(x,y)\end{array}\right.$

One dimension coordinate is mapped as

$\left\{\begin{array}{c}x\&prime;=x\&prime;\left(x\right)\\ y\&prime;=y\&prime;\left(x\right)\end{array}\right.;$

Calculate fused images G=[g (x, y)] barycenter (x _c, y _c), barycenter is adjusted into normalized image F=[f (x ', y ')] center (W ₂/ 2, H ₂/ 2):

$g_x\left(x\right)={\mathrm{\&Sigma;}}_{y=0}^{H_1-1}g(x,y)/{\mathrm{\&Sigma;}}_{x=0}^{W_1-1}{\mathrm{\&Sigma;}}_{y=0}^{H_1-1}g(x,y),x=0,\&CenterDot;\&CenterDot;\&CenterDot;,W_1-1,$

$g_y\left(y\right)={\mathrm{\&Sigma;}}_{x=0}^{W_1-1}g(x,y)/{\mathrm{\&Sigma;}}_{x=0}^{W_1-1}{\mathrm{\&Sigma;}}_{y=0}^{H_1-1}g(x,y),y=0,\&CenterDot;\&CenterDot;\&CenterDot;,H_1-1,$

$x_c={\mathrm{\&Sigma;}}_{x=0}^{W_1-1}x{g}_x\left(x\right),$

$y_c={\mathrm{\&Sigma;}}_{y=0}^{H_1-1}y{g}_y\left(y\right),$

G wherein _x(x) and g _y(y) be respectively fused images G=[g (x, y)] in the vertical direction with horizontal direction on picture element density;

According to centroid position (x _c, y _c), computed image G=[g (x, y)] monolateral second moment

With

${\mathrm{\&mu;}}_x^{+}={\mathrm{\&Sigma;}}_{x>x_c}{(x-x_c)}^2{g}_x\left(x\right)$

${\mathrm{\&mu;}}_x^{+}={\mathrm{\&Sigma;}}_{x<x_c}{(x-x_c)}^2{g}_x\left(x\right)$

${\mathrm{\&mu;}}_y^{+}={\mathrm{\&Sigma;}}_{y>y_c}{(y-y_c)}^2{g}_y\left(y\right)$

${\mathrm{\&mu;}}_y^{+}={\mathrm{\&Sigma;}}_{y<y_c}{(y-y_c)}^2{g}_y\left(y\right);$

The housing that input picture is set according to the monolateral second moment that calculates is $[x_c-2\sqrt{{\mathrm{\&mu;}}_x^{-}},x_c+2\sqrt{{\mathrm{\&mu;}}_x^{+}}]$ With $[y_c-2\sqrt{{\mathrm{\&mu;}}_y^{-}},y_c+2\sqrt{{\mathrm{\&mu;}}_y^{+}}].$ For the x axle, find the solution quadratic function u (x)=ax ²+ bx+c is three points on the x axle $(x_c-2\sqrt{{\mathrm{\&mu;}}_x^{-}},x_c,x_c+2\sqrt{{\mathrm{\&mu;}}_x^{+}})$ Be mapped as respectively (0,0.5,1), in like manner obtain the quadratic function u (y) of y axle three points on the y axle $(y_c-2\sqrt{{\mathrm{\&mu;}}_y^{-}},y_c,y_c+2\sqrt{{\mathrm{\&mu;}}_y^{+}})$ Be mapped as respectively (0,0.5,1); Obtain input picture G=[g (x, y)] be positioned at the pixel of the capable y of x row and output image F=[f (x ', y ')] be positioned at the coordinate mapping function of the pixel of x ' row y ' row:

$\left\{\begin{array}{c}x\&prime;=W_{2}u\left(x\right)\\ y\&prime;=H_{2}u\left(y\right)\end{array}\right.;$

Determine input picture G=[g (x according to the coordinate mapping function, y)] with normalized image F=[f (x ', y ')] the coordinate mapping relations, the input picture gray-scale value is passed through bilinear interpolation, obtain normalized image F=[f (x ', y ')] value.

