CN104537364A - Dollar bill denomination and edition identifying method based on texture analysis - Google Patents

Abstract

The invention discloses a dollar bill denomination and edition identifying method based on texture analysis. By referring to a conventional textural feature LBP obtaining method, noises of an image central point (a central pixel point of each 3*3 image block in each image) are deemed to be only associated with grey values of points in the neighborhood of the image central point, and variance of noise distribution is estimated through changes of the neighborhood points, so that the self-adaptive LBP threshold value is determined; chi-square distance matching comparison is performed on the obtained self-adaptive LBP threshold value of a tested sample and the self-adaptive LBP threshold value of a training sample with the edition and the denomination corresponding to those of the tested sample to finally identify the denomination and the edition of a tested dollar bill. The identifying method comprises the following three steps: (1) performing training sample feature extraction on existing 21 dollar bills which are circulating on the market to obtain the self-adaptive LBP threshold value of a training sample corresponding to each dollar bill; (2) performing tested sample feature extraction on a to-be-detected dollar bill to obtain the self-adaptive LBP threshold value of the tested sample; and (3) performing chi-square distance matching comparison on the obtained self-adaptive LBP threshold value of the tested sample and the self-adaptive LBP threshold value of the training sample with the edition and the denomination corresponding to those of the tested sample, and finally, identifying the denomination and the edition of the tested dollar bill.

Description

A kind of dollar bill denomination based on texture analysis and version recognition methods

Technical field

The present invention relates to financial machine and tool field, especially based on dollar bill denomination and the version recognition methods of texture analysis.Multi-optical spectrum paper money counting machine, cleaning-sorting machine and automatic teller machine can be widely used in financial circulation field to the Classification and Identification of dollar bill.

Background technology

Along with economic globalization development, the world currency is increasing in the circulation of various countries, and as the dollar of the topmost reserve currency in the whole world, its circulation increases by 42% in 5 years.Law-breaker, for seeking exorbitant profit, manufactures the dollar counterfeit money of various version, harm world financial safety.But the dollar bill circulated on the market now has 21 kinds, denomination comprises 1 yuan, 2 yuan, 10 yuan, 20 yuan, 50 yuan, 100 yuan, totally 7 kinds of denominations, and often kind of denomination comprises major part again, microcephaly, color version 3 versions.Because the dollar bill size of various denomination and version is identical, background texture, rich color, cannot determine dollar denominations and version by paper size or color.Traditional recognition technology needs manually to set identification point, and these class methods are low to noise immunity, to illumination not robust, simultaneously also very sensitive to skew, slightly pollutes or offsets, differentiate easily to make mistakes when characteristic area.

The LBP feature of determination LBP operator by obtaining of traditional LBP threshold value, it is the common technology identified image at present, the method is passed through in certain 3 × 3 the regional area in an image block, with central pixel point gray-scale value for threshold value, neighborhood eight pixels are made comparisons with central pixel point respectively, when being greater than central pixel point gray-scale value, value is 1, is less than, and value is 0.And then the weights being multiplied by correspondence position respectively obtain the value of the LBP pattern of central pixel point.The significant challenge that dollar identifies is by the new and old noise caused with stain of dollar bill, the threshold value of tradition LBP is the gray-scale value of central pixel point, noise can make the gray-scale value of central pixel point change, namely threshold value changes, and this can make corresponding LBP value that great variety occurs.Bank note dollar in circulation, owing to there is the difference of new and old edition and nominal value stain, therefore cause during machine examination and produce noise, the gray-scale value of central pixel point is changed, cause the LBP threshold value generation great variety obtained, finally affect the mistake of machine recognition.

Therefore, seeking one, to have noise resisting ability strong, portable high general for multi-optical spectrum paper money counting machine, cleaning-sorting machine, and the method for the identification dollar denominations that is suitable for of automatic teller machine and version has become the expectation of financial circulation field volume.

Summary of the invention

Object of the present invention: being intended to propose one, to have noise resisting ability strong, portable high general for multi-optical spectrum paper money counting machine, cleaning-sorting machine, and the method for the identification dollar denominations that is suitable for of automatic teller machine and version.

