CN102496021A - Wavelet transform-based thresholding method of image - Google Patents

Abstract

The invention provides a wavelet transform-based thresholding method of an image, and belongs to the field of image processing. In the method, a gray scale of a complete natural scenery image is subjected to wavelet decomposition by utilizing the excellent denoising characteristic of wavelet; foreground characters in the image are removed as noise by virtue of low-pass filter so as to acquire approximate background distribution and foreground distribution; a global threshold is calculated according to the foreground distribution, the global threshold and the background distribution are superposed to form a local threshold which is finally used for image thresholding. The thresholding method provided by the invention can be used for quickly and effectively separating the character part as the foreground to eliminate interference of a complex background, and provides advantages to subsequent character cutting and identifying work. According to the thresholding method provided by the invention, the problem of OCR (optical character recognition) of the natural scenery image can be effectively solved.

Description

Image binaryzation method based on wavelet transformation

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of image binaryzation method based on wavelet transformation.

Background technology

For the image that contains literal, the purpose of binaryzation is normally separated word segment as prospect.The effect of binaryzation directly has influence on follow-up literal and cuts apart and identification.Than file and picture binary coding method, natural scene image is carried out binaryzation require adaptation of methods property stronger, and can handle multiple complex situations simultaneously.

Research to image binaryzation in recent years deepens continuously, but also considerably less to the research of natural scene image binaryzation specially, only has the people that the natural scene image of monocase has been carried out binaryzation.But for a complete natural scene image, how navigating to character area itself is exactly a difficult problem.Therefore how research is carried out binaryzation to complete image and is had more universal significance.

Wavelet theory (Wavelet theory) is considered to the important breakthrough on mathematical analysis and method in recent years, under the scientist of different ambits makes joint efforts, nowadays solid mathematical theory basis and broad application background has been arranged.At mathematics circle, wavelet analysis is counted as the milestone on the Fourier analysis development history, and it is the perfect adaptation of functional analysis, Fourier analysis, batten analysis, harmonic analysis, numerical analysis.The place that wavelet analysis is superior to the Fourier conversion is; It all has good localization property in time domain and frequency domain; Through changing sampling step length, can focus on any details of object, make people both can see " forest "; Can see " trees " again, so be called as " school microscop ".Wavelet transformation use sharp-pointed gradually temporal resolution for the radio-frequency component of signal in case move nearly observation signal become branch soon, for low-frequency component use sharp-pointed gradually frequency resolution in case move the remote viewing signal become branch (overall variation trend) slowly.The signal analysis representation feature of small echo this " not only seeing trees but also see forest " is very effective to analyzing non-stationary signal.At present, wavelet transformation is widely used in the every field of signal Processing: handle like voice signal, and Digital Image Processing, Digital Video Processing, nonlinear properties processing etc. have become strong tool in scientific research and the practical application at present.

Wavelet filtering is the important application of wavelet transformation in image processing field.The basic thought of wavelet filteration method is: through the multilayer wavelet transformation, represent the absolute value of wavelet coefficient of original image information bigger to original image, and represent the absolute value of wavelet coefficient of noise signal less relatively.Through threshold value is set, with absolute value less than or filter greater than the wavelet coefficient of threshold value, thereby reach the effect of filtering.This research utilizes the good denoising characteristic of small echo just, with the prospect literal in the image as noise removal.Literal identification in the natural scene image can be by means of existing OCR technology; But different with document is; This literal is embedded in the middle of the complex background, and how eliminating the complex background interference better is the key issue that the natural scene image binaryzation will solve.

Summary of the invention

The objective of the invention is to be directed against specially the research of natural scene image binaryzation, further explore how complete image is carried out binaryzation, propose a kind of image binaryzation method based on wavelet transformation.

The present invention utilize wavelet transformation with the prospect literal in the image as noise removal; Thereby background distributions that obtains being similar to and prospect distribute; Form local threshold according to prospect Distribution calculation overall situation binary-state threshold and with global threshold and background distributions stack back again, finally be used for image binaryzation.

The concrete generative process of the inventive method comprises the steps:

Step 1, read in a width of cloth natural scene coloured image, be converted into gray-scale map.