1.3 see also the framework based on the feature extraction of histogram of gradients shown in Fig. 4.

Two 3 * 3 templates utilizing the Sobel operator are computed image F=[f (x, y) respectively] in each locational gradient, two 3x3 templates of Sobel operator are as shown in Figure 7.For image F=[f (x, y)], it is tried to achieve by following formula along x axle and the axial first order derivative component of y respectively:

g _x(x，y)＝f(x+1，y-1)+2f(x+1，y)+f(x+1，y+1)

-f(x-1，y-1)-2f(x-1，y)-f(x-1，y+1)，

g _y(x，y)＝f(x-1，y+1)+2f(x，y+1)+f(x+1，y+1)

-f(x-1，y-1)-2f(x，y-1)-f(x+1，y-1).

x＝0，...，W ₂-1，y＝1，...，H ₂-1；

Computed image F=[f (x, y)] gradient intensity mag (x, y) and the deflection of position (x, y)

(x, y) is respectively:

$\mathrm{mag}(x,y)={[g_x^2(x,y)+g_y^2(x,y)]}^{1/2},$

Define L reference direction, the situation of L=4 and L=8 as Fig. 8 and shown in.Utilize parallelogram law to be decomposed into from its two nearest reference direction, as shown in Figure 9 gradient.Be W with size after the normalization ₂* H ₂Image F=[f (x, y)] be divided into R * R mutually disjoint rectangular area, set up the gradient orientation histogram of L dimension for each rectangular area.Image F=[f (x, y)] in the gradient orientation histogram of nearest 4 rectangular areas of gradient pair and this pixel of each pixel contribution is arranged.Be 4 rectangular areas (from top to bottom, from left to right being numbered respectively 1,2,3 and 4) that pixel is nearest with it as shown in figure 10, wherein each little rectangle frame represents a pixel, and 4 * 4 little rectangle frames form a large rectangle zone.In the horizontal direction, the distance at pixel and center, rectangular area is respectively d _HlAnd d _HrIn the vertical direction, the distance at pixel and center, rectangular area is respectively d _VtAnd d _VbIf the intensity of the component of pixel gradient on the l direction is g _l, then the gradient of this pixel is to the 1st, 2, and the contribution margin of the l of the gradient orientation histogram of 3 and 4 rectangular areas dimension is respectively g _l* d _Hr* d _Vb/ ((d _Hl+ d _Hr) * (d _Vt+ d _Vb)), g _l* d _Hl* d _Vb/ ((d _Hl+ d _Hr) * (d _Vt+ d _Vb)), g _l* d _Hr* d _Vt/ ((d _Hl+ d _Hr) * (d _Vt+ d _Vb)) and g _l* d _Hl* d _Vt/ ((d _Hl+ d _Hr) * (d _Vt+ d _Vb)).Utilize this method to calculate the gradient of each pixel to the contribution of the gradient orientation histogram of rectangular area adjacent thereto, try to achieve each rectangular area gradient orientation histogram, obtained at last the R * R of character picture * L dimensional feature.

1.4 see also the framework of the transformation matrix of asking Feature Dimension Reduction shown in Fig. 5:

1.4.1 principal component analysis (PCA)

The high dimensional feature vector comprises the feature that is mutually related, and it is large that it is processed operand, utilizes principal component analysis that the high dimensional feature vector is carried out principal component analysis (PCA), finds the solution PCA dimensionality reduction matrix P _PCAIf the character feature that extracts from n training sample is x _i, i=1 ..., n, x _iDimension m=R * R * L; The Scatter Matrix of training sample character feature is:

$\mathrm{\&Sigma;}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x})}^T(x_i-\stackrel{\&OverBar;}{x}),\stackrel{\&OverBar;}{x}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i$