This dollar bill denomination based on texture analysis and version recognition methods, with reference to traditional textural characteristics LBP acquisition methods, it is characterized in that: theoretical with reference to Markov random field, the noise of the central pixel point of 3 × 3 image blocks each inside image is considered as only relevant with the gray-scale value of the point in its neighborhood, the variance of estimating noise distribution is carried out with the change of neighborhood point, determine adaptive LBP threshold value thus, and the adaptive threshold LBP feature of training sample corresponding with its version and denomination for the adaptive threshold LBP feature of the tested sample obtained is carried out card side's distance matching ratio pair, the denomination of the tested dollar bill of final identification and version.

This recognition methods comprises following three concrete steps:

1) by carrying out training sample feature extraction to the existing 21 kinds of dollar bills circulated on market, obtaining 21 kinds of each self-corresponding training sample adaptive threshold LBP features of dollar bill;

2) test sample book feature extraction carried out to the dollar bill that need detect, obtain tested sample adaptive threshold LBP feature;

3) by the tested sample adaptive threshold LBP feature obtained with and the corresponding training sample adaptive threshold LBP feature of its version, denomination carry out card side's distance matching ratio pair, the finally denomination of the tested dollar bill of identification and version.

The defining method of described training sample adaptive threshold LBP and test sample book adaptive threshold LBP is as follows:

1) according to formula:

${\mathrm{\σ}}_P=\sqrt{\frac{1}{P-1}\underset{p=0}{\overset{p-1}{\mathrm{\Σ}}}{(g_p-\mathrm{\μ})}^2}---\left(2\right)$

$\mathrm{\μ}=\frac{1}{P}\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}{g}_p---\left(3\right)$

Thr＝g _c-ασ _P(4)

The step represented is carried out;

Wherein:

S represents that S represents sign function, is defined as $s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;1\\ 0.x<0\end{array}\right.;$

G _prepresent that its coordinate is (x _c+ Rcos (2 π p/P), y _c-Rsin (2 π p/P));

P represents that radius is the neighborhood system that the pixel within the scope of R is formed;

σ _prepresent gray-scale value variance; a

A represents LBP parameter;

G _ccentered by pixel gray-scale value.

The training process of described training sample is as follows:

1) the upper left point coordinate x=1 of initialization LBP parameter alpha=0 and image block, y=1.Being fixed α=0, slides from left to right from top to bottom from bank note upper left side in image block position, asks in each position, the adaptive threshold LBP statistic histogram S of each training sample _i, i=1,2 ... 21, and calculate the card side distance χ between every two statistic histograms ²;

${\mathrm{\&chi;}}^2(S_i,S_j)=\underset{k=1}{\overset{B}{\mathrm{\&Sigma;}}}\frac{{(S_{\mathrm{ik}}-S_{\mathrm{jk}})}^2}{S_{\mathrm{ik}}+S_{\mathrm{jk}}}---\left(5\right)$

S _i, i=1,2 ... 21, i is i-th sample, is the adaptive threshold LBP statistic histogram of each training sample;

B is the dimension of adding up the directly side of putting figure;

K is the kth dimension of statistic histogram; We choose the final image block position of the maximum image block position of the mean distance of all distances as current iteration;

2) still image block invariant position, α is [-0.5, + 0.5] mobile in, ask for each α value place, the adaptive threshold LBP statistic histogram of each training sample also calculates corresponding card side's distance (the same), chooses the final α value of the maximum α value of the mean distance of all distances as current iteration;

3) α that asks for of fixing step 2, asks for local optimum image block position by step 1, loop iteration successively, tries to achieve the image block position (x that the overall situation is best _opt,y _opt) and α _optvalue, and calculate and often open training sample at optimized image block position (x _opt, y _opt) and best a _opttime adaptive threshold LBP statistic histogram i=1,2 ... 21; For test process.

As follows to the identifying of tested coin:

1) to any tested coin dollar bill image newly collected, at its corresponding optimized image block position x _opt, y _optwith best a _optplace asks for its adaptive threshold LBP statistic histogram S _test;

2) each tested coin adaptive threshold LBP statistic histogram S is calculated _testto the adaptive threshold LBP statistic histogram of each training sample i=1,2 ... card side's distance of 21 i=1,2 ..., 21, and get its minimum value

3) get and confirm as minimum value the corresponding training sample i at place, this i are denomination or the version classification of the dollar bill of the tested sample that we judge.