Step 2, background distributions are similar to, and earlier gray-scale map are made L layer wavelet decomposition, obtain the detail coefficients of L layer approximation coefficient LL and three directions, are respectively level detail coefficient HL, vertical detail coefficient LH and diagonal detail coefficient HH.According to a large amount of experiments, it is best that the decomposition number of plies L of small echo gets 6 layers of effect; Pass through the image of the level and smooth word segment of LPF again, and do 1 layer of wavelet reconstruction, obtain the thumbnail of background distributions; Utilize image interpolation that the background distributions thumbnail is amplified to original image size at last, promptly obtain the background distributions figure that is similar to.

Step 3, prospect distribute approximate, and the difference image that the former gray-scale map in background distributions figure and the step 1 is obtained as difference operation is the prospect distribution plan.

Step 4, the threshold value of choosing on a kind of overall binarization method calculating difference image are global threshold.

Step 5, with the background distributions figure stack that global threshold and step 2 obtain, can obtain the binaryzation local threshold of each pixel in the former gray-scale map.

Step 6, the binary-state threshold that obtains according to step 5 are bianry image with the former greyscale image transitions in the step 1.

Wherein, step 1 is described transfers coloured image to gray-scale map GRAY and adopts following formula:

GRAY(x，y)＝0.2989*R(I(x，y))+0.5870*G(I(x，y))+0.1140*B(I(x，y))

Wherein, R (.), G (.) and B (.) represent to get redness, green and blue component respectively.

The approximate employing of the described background distributions of step 2 anti-symmetrical bi-orthogonal wavelet is done after the multilayer wavelet decomposition to gray level image; According to the visu thrink method setting threshold that proposes by Dohono; Only the Word message in the top decomposition result is handled, then top decomposition result is carried out reconstruct, get rid of the interference of noise and contextual factor; Adopt bicubic interpolation that the background thumbnail that reconstruct obtains is amplified to original image size at last, promptly obtain the background distributions figure BG that is similar to.

The described prospect of step 3 distributes approximate owing to consider that it is that image is done smoothing processing in essence that the small echo filter is made an uproar; Some pixel (the especially background around the literal) can deepen (it is big that the value of pixel becomes) when filter is made an uproar; When making difference operation, negative value might occur, all directly be changed to 0 with this situation that negative value occurs this moment; Therefore, adopt following formula calculating prospect distribution FG:

$\mathrm{FG}(x,y)=\left\{\begin{array}{c}\mathrm{BG}(x,y)-\mathrm{GRAY}(x,y),\mathrm{ifGRAY}(x,y)<\mathrm{BG}(x,y)\\ 0,\mathrm{otherwise}\end{array}\right..$

The described global threshold of step 4 calculates, and adopts fairly simple global threshold computing method: the average μ and the standard deviation sigma of all pixels during at first calculating prospect distributes; Again [| μ-σ |, | μ+σ |] the global threshold GT that looks on the interval gray-scale value to distribute as prospect with minimum projection's number, its computing method shown in following formula,

$\mathrm{\&mu;}=\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*H\left(i\right)/(M*N)$

$\mathrm{\&sigma;}=\sqrt{\frac{1}{M*N-1}\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*{(H\left(i\right)-\mathrm{\&mu;})}^2}$

$\mathrm{GT}=\arg \underset{i=\mathrm{\&mu;}-\mathrm{\&sigma;}}{\overset{\mathrm{\&mu;}+\mathrm{\&sigma;}}{\min }}\left(H\left(i\right)\right)$

Wherein, M and N are respectively the line number and the columns of prospect distribution plan, and H (i) expression prospect distribution grey level histogram is calculated as follows:

$H\left(i\right)=\underset{y=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}A(i,\mathrm{FG}(x,y)),i\&Element;\left[\mathrm{0,255}\right],A(m,n)=\left\{\begin{array}{c}1,\mathrm{ifm}=\mathrm{<figure-callout\; id="0"\; label="n"\; filenames="HDA0000111442270000012.png,HDA0000111442270000013.png"\; state="\{\{state\}\}">n</figure-callout>}\\ 0,\mathrm{otherwise}\end{array}\right..$

The described local threshold of step 5 calculates according to following formula global threshold and background distributions is superposeed, obtain each pixel in the former gray-scale map (x, binaryzation local threshold LT y) (x, y),

LT(x，y)＝GT+BG(x，y)。

The local binary-state threshold that the binaryzation of the described gray level image of step 6 calculates according to step 5; To each the pixel (x in the former gray-scale map; Y), if gray-scale value less than the local binary-state threshold LT of this point (x, y); Then this point is turned to the foreground point by two-value, otherwise this point is turned to background dot by two-value.