Scatter Matrix is carried out Eigenvalues Decomposition is:

∑＝UΛU ^T

U=[u wherein ₁, u ₂..., u _m] be orthogonal matrix, Λ=diag (λ ₁, λ ₂..., λ _m) be diagonal matrix, λ ₁〉=λ ₂〉=... 〉=λ _mBe eigenwert.If will preserve the energy of r% behind the principal component analysis PCA dimensionality reduction, then the principal direction number l of principal component analysis preservation is

$l=\arg \underset{k}{\min }(\frac{{\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{\&lambda;}}_i}{{\mathrm{\&Sigma;}}_{i=1}^m{\mathrm{\&lambda;}}_i}\&GreaterEqual;r)$

The transformation matrix that principal component analysis obtains is P _PCA=[u ₁, u ₂..., u _l], to character feature x _i, carry out the l dimension character feature z after dimensionality reduction obtains dimensionality reduction _i=(P _PCA) ^Tx _i, i=1 ..., n, (P _PCA) ^TExpression P _PCATransposed matrix;

1.4.2 the character feature behind the training sample dimensionality reduction is carried out linear discriminant analysis (LDA), finds the solution transformation matrix W:

If character class number to be identified in the recognition system is C, the i class comprises n _iIndividual training sample.Calculate i class character sample characteristic mean μ _iWith all sample characteristics average μ:

${\mathrm{\&mu;}}_i=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_k^i,$ ${\mathrm{\&mu;}}_i=\frac{1}{n}\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_k^i,$ $n=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{n}_i$

Scatter Matrix S between compute classes _bWith Scatter Matrix S in the class _w:

$S_b=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{n_i}{n}({\mathrm{\&mu;}}_i-\mathrm{\&mu;}){({\mathrm{\&mu;}}_i-\mathrm{\&mu;})}^T$

$S_w=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\left(\frac{n_i}{n}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}(z_k^i-{\mathrm{\&mu;}}_i){(z_k^i-{\mathrm{\&mu;}}_i)}^T\right)$

A transformation matrix W is sought in linear discriminant analysis so that after the conversion between class dispersion as far as possible large, dispersion is as far as possible little in the class simultaneously, utilizes the maximization criterion

$J=\frac{\mathrm{tr}\left(W^T{S}_{b}W\right)}{\mathrm{tr}\left(W^T{S}_{w}W\right)}$

Represent.LDA can solve by finding the solution the generalized eigenvector problem:

S _bw＝λS _ww

If vectorial w ₁..., w _d..., w _lBe the solution of generalized eigenvector problem, the generalized eigenvalue λ of their correspondences ₁〉=... 〉=λ _d〉=... 〉=λ _l, the solution proper vector of d generalized eigenvector problem forms W, i.e. W=[w before selecting ₁..., w _d].

1.5 see also the classifier design shown in Fig. 6 and character recognition framework:

Utilize transformation matrix W to i character type characteristic mean μ _iCarry out dimensionality reduction, and the feature behind the normalization dimensionality reduction

${\mathrm{\&mu;}}_i^{*}=W^T{\mathrm{\&mu;}}_i,{\stackrel{\&OverBar;}{\mathrm{\&mu;}}}_i={\mathrm{\&mu;}}_i^{*}/\sqrt{{\left({\mathrm{\&mu;}}_i^{*}\right)}^T{\mathrm{\&mu;}}_i^{*}}$

Preserve transformation matrix P=WP _PCA, the coding of each character type and characteristic of correspondence μ thereof _iIn the File of identification storehouse.