This dollar bill denomination based on texture analysis proposed according to above technical scheme and version recognition methods, for the illumination of dollar bill identification, skew and noise problem propose a kind of LBP feature of adaptive threshold, adopt the thought of EM (EM algorithm) algorithm to solve the globally optimal solution of characteristic parameter and localized mass parameter at training process simultaneously, experiment proves that this method discrimination can reach 100%, noise resisting ability is strong, portable high, multi-optical spectrum paper money counting machine can be widely used in, cleaning-sorting machine, automatic teller machine.

Accompanying drawing explanation

Fig. 1 operating process schematic diagram of the present invention;

The computation process schematic diagram of Fig. 2 tradition LBP;

Fig. 3 is the neighborhood system distribution schematic diagram of different P, R value correspondence.

Embodiment

The dollar bill denomination based on texture analysis that the present invention proposes and version recognition methods, its main points are: the adaptive threshold LBP operator adopting a kind of new proposition, extract can obviously distinguish each denomination from the bank note White-light image gathered, the LBP feature of version, comes various different bank note classification according to LBP characteristic statistics histogrammic card side distance.Although what this method adopted is local shape factor, a kind of method that have employed the EM of being similar to algorithm idea asks for two key parameter α of this method and the globally optimal solution of image block position.Be characterized in that the method is classified from the local obtained to bank note on the one hand, machine extracts parameter global optimum solution by autonomous learning on the other hand, do not need artificial setting, compensate for one-sidedness and the subjectivity of artificial setting differential point, thus cause false distinguishing scarce capacity, the problems such as poor stability.

Set forth the present invention further below in conjunction with Figure of description, and provide embodiments of the invention.

This dollar bill denomination based on texture analysis as shown in Figure 1 and version recognition methods, with reference to traditional textural characteristics LBP acquisition methods, the noise of the central pixel point of 3 × 3 image blocks each inside image graph picture is considered as only relevant with the gray-scale value of the point in its neighborhood, carry out the variance of estimating noise distribution with the change of neighborhood point, determine adaptive LBP threshold value thus; And the adaptive threshold LBP of training sample corresponding with its version and denomination for the adaptive threshold LBP of the tested sample obtained is carried out card side's distance matching ratio pair, final denomination and the version identifying tested dollar bill;

This recognition methods comprises following three steps:

1) by carrying out training sample feature extraction to the existing 21 kinds of dollar bills circulated on market, obtaining 21 kinds of each self-corresponding training sample adaptive threshold LBP of dollar bill;

2) test sample book feature extraction carried out to the dollar bill that need detect, obtain tested sample adaptive threshold LBP;

3) by the tested sample adaptive threshold LBP obtained with and the corresponding training sample adaptive threshold LBP of its version, denomination carry out card side's distance matching ratio pair, the finally denomination of the tested dollar bill of identification and version.

In view of a kind of inventive improvements that the present invention is to traditional image recognition technology based on LBP threshold value.

In traditional LBP threshold value determination method (see Fig. 2, Fig. 3 Suo Shi): by certain 3 × 3 the regional area in an image block, with central pixel point gray-scale value for threshold value, neighborhood eight pixels are made comparisons with central pixel point respectively, when being greater than central pixel point gray-scale value, value is 1, is less than, and value is 0.And then the weights being multiplied by correspondence position respectively obtain the value of the LBP pattern of central pixel point.

First define the joint distribution that texture T is the individual pixel gray scale of P+1 (P>0) in regional area, T=t (g _c, g _0,, g _p+1), g _cfor the pixel value of texture region intermediate point, g _p(p=0,2 ... P-1) be respectively with intermediate point in neighborhood, radius is the pixel value of P the point of the upper interval angles 2 π/P of surrounding of R, these points constitute the neighborhood system (Fig. 2 gives different P, neighborhood system during R value) of a circular symmetric.Due to g _pcoordinate be (x _c+ Rcos (2 π p/P), y _c-Rsin (2 π p/P)), so neighborhood point that can not be all can drop on the position of pixel, when neighborhood point does not drop on the position of pixel, generally obtain the gray-scale value of this neighborhood point by bilinear interpolation.

The neighborhood system distribution schematic diagram corresponding for different P, R value that Fig. 3 provides.