Advantage of the present invention and good effect:

The method that the present invention proposes; Gray-scale map to complete natural scene image carries out wavelet decomposition; The prospect that the filtering literal obtains being similar to distributes, and forms the local threshold that finally is used for image binaryzation according to prospect Distribution calculation overall situation binary-state threshold and with global threshold and background distributions stack back again.In order to verify the validity of the inventive method, chosen Otsu, Pavlidis, Liu, Bernsen, Niblack, Sauvola, Gatos, Multi-Scale, 9 kinds of important binarization methods such as block compare test.These binarization methods had both comprised global approach, and partial approach is also arranged; The existing generally acknowledged method of similar research that obtains early also has the up-to-date proposition binarization method of researcher.When selecting, method covers comprehensively.Test is carried out on the ICDAR2003 sample set, and the result shows that the inventive method all is superior to above-mentioned 9 kinds of methods on recall rate, accuracy rate, three indexs of F-measure, and the processing time of single image about 1 second, satisfy application request.It is thus clear that the binarization method based on wavelet transformation that the present invention proposes is adapted to the complicated scene image of background more.

Description of drawings

Fig. 1 is the binaryzation process synoptic diagram based on wavelet transformation;

Fig. 2 is the row extended matrix;

Fig. 3 is the partiting row sampling matrix.

Fig. 4 is a binaryzation process synoptic diagram, wherein, (a) is the gray-scale map after the conversion, (b) individual layer wavelet decomposition result, (c) individual layer wavelet decomposition figure, (d) background distributions figure, (e) prospect distribution plan, (f) prospect distribution grey level histogram, (g) binary map.

Embodiment

Fig. 1 has provided idiographic flow of the present invention, combines embodiment of the invention further explain at present:

1. coloured image transfers gray level image to

Get a fabric width and the high colored natural scene image I that is respectively W=388 and H=543, transfer it to gray level image GRAY according to following formula earlier.Have for

y ∈ [1, H]:

GRAY(x，y)＝0.2989*R(I(x，y))+0.5870*G(I(x，y))+0.1140*B(I(x，y))

Wherein, R (.), G (.) and B (.) represent to get redness, green and blue component respectively.

Gray-scale map after the conversion is shown in (a) among Fig. 4.

2. background distributions is approximate

Earlier gray level image GRAY is made L layer wavelet decomposition, obtain the detail coefficients of L layer approximation coefficient LL and three directions, be respectively level detail coefficient HL, vertical detail coefficient LH and diagonal detail coefficient HH, shown in (b) among Fig. 4.According to a large amount of experiments, it is best that the decomposition number of plies L of small echo gets 6 layers of effect, adopts anti-symmetrical bi-orthogonal wavelet in the present embodiment, and decomposing number of plies L is 6; Pass through the image of the level and smooth word segment of LPF again, and do 1 layer of wavelet reconstruction, obtain the thumbnail of background distributions; Utilize image interpolation that the background distributions thumbnail is amplified to original image size at last, promptly obtain the background distributions figure BG that is similar to, shown in (d) among Fig. 4.The concrete disposal route in each step is following:

1) wavelet decomposition

Gray level image is carried out 1 layer of wavelet decomposition, be about to the detail coefficients that gray level image is decomposed into approximation coefficient and three directions, shown in (b) among Fig. 4; It is exactly that the approximation coefficient LL among the L-1 layer result is decomposed that gray level image is made L layer wavelet decomposition, obtains new approximation coefficient and detail coefficients.The present invention utilizes anti-symmetrical bi-orthogonal wavelet to make wavelet decomposition; Low pass that it is corresponding and Hi-pass filter matrix of coefficients are respectively

and

and establish that matrix f to be decomposed is of a size of M*N in the present embodiment; M and N be the height and width of correspondence image respectively; M=H=543, N=W=388.The decomposition step of wavelet transformation is following:

Fig. 2 is the row extended matrix

Fig. 3 is the partiting row sampling matrix

Respectively expand the matrix f that row obtain M capable (N+2) row by following formula about with matrix f when decomposing (1 layer f=GRAY) _Ar:

$f_{\mathrm{ar}}=f*{\left(E_r^N\right)}^T$

We get the partial pixel dot matrix f among the matrix f to be decomposed ^*Operation is done and being specified, the process for the row expansion as follows, below the partial pixel dot matrix f that all takes out with the symbolic representation of upper right corner band " * " ^*Processing procedure.