(5.2) character recognition

The realization of 2 recognition systems

From the File of character recognition storehouse, read transformation matrix P, the coding of each character type and characteristic of correspondence μ thereof _iBianry image and gray level image to each character to be identified merge, and the image after merging is carried out normalization, carry out the multidimensional characteristic a that feature extraction obtains character picture.Utilize transformation matrix P that the multidimensional characteristic a of character picture is carried out feature b=P after Feature Dimension Reduction obtains dimensionality reduction ^TA, P ^TTransposed matrix for transformation matrix P.Feature normalization behind the dimensionality reduction is obtained

$\stackrel{\&OverBar;}{b}=b/\sqrt{b^{T}b}.$

Normalization center vector { the μ of order computation b and each character type _i} _1≤i≤CCosine distance { d _i} _1≤i≤C

d _i＝1-y ^Tμ

The minimum class of distance is the recognition result of character picture.

The above; only be the embodiment among the present invention; but protection scope of the present invention is not limited to this; anyly be familiar with the people of this technology in the disclosed technical scope of the present invention; can understand conversion or the replacement expected; all should be encompassed in of the present invention comprising within the scope, therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.

Claims (5)

1. a method that merges the character recognition of bianry image and gray level image is characterized in that, the fused images of described bianry image and gray level image is processed, and carries out character recognition, and it may further comprise the steps:

Step 1: the bianry image of establishing the single character picture that obtains after the pre-service is B ₀=[b ₀(x, y)], the value that wherein is positioned at the pixel of the capable y row of x is b ₀(x, y), b ₀(x, y) is 0 or 1, and the size of image is W ₁* H ₁The gray level image of character is G _c=[g _c(x, y)], the value that is positioned at the pixel of the capable y row of x is g _c(x, y), 0≤g _c(x, y)≤255; Bianry image B with character picture ₀With gray level image G _cMerge the image G=[g (x, y) after obtaining merging], the value that is positioned at the pixel of the capable y row of x is g (x, y), 0≤g (x, y)≤255;

Step 2: extracting fused images G=[g (x, y)] feature before, carry out first fused images G=[g (x, y)] the position and the normalized of size; The input picture that image normalization is processed is G=[g (x, y)], the output image after the normalization is F=[f (x ', y ')], its size is respectively W ₁* H ₁And W ₂* H ₂Input picture G=[g (x, y)] pixel that is positioned at the capable y of x row will be mapped to F=[f (x ', y ')] being positioned at the pixel of x ' row y ' row, the coordinate by input picture and output image shines upon to realize image normalization:

$\left\{\begin{array}{c}{x}^{\&prime;}=x^{\&prime;}(x,y)\\ y^{\&prime;}=y^{\&prime;}(x,y)\end{array}\right.$

One dimension coordinate is mapped as:

$\left\{\begin{array}{c}{x}^{\&prime;}=x^{\&prime;}\left(x\right)\\ y^{\&prime;}=y^{\&prime;}\left(x\right)\end{array}\right.;$

Step 3: the feature of extracting the histogram of gradients of normalized image based on histogram of gradients;

Step 4: utilize principal component analysis and linear discriminant analysis that the feature of the histogram of gradients of normalized image is carried out dimension-reduction treatment, obtain the transformation matrix of Feature Dimension Reduction;

Step 5: set up character feature template library, read the transformation matrix of Feature Dimension Reduction and character is identified;

The fusion of described bianry image and gray level image comprises:

Step 11: to gray level image G _c=[g _c(x, y)] pixel point value g _c(x, y) utilizes traditional maximum variance between clusters to try to achieve threshold value th _Ostu, ask a global threshold th, make th=a * th _Ostu, a is a constant; Utilize total head threshold value th to gray level image G _c=[g _c(x, y)] pixel point value g _c(x, y) carries out threshold processing, obtains bianry image B _g=[b _g(x, y)], the pixel point value b of bianry image _g(x, y) is defined as:

$b_g(x,y)=\left\{\begin{array}{cc}1& g_c(x,y)>\mathrm{th}\\ 0& g_c(x,y)\&le;\mathrm{th}\end{array}\right.,$ x＝0，...，W ₁-1，y＝0，...，H ₁-1；