The gray-scale value of neighborhood point deducts the gray-scale value of intermediate point, and such local grain T is exactly central point g _cwith central point g _cwith neighborhood point g _i, i=0,1 ..., the joint distribution of P-1 difference

T=t (g _c, g ₀-g _c, g ₁-g _c..., g _p-1-g _c) formula (A)

G _ccentered by pixel gray-scale value, g _i, i=0,1 ..., P-1 is neighborhood territory pixel point gray-scale value, and t is a kind of distribution function.Assuming that the difference g of neighborhood point and intermediate point _i-g _c, i=0,1 ... P-1 and intermediate point g _cvalue be separate, carrying out exchange conversion to above formula can obtain:

T ≈ t (g _c) t (g ₀-g _c, g ₁-g _c. ..., g _p-1-g _c) formula (B)

In order to obtain the gray scale translation invariance of image, we only pay close attention to grey scale change, and texture information is also mainly show in grey scale change, so t (g _c) having little significance to texture analysis, substantially negligible.Texture information is mainly included in associating difference profile below,

T ≈ t (g ₀-g _c, g ₁-g _c..., g _p-1-g _c) formula (C)

LBP adopts g _p-g _csymbol replace actual value, the joint distribution in such above formula can not be subject to the image no matter change of gray scale translation or the impact of number of greyscale levels change.

T≈t(s(g ₀-g _c),s(g ₁-g _c),…,s(g _P-1-g _c))

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;1\\ 0.x<0\end{array}\right.$ Formula (D)

In order to each LBP mechanism can be distinguished, to each s (g _p-g _c) be multiplied by corresponding weights 2 ^p, often kind of local binary patterns can use 0-2 like this ^pthe value of-1 represents,

${\mathrm{LBP}}_{P,R}=\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}s(g_p-g_c)2^p$ Formula (E)

P is neighborhood territory pixel point number, and R is the radius of neighbourhood, and be the neighborhood of 8 for P like this, one has 2 ⁸individual possible values.For the unchanged smooth region of texture, the value of LBP pattern is just 0, and for a speckle regions, the value of LBP pattern just reaches maximum.LBP can detect edge, bright spot, the various texture pattern such as dim spot.The improvement made LBP is a lot, has consistent pattern LBP (Uniform LBP), invariable rotary LBP (Rotation Invariant LBP) etc.

Because LBP has robustness for illumination variation, the dollar bill gradation of image translation affected by daylighting can not affect the LBP value of picture block.Therefore the statistic histogram of all for image block pixel LBP values is used as the textural characteristics of picture block.Can be down to minimum by being offset by image putting position the feature deviation that the image block translation that causes causes like this.

The reason that the technical program releases adaptive threshold LBP is: in the dollar identification of reality, and significant challenge is by the new and old noise caused with stain of dollar bill.The threshold value of tradition LBP is the gray-scale value of central pixel point, and noise can make the gray-scale value of central pixel point change, and namely threshold value changes, and this can make corresponding LBP value that great variety occurs.For this reason, the applicant adopts a kind of adaptive threshold LBP.

According to Markov random field, in image, certain any gray-scale value is only relevant with the gray-scale value of the point in its neighborhood, and noise suppose central point is also relevant with the gray-scale value of the point in its neighborhood, and we carry out with the change of neighborhood point the variance that estimating noise distributes.Suppose that noise profile obeys Gaussian distribution N (0 a, σ ²), its variance neighborhood territory pixel point gray-scale value variances sigma _pbe similar to, μ is the average of neighborhood point.

${\mathrm{\&sigma;}}_P=\sqrt{\frac{1}{P-1}\underset{p=0}{\overset{p-1}{\mathrm{\&Sigma;}}}{(g_p-\mathrm{\&mu;})}^2}$

$\mathrm{\&mu;}=\frac{1}{P}\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}{g}_p$

Therefore in a local neighborhood, the noise n of central pixel point gray-scale value is approximately

n＝ασ _P

α is a constant, and n is noise figure, and the threshold value Thr calculating corresponding adaptive LBP is like this

Thr＝g _c-ασ _P

The value of such local binary patterns is by formula

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;1\\ 0.x<0\end{array}\right.$

G _ccentered by pixel gray-scale value, when α=0, just deteriorate to traditional LBP, α initial value gets 0, its test occurrence be by training sample training obtain.By the high-speed scanning device of multi-optical spectrum paper money counting machine, obtain each totally 21 the image training samples of the dollar bill of different denomination different editions.Training process obtains got 50*50 picture block (due to the treatable data set finite capacity of paper money counter picture processing chip, here too large picture block should not be got) particular location on dollar bill (is considered as upper left point coordinate (x here, y)), with the parameter alpha inside adaptive threshold LBP, Two Variables can not solve simultaneously, and we adopt a kind of mode of the EM of being similar to algorithm to train to solve here.