$f^{*}=\left[\begin{array}{cccc}97& 96& 95& 94\\ 97& 96& 95& 94\\ 96& 96& 95& 94\end{array}\right]$

${f_{\mathrm{ar}}}^{*}=\left[\begin{array}{cccccc}97& 97& 96& 95& 94& 94\\ 97& 97& 96& 95& 94& 94\\ 96& 96& 96& 95& 94& 94\end{array}\right]$

Press following formula with f _ArRespectively with

With

Make convolution, obtain the matrix f of M capable (N+1) row _HAnd f _G:

$f_H=f_{\mathrm{Ar}}\&CircleTimes;{\tilde{H}}^T,$ Promptly $f_H(x-2,y-1)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{Ar}}(k,y-1)\tilde{H}(x-k,1)$

$f_G=f_{\mathrm{Ar}}\&CircleTimes;{\tilde{G}}^T,$ Promptly $f_G(x-2,y-1)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{Ar}}(k,y-1)\tilde{G}(x-k,1)$

${f_H}^{*}=\left[\begin{array}{ccccc}137.2& 136.5& 135.1& 133.6& 132.9\\ 137.2& 136.5& 135.1& 133.6& 132.9\\ 135.8& 135.8& 135.1& 133.6& 132.9\end{array}\right]$

${f_G}^{*}=\left[\begin{array}{ccccc}0& 0.7& 0.7& 0.7& 0\\ 0& 0.7& 0.7& 0.7& 0\\ 0& 0& 0.7& 0.7& 0\end{array}\right]$

Get f respectively by following formula _HAnd f _GIn even column and do the row expansion, obtain the matrix f of (M+2) row (N+1)/2 row _EHAnd f _EG:

$f_{\mathrm{EH}}=E_r^M*f_H*{\left(S_r^{N+1}\right)}^T$

${f_{\mathrm{EH}}}^{*}=\left[\begin{array}{cc}136.5& 133.6\\ 136.5& 133.6\\ 136.5& 133.6\\ 135.8& 133.6\\ 135.8& 133.6\end{array}\right]$

$f_{\mathrm{EG}}=E_r^M*f_G*{\left(S_r^{N+1}\right)}^T$

${f_{\mathrm{EG}}}^{*}=\left[\begin{array}{cc}0.7& 0.7\\ 0.7& 0.7\\ 0.7& 0.7\\ 0& 0.7\\ 0& 0.7\end{array}\right]$

Press following formula with f _EHAnd f _EGRespectively with

With

Make convolution, can obtain the matrix of four (M+1) row ((N+1)/2) row:

$\mathrm{MA}=f_{\mathrm{EH}}\&CircleTimes;\tilde{H},$ Promptly $\mathrm{MA}(x,y)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{EH}}(k,y-1)\tilde{H}(x-k,1)$

${\mathrm{MA}}^{*}=\left[\begin{array}{cc}193.0& 189\\ 193.0& 189\\ 192.5& 189\\ 192.0& 189\end{array}\right]$

${\mathrm{MD}}^H=f_{\mathrm{EH}}\&CircleTimes;\tilde{G},$ Promptly $M{D}^H(x,y)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{EH}}(k,y-1)\tilde{G}(x-k,1)$

${\mathrm{MD}}^{H*}=\left[\begin{array}{cc}0& 0\\ 0& 0\\ 0.5& 0\\ 0& 0\end{array}\right]$

${\mathrm{MD}}^V=f_{\mathrm{EG}}\&CircleTimes;\tilde{H},$ Promptly $M{D}^V(x,y)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{EG}}(k,y-1)\tilde{H}(x-k,1)$

${\mathrm{MD}}^{V*}=\begin{array}{}\left[\begin{array}{cc}1.0& 1.0\\ 1.0& 1.0\\ 0.5& 1.0\\ 0& 1.0\end{array}\right]\end{array}$