Step 12: utilize two-value morphology to bianry image B _oCarry out condition expansion, establishing D is the image of a 3x3, and the pixel value of its each position is 1; To bianry image B _oCarrying out condition expansion is:

$B_{i+1}=(B_i\&CirclePlus;)\&cap;B_g,$ i＝1，2，...，N，B ₁＝B ₀

According to following formula to bianry image B _iRepeatedly carry out condition expansion, until B _I+1=B _iOr reach maximum iterations, establishing the bianry image that obtains at last is B=[b (x, y)];

Step 13: obtain the fused images G=[g (x, y) for single character recognition], g (x, y) is defined as:

$g(x,y)=\left\{\begin{array}{cc}{g}_c(x,y)& b(x,y)=1\\ 0& b(x,y)=0\end{array}\right.,$ x＝0，...，W ₁-1，y＝0，...，H ₁-1。

2. the method for the character recognition of described fusion bianry image and gray level image according to claim 1 is characterized in that the normalization of image comprises:

Step 21: calculate fused images G=[g (x, y)] barycenter (x _c, y _c), barycenter is adjusted into normalized image F=[f (x ', y ')] center (W ₂/ 2, H ₂/ 2):

$g_x\left(x\right)={\mathrm{\&Sigma;}}_{y=0}^{H_1-1}g(x,y)/{\mathrm{\&Sigma;}}_{x=0}^{W_1-1}{\mathrm{\&Sigma;}}_{x=0}^{H_1-1}g(x,y),$ x＝0，...，W ₁-1，

$g_y\left(y\right)={\mathrm{\&Sigma;}}_{x=0}^{W_1-1}g(x,y)/{\mathrm{\&Sigma;}}_{x=0}^{W_1-1}{\mathrm{\&Sigma;}}_{y=0}^{H_1-1}g(x,y),$ y＝0，...，H ₁-1，

$x_c={\mathrm{\&Sigma;}}_{x=0}^{W_1-1}{\mathrm{xg}}_x\left(x\right),$

$y_c={\mathrm{\&Sigma;}}_{y=0}^{H_1-1}{\mathrm{yg}}_y\left(y\right),$

G wherein _x(x) and g _y(y) be respectively fused images G=[g (x, y)] in the vertical direction with horizontal direction on picture element density;

Step 22: according to centroid position (x _c, y _c), computed image G=[g (x, y)] monolateral second moment

With

${\mathrm{\&mu;}}_x^{+}={\mathrm{\&Sigma;}}_{x>x_c}{(x-x_c)}^2{g}_x\left(x\right)$

${\mathrm{\&mu;}}_x^{-}={\mathrm{\&Sigma;}}_{x<x_c}{(x-x_c)}^2{g}_x\left(x\right)$

${\mathrm{\&mu;}}_y^{+}={\mathrm{\&Sigma;}}_{y>y_c}{(y-y_c)}^2{g}_y\left(y\right)$

${\mathrm{\&mu;}}_y^{-}={\mathrm{\&Sigma;}}_{y<y_c}{(y-y_c)}^2{g}_y\left(y\right)$

Step 23: the housing that input picture is set according to the monolateral second moment that calculates is:

$[x_c-2\sqrt{{\mathrm{\&mu;}}_x^{-}},x_c=2\sqrt{{\mathrm{\&mu;}}_x^{+}}]$ With $[y_c-2\sqrt{{\mathrm{\&mu;}}_y^{-}},y_c+2\sqrt{{\mathrm{\&mu;}}_y^{+}}];$

For the x axle, find the solution quadratic function u (x)=ax ²+ bx+c is three points on the x axle

Be mapped as respectively (0,0.5,1), in like manner obtain the quadratic function u (y) of y axle three points on the y axle

Be mapped as respectively (0,0.5,1); Obtain input picture G=[g (x, y)] be positioned at the pixel of the capable y of x row and output image F=[f (x ', y ')] be positioned at the coordinate mapping function of the pixel of x ' row y ' row:

$\left\{\begin{array}{c}{x}^{\&prime;}=W_{2}u\left(x\right)\\ y^{\&prime;}=H_{2}u\left(y\right)\end{array},\right.$ W ₂, H ₂Be respectively output image F=[f (x ', y ')] wide and high;

Step 24: finally obtain normalized image F=[f (x ', y ') by bilinear interpolation] value.