This training process is as follows:

1, the upper left point coordinate x=1 of initialization LBP parameter alpha=0 and image block, y=1.Being fixed α=0, slides from left to right from top to bottom from bank note upper left side in image block position, asks in each position, the adaptive threshold LBP statistic histogram S of each training sample _i, i=1,2 ... 21, and calculate the card side distance χ between every two statistic histograms ².

${\mathrm{\&chi;}}^2(S_i,S_j)=\underset{k=1}{\overset{B}{\mathrm{\&Sigma;}}}\frac{{(S_{\mathrm{ik}}-S_{\mathrm{jk}})}^2}{S_{\mathrm{ik}}+S_{\mathrm{jk}}}$

S _i, i=1,2 ... 21, i is i-th sample, and be the adaptive threshold LBP statistic histogram of each training sample, B is the dimension of statistic histogram, and k is the kth dimension of statistic histogram.We choose the final image block position of the maximum image block position of the mean distance of all distances as current iteration.

2, still image block invariant position, α is [-0.5, + 0.5] mobile in, ask for each α value place, the adaptive threshold LBP statistic histogram of each training sample also calculates corresponding card side's distance (the same), chooses the final α value of the maximum α value of the mean distance of all distances as current iteration.

3, the α that asks for of fixing step 2, asks for local optimum image block position by step 1.Loop iteration successively, tries to achieve the image block position (x that the overall situation is best _opt,y _opt) and α _optvalue, and calculate and often open training sample at optimized image block position (x _opt, y _opt) and best a _opttime adaptive threshold LBP statistic histogram i=1,2 ... 21.For follow-up test identifying.

The operating process of the inventive method is described in detail below in conjunction with a kind of specific embodiments adopted

If get the dollar bill one totally 21 of the various version of various denomination, be considered as 12 class dollar bills here, obtains its gray level image, actual recognition methods is expressed as follows with reference to training process above:

The first step is the particular location (x obtaining taking image block by training _opt,y _opt) and adaptive LBP parameter a _opt.Remove outside the adaptive threshold of our proposition, that this programme is selected is consistent LBP (Uniform LBP).

Second step, by these 21 different training bank note at the global optimum position (x that upper step is asked for _opt,y _opt) place takes the image block of 50*50, and use and above walk the auto-adaptive parameter a asked for _opt, the adaptive threshold LBP characteristic statistics histogram of each image block is obtained.

3rd step, newly adds arbitrarily a dollar bill, at optimal location (x _opt,y _opt) place takes the image block of 50*50, the same auto-adaptive parameter a using the first step to ask for _optthe adaptive threshold LBP characteristic statistics histogram of this image block is obtained, and mate one by one with the adaptive threshold LBP statistic histogram of second 21 class asked for, the classification of the training dollar bill the most similar to newly adding dollar bill is judged as the classification newly adding dollar bill.

Test process

1, when newly collecting any tested dollar bill image, we are at optimized image block position x _opt, y _optwith best a _optplace asks for its adaptive threshold LBP statistic histogram S _test

2, S is calculated _testto the adaptive threshold LBP statistic histogram of the training sample of each correspondence i=1,2 ... card side's distance of 21 i=1,2 ..., 21, and get its minimum value

3, minimum value is got corresponding training sample i, the i at place are the classification of the test sample book that we judge.

Show that the recognition methods of the dollar bill that the present invention proposes has following characteristics through actual experiment:

1, global optimum's parametric solution is carried out to banknote image, local shape factor, avoid differentiating some region detection and the discriminating power that causes is not enough;

2, training sample quantity demand of the present invention is few, and every class training sample only needs one, but can obtain the discrimination of 100%.

3, machine is asked for by the optimized parameter being similar to EM algorithm, makes up the subjectivity of artificial selection parameter;

4, LBP has robustness to illumination, and adaptive threshold LBP in this paper has robustness to noise, and the use of statistic histogram can solve the deviation that small translation brings.