${\mathrm{MD}}^D=f_{\mathrm{EG}}\&CircleTimes;\tilde{G},$ Promptly $M{D}^D(x,y)=\underset{k=x-2}{\overset{x-1}{\mathrm{\&Sigma;}}}{f}_{\mathrm{EG}}(k,y-1)\tilde{G}(x-k,1)$

${\mathrm{MD}}^{D*}=\left[\begin{array}{cc}0& 0\\ 0& 0\\ 0.5& 0\\ 0& 0\end{array}\right]$

Wherein, k is the filtering window span.

Press following formula again to matrix M A, MD ^H, MD ^V, MD ^DAt a distance from the row sampling, can obtain 1 layer of wavelet decomposition result, shown in (b) among Fig. 4, LL is the approximation coefficient of the 1st layer of wavelet decomposition, and HL is that level detail coefficient, LH are the vertical detail coefficient, and HH is the diagonal detail coefficient.Each matrix of consequence is ((M+1)/2) row ((N+1)/2) row, is followed successively by A, the D of matrix of the partial pixel point composition of taking-up from left to right ^H, D ^VAnd D ^D:

$A=\mathrm{MA}*{\left(S_r^{N+1}\right)}^T,$ $D^H={\mathrm{MD}}^H*{\left(S_r^{N+1}\right)}^T,$ $D^V={\mathrm{MD}}^V*{\left(S_r^{N+1}\right)}^T,$ $D^D={\mathrm{MD}}^D*{\left(S_r^{N+1}\right)}^T,A,D^H,$ D ^V, D ^DUpper left, upper right, down left and lower right-most portion, i.e. LL, HL, LH, HH shown in (b) 1 layer of wavelet decomposition result among Fig. 4 in the difference corresponding diagram 4 in (c).

$A^{*}=\left[\begin{array}{cc}193.0& 189\\ 192.0& 189\end{array}\right]$ $D^{H*}=\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right]$ $D^{V*}=\left[\begin{array}{cc}1.0& 1.0\\ 0& 1.0\end{array}\right]$ $D^{D*}=\left[\begin{array}{cc}0& 0\\ 0& 0\end{array}\right]$

By above-mentioned steps A is carried out wavelet decomposition, promptly obtain 1 layer of wavelet decomposition result, the rest may be inferred, can calculate 6 layers of wavelet decomposition result.Because the gray-scale map text color is shallow than background color, need in the present embodiment doing wavelet decomposition after its inverse again.

2) wavelet filtering

The wavelet filtering process be exactly with among the wavelet filtering result greater than threshold value T _nDetail coefficients be changed to 0, keep less than T _nValue, that is:

$D^H(x,y)=\left\{\begin{array}{c}{D}^H(x,y),\mathrm{if}{D}^H(x,y)<{\mathrm{<figure-callout\; id="0"\; label="T\; n"\; filenames="HDA0000111442270000012.png,HDA0000111442270000013.png"\; state="\{\{state\}\}">T</figure-callout>}}_n\\ 0,\mathrm{otherwise}\end{array}\right.$

$D^V(x,y)=\left\{\begin{array}{c}{D}^V(x,y),\mathrm{if}{D}^V(x,y)<{\mathrm{<figure-callout\; id="0"\; label="T\; n"\; filenames="HDA0000111442270000012.png,HDA0000111442270000013.png"\; state="\{\{state\}\}">T</figure-callout>}}_n\\ 0,\mathrm{otherwise}\end{array}\right.$

$D^D(x,y)=\left\{\begin{array}{c}{D}^D(x,y),\mathrm{if}{D}^D(x,y)<{\mathrm{<figure-callout\; id="0"\; label="T\; n"\; filenames="HDA0000111442270000012.png,HDA0000111442270000013.png"\; state="\{\{state\}\}">T</figure-callout>}}_n\\ 0,\mathrm{otherwise}\end{array}\right.$

The threshold value the most frequently used according to wavelet filtering, the visu thrink method preset threshold that is promptly proposed by Dohono is handled the Word message in the top decomposition result, word segment.Threshold value calculation method is formula as follows:

T _n＝σ _n*sqrt(2lnN)

Wherein, σ _n=c/0.6745, c are the intermediate value of the absolute value of small echo detail coefficients, N=3*H _L* W _L, H _LAnd W _LBe the row and column that obtains matrix of coefficients after the L layer wavelet decomposition, the threshold value that calculates in the present embodiment is 204.6.Detail coefficients after 6 layers of wavelet decomposition is as follows, and we find after doing filtering according to threshold value that because the partial pixel point that we take out belongs to background pixel, the value of detail coefficients remains less than threshold value the most at last, and word segment will be filtered.