3. the method for the character recognition of described fusion bianry image and gray level image according to claim 1 is characterized in that, the step of the feature of the described histogram of gradients of extracting normalized image based on histogram of gradients comprises:

Step 31: two 3 * 3 templates utilizing the Sobel operator are computed image F=[f (x, y) respectively] in each locational gradient; For image F=[f (x, y)], it is tried to achieve by following formula along x axle and the axial first order derivative component of y respectively:

g _x(x，y)＝f(x+1，y-1)+2f(x+1，y)+f(x+1，y+1)

-f(x-1，y-1)-2f(x-1，y)-f(x-1，y+1)，

g _y(x，y)＝f(x-1，y+1)+2f(x，y+1)+f(x+1，y+1)，

-f(x-1，y-1)-2f(x，y-1)-f(x+1，y-1).

x＝0，...，W ₂-1，y＝0，...，H ₂-1；

Step 32: image F=[f (x, y)] gradient intensity mag (x, y) and the deflection of position (x, y) Be respectively:

$\mathrm{mag}(x,y)=[g_x^2(x,y)+g_x^2(x,y){]}^{1/2},$

Step 33: defining L reference direction, utilize parallelogram law to be decomposed into from its two nearest reference direction gradient, is W with size after the normalization ₂* H ₂Image F=[f (x, y)] be divided into R * R mutually disjoint rectangular area, set up the gradient orientation histogram of L dimension for each rectangular area; Image F=[f (x, y)] in the gradient orientation histogram of nearest 4 rectangular areas of gradient pair and this pixel of each pixel contribution is arranged; Calculate the gradient of each pixel to the contribution of the gradient orientation histogram of rectangular area adjacent thereto, try to achieve each rectangular area gradient orientation histogram, obtained at last the R * R of character picture * L dimensional feature.

4. the method for the character recognition of described fusion bianry image and gray level image according to claim 1 is characterized in that the described step of utilizing principal component analysis and linear discriminant analysis that the feature of the histogram of gradients of normalized image is carried out dimension-reduction treatment comprises:

Step 41: the high dimensional feature vector is carried out principal component analysis (PCA), find the solution PCA dimensionality reduction matrix P _PCA:

The high dimensional feature vector comprises the feature that is mutually related, and it is large that it is processed operand, utilizes principal component analysis that the high dimensional feature vector is carried out principal component analysis (PCA), finds the solution PCA dimensionality reduction matrix P _PCAIf the character feature that extracts from n training sample is x _i, i=1 ..., n, x _iDimension m=R * R * L; The Scatter Matrix of training sample character feature is:

$\mathrm{\&Sigma;}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x})}^T(x_i-\stackrel{\&OverBar;}{x}),$ $\stackrel{\&OverBar;}{x}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i,$

Scatter Matrix is carried out Eigenvalues Decomposition is:

∑＝UΛU ^T

U=[u wherein ₁, u ₂..., u _m] be orthogonal matrix, Λ=diag (λ ₁, λ ₂..., λ _m) be diagonal matrix, λ ₁〉=λ ₂〉=... 〉=λ _mBe eigenwert, establish the energy that will preserve r% behind the principal component analysis PCA dimensionality reduction, then the principal direction number l that preserves of principal component analysis is:

$l=\arg \underset{k}{\min }(\frac{{\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{\&lambda;}}_i}{{\mathrm{\&Sigma;}}_{i=1}^m{\mathrm{\&lambda;}}_i}\&GreaterEqual;r)$