5, the physical significance of card side's distance coupling is: it should be different for increasing and reduce certain statistic and increase and reduce identical statistic to the impact that feature decision causes in low frequency dimension in high frequency dimension.Adopt card side's distance can increase the identification of LBP statistical nature, in experimental result, card side's distance matching ratio Euclidean distance matching effect is good.

Claims (4)

1. the dollar bill denomination based on texture analysis and version recognition methods, with reference to traditional textural characteristics LBP acquisition methods, the noise of the central pixel point of 3 × 3 image blocks each inside image is considered as only relevant with the gray-scale value of the point in its neighborhood, carry out the variance of estimating noise distribution with the change of neighborhood point, determine adaptive LBP threshold value thus; And the adaptive threshold LBP of training sample corresponding with its version and denomination for the adaptive threshold LBP of the tested sample obtained is carried out card side's distance matching ratio pair, final denomination and the version identifying tested dollar bill;

This recognition methods comprises following three tool steps:

1) by carrying out training sample feature extraction to the existing 21 kinds of dollar bills circulated on market, obtaining 21 kinds of each self-corresponding training sample adaptive threshold LBP of dollar bill;

2) test sample book feature extraction carried out to the dollar bill that need detect, obtain test sample book adaptive threshold LBP;

3) training sample adaptive threshold LBP corresponding with its version and denomination for the test sample book adaptive threshold LBP of acquisition is carried out card side's distance matching ratio pair, final denomination and the version identifying tested dollar bill.

2. a kind of dollar bill denomination based on texture analysis and version recognition methods as claimed in claim 1, is characterized in that: the defining method of described training sample adaptive threshold LBP and test sample book adaptive threshold LBP is as follows:

According to formula:

Thr＝g _c-ασ _P

(4)

The step represented is carried out;

Wherein:

S represents sign function,

μ represents the average of neighborhood point;

G _prepresent that its coordinate is (x _c+ Rcos (2 π p/P), y _c-Rsin (2 π p/P));

P represents that radius is the neighborhood system that the pixel within the scope of R is formed;

σ _prepresent gray-scale value variance; a

A represents LBP parameter;

G _ccentered by pixel gray-scale value.

3. a kind of dollar bill denomination based on texture analysis and version recognition methods as claimed in claim 1, is characterized in that: the training process of described training sample is as follows:

1) the upper left point coordinate x=1 of initialization LBP parameter alpha=0 and image block, y=1.Being fixed α=0, slides from left to right from top to bottom from bank note upper left side in image block position, asks in each position, the adaptive threshold LBP statistic histogram S of each training sample _i, i=1,2 ... 21, and calculate the card side distance χ between every two statistic histograms ²;

S _i, i=1,2 ... 21, i is i-th sample, is the adaptive threshold LBP statistic histogram of each training sample;

B is the dimension of adding up the directly side of putting figure;

K is the kth dimension of statistic histogram; We choose the final image block position of the maximum image block position of the mean distance of all distances as current iteration;

2) still image block invariant position, α is [-0.5, + 0.5] mobile in, ask for each α value place, the adaptive threshold LBP statistic histogram of each training sample also calculates corresponding card side's distance (the same), chooses the final α value of the maximum α value of the mean distance of all distances as current iteration;

3) α that asks for of fixing step 2, asks for local optimum image block position by step 1, loop iteration successively, tries to achieve the image block position (x that the overall situation is best _opt,y _opt) and α _optvalue, and calculate and often open training sample at optimized image block position (x _opt, y _opt) and best a _opttime adaptive threshold LBP statistic histogram i=1,2 ... 21; For test process.

4. a kind of dollar bill denomination based on texture analysis and version recognition methods as claimed in claim 1, is characterized in that: as follows to the identifying of tested coin:

1) to any tested coin dollar bill image newly collected, at its corresponding optimized image block position x _opt, y _optwith best a _optplace asks for its adaptive threshold LBP statistic histogram S _test;

2) each tested coin adaptive threshold LBP statistic histogram S is calculated _testto the adaptive threshold LBP statistic histogram of each training sample i=1,2 ... card side's distance of 21 i=1,2 ..., 21, and get its minimum value

3) get and confirm as minimum value the corresponding training sample i at place, this i are denomination or the version classification of the dollar bill of the tested sample that we judge.