$\begin{array}{ccc}{D}^{H*}=\left[2.1\right]& D^{V*}=\left[7.9\right]& D^{D*}=[-1.2]\end{array}$

3) wavelet reconstruction

Wavelet reconstruction is the inverse operation of wavelet decomposition, to top coefficient of wavelet decomposition A _L,

Utilize coefficient of wavelet decomposition with

With

Associate matrix

With

The convolution sum decomposition result is carried out reconstruct, shown in following formula:

$f_r=F_{H_L}*((E_{H_L}*A\&CircleTimes;H)*{\left(E_{W_L}\right)}^T\&CircleTimes;H^T)*{\left(F_{W_L}\right)}^T$

$+F_{H_L}*((E_{H_L}*D^H\&CircleTimes;G)*{\left(E_{W_L}\right)}^T\&CircleTimes;H^T)*{\left(F_{W_L}\right)}^T$

$+F_{H_L}*((E_{H_L}*D^V\&CircleTimes;H)*{\left(E_{W_L}\right)}^T\&CircleTimes;G^T)*{\left(F_{W_L}\right)}^T$

$+F_{H_L}*((E_{H_L}*D^D\&CircleTimes;G)*{\left(E_{W_L}\right)}^T\&CircleTimes;G^T)*{\left(F_{W_L}\right)}^T$

${f_r}^{*}=\left[\begin{array}{cc}95& 95\\ 95& 95\end{array}\right]$

Wherein,

4) image interpolation

The present invention adopts bicubic interpolation, utilizes and to treat that the gray scale of 16 points is done cubic interpolation around the sampled point, considers that not only the gray scale of 4 direct neighbor points influences, and considers the influence of gray-value variation rate between each adjoint point.The bicubic interpolation algorithm need be chosen the interpolation basis function and come fitting data, and its form is shown below:

$S\left(w\right)=\left\{\begin{array}{cc}1-2{\left|w\right|}^2++{\left|w\right|}^3& \left|w\right|<1\\ 4-8\left|w\right|+5{\left|w\right|}^2-{\left|w\right|}^3& 1\&le;\left|w\right|<2\\ 0& \left|w\right|\&GreaterEqual;2\end{array}\right.$

Carry out bicubic interpolation according to following formula, can obtain the image array f after the interpolation _R, promptly approximate background distributions figure BG is shown in (d) among Fig. 4.f _RIdentical with the original image matrix size:

f _R(i+u，j+v)＝A*B*C

${f_R}^{*}=\left[\begin{array}{cccc}95& 95& 95& 95\\ 95& 95& 95& 95\\ 95& 95& 95& 95\end{array}\right]$

Wherein, A, B, C are matrix, and its form is following:

A＝[S(1+u)?S(u)?S(1-u)?S(2-u)]

$B=\left[\begin{array}{cccc}f(i-1,j-2)& f(i,j-2)& f(i+1,j-2)& f(i+2,j-2)\\ f(i-1,j-1)& f(i,j-1)& f(i+1,j-1)& f(i+2,j-1)\\ f(i-1,j)& f(i,j)& f(i+1,j)& f(i+2,j)\\ f(i-1,j+1)& f(i,j+1)& f(i+1,j+1)& f(i+1,j+1)\end{array}\right]$

C＝[S(1+v)?S(v)?S(1-v)?S(2-v)] ^T

3. prospect distributes approximate

Background distributions figure BG and former gray-scale map GRAY are made difference operation can obtain prospect distribution plan FG.Obviously, in prospect distributed, the value of former gray-scale map background pixel point should level off to 0, the value of foreground pixel point should be away from 0.Because it is that image is done smoothing processing in essence that the small echo filter is made an uproar; Some pixel (the especially background around the literal) can deepen (it is big that the value of pixel becomes) when filter is made an uproar; When making difference operation, negative value might occur, all directly be changed to 0 with this situation that negative value occurs this moment.Therefore, the prospect that finally is calculated as follows distributes:

$\mathrm{FG}(x,y)=\left\{\begin{array}{c}\mathrm{BG}(x,y)-G(x,y),\mathrm{ifG}(x,y)<\mathrm{BG}(x,y)\\ 0,\mathrm{otherwise}\end{array}\right.$

The prospect distribution plan that calculates in the present embodiment is obtained as difference operation according to following formula with (d) background distributions figure by (a) former gray-scale map among Fig. 4 shown in (e) among Fig. 4.

4. global threshold calculates

In prospect distributed, a large amount of background pixel point values were 0 or approach 0 that the value of foreground pixel point is then away from 0.Because in the prospect distribution plan, preceding background difference is bigger, therefore adopt fairly simple global threshold computing method here: the average μ and the standard deviation sigma of all pixels during at first calculating prospect distributes; Again [| μ-σ |, | μ+σ |] the global threshold GT that looks on the interval gray-scale value to distribute as prospect with minimum projection's number, its computing method are shown in following formula.

$\mathrm{\&mu;}=\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*H\left(i\right)/(M*N)$

$\mathrm{\&sigma;}=\sqrt{\frac{1}{M*N-1}\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*{(H\left(i\right)-\mathrm{\&mu;})}^2}$

$\mathrm{GT}=\arg \underset{i=\mathrm{\&mu;}-\mathrm{\&sigma;}}{\overset{\mathrm{\&mu;}+\mathrm{\&sigma;}}{\min }}\left(H\left(i\right)\right)$

H (i) expression prospect distribution grey level histogram, its computing method are shown below:

$H\left(i\right)=\underset{y=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}A(i,\mathrm{FG}(x,y)),i\&Element;\left[\mathrm{0,255}\right]$

Wherein, $A(m,n)=\left\{\begin{array}{c}1,\mathrm{Ifm}=\mathrm{<figure-callout\; id="0"\; label="n"\; filenames="HDA0000111442270000012.png,HDA0000111442270000013.png"\; state="\{\{state\}\}">n</figure-callout>}\\ 0,\mathrm{Otherwise}\end{array}\right..$

Present embodiment calculates prospect distribution grey level histogram according to above-mentioned formula, shown in (f) among Fig. 4.The average μ of all pixels=15 in the prospect distribution plan, standard deviation sigma=30, [| μ-σ |, | μ+σ |] the global threshold GT=45 that finds on the interval, such as among the figure (f) mark.

5. local threshold calculates

By following formula global threshold GT and background distributions BG are superposeed, can obtain each pixel among the former gray-scale map GRAY (x, binaryzation local threshold LT y) (x, y).The place that literal is arranged in former gray-scale map, its local threshold is bigger, makes can well be kept in the binary image Chinese words; And be the place of background in former gray-scale map, its local threshold is less, makes most of background in binary image, remain background.

LT(x，y)＝GT+BG(x，y)

6. binaryzation

Convert gray level image GRAY into bianry image BW by following formula.(x, y), (x, y), then this point is turned to the foreground point by two-value, otherwise this point is turned to background dot by two-value if gray-scale value is less than the local binary-state threshold LT of this point for each pixel among the former gray-scale map GRAY.The binary map that (g) is converted to for (a) among Fig. 4.

$\mathrm{BW}(x,y)=\left\{\begin{array}{c}1,\mathrm{ifGRAY}(x,y)<\mathrm{LT}(x,y)\\ 0,\mathrm{otherwise}\end{array}\right.$

Claims (7)

1. the image binaryzation method based on wavelet transformation is characterized in that this method comprises the steps:

Step 1, read in a width of cloth natural scene coloured image, be converted into gray-scale map;

Step 2, background distributions are similar to, and earlier gray-scale map are made L layer wavelet decomposition, obtain the detail coefficients of L layer approximation coefficient LL and three directions, are respectively level detail coefficient HL, vertical detail coefficient LH and diagonal detail coefficient HH; Pass through the image of the level and smooth word segment of LPF again, and do 1 layer of wavelet reconstruction, obtain the thumbnail of background distributions; Utilize image interpolation that the background distributions thumbnail is amplified to original image size at last, promptly obtain the background distributions figure that is similar to;