The transformation matrix that principal component analysis obtains is P _PCA=[u ₁, u ₂..., u _l], to character feature x _i, carry out the l dimension character feature z after dimensionality reduction obtains dimensionality reduction _i=(P _PCA) ^Tx _i, i=1 ..., n, (P _PCA) ^TExpression P _PCATransposed matrix;

Step 42: the character feature behind the training sample dimensionality reduction is carried out linear discriminant analysis, find the solution transformation matrix W:

If character class number to be identified in the recognition system is C, the i class comprises n _iIndividual training sample; Calculate i class character sample characteristic mean μ _iWith all sample characteristics average μ:

${\mathrm{\&mu;}}_i=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_k^i,$ $\mathrm{\&mu;}=\frac{1}{n}\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_k^i,$ $n=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}{n}_i$

Scatter Matrix S between compute classes _bWith Scatter Matrix S in the class _w:

$S_b=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{n_i}{n}({\mathrm{\&mu;}}_i-\mathrm{\&mu;}){({\mathrm{\&mu;}}_i-\mathrm{\&mu;})}^T$

$S_w=\underset{i=1}{\overset{C}{\mathrm{\&Sigma;}}}\left(\frac{n_i}{n}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}(z_k^i-{\mathrm{\&mu;}}_i){(z_k^i-{\mathrm{\&mu;}}_i)}^T\right)$

A transformation matrix W is sought in linear discriminant analysis so that after the conversion between class dispersion as far as possible large, dispersion is as far as possible little in the class simultaneously, utilizes the maximization criterion

$J=\frac{\mathrm{tr}\left(W^T{S}_{b}W\right)}{\mathrm{tr}\left(W^T{S}_{w}W\right)}$ Represent;

Solve linear discriminant analysis by finding the solution the generalized eigenvector problem:

S _bw＝λS _ww

If vectorial w ₁..., w _d..., w _lBe the solution of generalized eigenvector problem, the generalized eigenvalue λ of their correspondences ₁〉=... 〉=λ _d〉=... 〉=λ _l, the solution proper vector of d generalized eigenvector problem forms W, i.e. W=[w before selecting ₁..., w _d].

5. the method for the character recognition of described fusion bianry image and gray level image according to claim 4 is characterized in that described character recognition comprises:

Step 51: design category device

Utilize transformation matrix W to i character type characteristic mean μ _iCarry out dimensionality reduction, and the feature behind the normalization dimensionality reduction ${\mathrm{\&mu;}}_i^{*}=W^T{\mathrm{\&mu;}}_i,$ ${\stackrel{\&OverBar;}{\mathrm{\&mu;}}}_i={\mathrm{\&mu;}}_i^{*}/\sqrt{{\left({\mathrm{\&mu;}}_i^{*}\right)}^T{\mathrm{\&mu;}}_i^{*}};$

Preserve transformation matrix P=WP _PCA, the coding of each character type and corresponding character type feature thereof

In the file in the character recognition storehouse;

Step 52: character recognition

Read coding and the corresponding character type feature thereof of transformation matrix P, each character type in the file from the character recognition storehouse

Bianry image and gray level image to each character to be identified merge, and the image after merging is carried out normalization, carry out the multidimensional characteristic a that feature extraction obtains character picture; Utilize transformation matrix P that the multidimensional characteristic a of character picture is carried out feature b=P after Feature Dimension Reduction obtains dimensionality reduction ^TA, P ^TTransposed matrix for transformation matrix P;

Feature b normalization behind the dimensionality reduction is obtained

Order computation character normalization feature

Normalization center vector with each character type Cosine distance { d _i} _1≤i≤C

$d_i=1-{\stackrel{\&OverBar;}{y}}^T\stackrel{\&OverBar;}{\mathrm{\&mu;}}$

The minimum class of distance is the recognition result of character picture.

Images