Step 3, prospect distribute approximate, and the difference image that the former gray-scale map in background distributions figure and the step 1 is obtained as difference operation is the prospect distribution plan;

Step 4, the threshold value of choosing on a kind of overall binarization method calculating difference image are global threshold;

Step 5, with the background distributions figure stack that global threshold and step 2 obtain, can obtain the binaryzation local threshold of each pixel in the former gray-scale map;

Step 6, the binary-state threshold that obtains according to step 5 are bianry image with the former greyscale image transitions in the step 1.

2. method according to claim 1 is characterized in that step 1 is described and transfers coloured image to gray-scale map GRAY and adopt following formula:

GRAY(x，y)＝0.2989*R(I(x，y))+0.5870*G(I(x，y))+0.1140*B(I(x，y))

Wherein, R (.), G (.) and B (.) represent to get redness, green and blue component respectively.

3. method according to claim 1; It is characterized in that the approximate anti-symmetrical bi-orthogonal wavelet that adopts of the described background distributions of step 2 does after the multilayer wavelet decomposition gray level image; According to the visu thrink method setting threshold that proposes by Dohono; Only the Word message in the top decomposition result is handled, then top decomposition result is carried out reconstruct, get rid of the interference of noise and contextual factor; Adopt bicubic interpolation that the background thumbnail that reconstruct obtains is amplified to original image size at last, promptly obtain the background distributions figure BG that is similar to.

4. method according to claim 1; It is characterized in that the described prospect of step 3 distributes approximate owing to consider that it is that image is done smoothing processing in essence that the small echo filter is made an uproar; Some pixel can deepen pixel when filter is made an uproar value becomes big; Especially the background pixel point around the literal negative value might occur when making difference operation, and all directly be changed to 0 with this situation that negative value occurs this moment; Therefore, adopt following formula to calculate prospect distribution plan FG:

$\mathrm{FG}(x,y)=\left\{\begin{array}{c}\mathrm{BG}(x,y)-\mathrm{GRAY}(x,y),\mathrm{ifGRAY}(x,y)<\mathrm{BG}(x,y)\\ 0,\mathrm{otherwise}\end{array}\right..$

5. method according to claim 1 is characterized in that the described global threshold of step 4 calculates, and adopts fairly simple global threshold computing method: the average μ and the standard deviation sigma of all pixels during at first calculating prospect distributes; Again [| μ-σ |, | μ+σ |] the global threshold GT that looks on the interval gray-scale value to distribute as prospect with minimum projection's number, its computing method shown in following formula,

$\mathrm{\&mu;}=\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*H\left(i\right)/(M*N)$

$\mathrm{\&sigma;}=\sqrt{\frac{1}{M*N-1}\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}i*{(H\left(i\right)-\mathrm{\&mu;})}^2}$

$\mathrm{GT}=\arg \underset{i=\mathrm{\&mu;}-\mathrm{\&sigma;}}{\overset{\mathrm{\&mu;}+\mathrm{\&sigma;}}{\min }}\left(H\left(i\right)\right)$

Wherein, M and N are respectively the line number and the columns of prospect distribution plan, and H (i) expression prospect distribution grey level histogram is calculated as follows:

$H\left(i\right)=\underset{y=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}A(i,\mathrm{FG}(x,y)),i\&Element;\left[\mathrm{0,255}\right],$ $A(m,n)=\left\{\begin{array}{c}1,\mathrm{ifm}=n\\ 0,\mathrm{otherwise}\end{array}\right..$

6. method according to claim 1 is characterized in that the described local threshold of step 5 calculates according to following formula global threshold and background distributions are superposeed, obtain each pixel in the former gray-scale map (x, binaryzation local threshold LT y) (x, y),

LT(x，y)＝GT+BG(x，y)。

7. method according to claim 1; It is characterized in that the local binary-state threshold that the binaryzation of the described gray level image of step 6 calculates according to step 5, to each pixel in the former gray-scale map (x, y); If gray-scale value is less than the local binary-state threshold LT (x of this point; Y), then this point is turned to the foreground point by two-value, otherwise this point is turned to background dot by two-value.

Images