CN104298989A - Counterfeit identifying method and counterfeit identifying system based on zebra crossing infrared image characteristics - Google Patents

Abstract

The invention relates to RMB counterfeit identifying technologies, and provides a counterfeit identifying method and a counterfeit identifying system based on zebra crossing infrared image characteristics. The counterfeit identifying method comprises the steps of A, acquiring an infrared image of RMB and carrying out preprocessing on the acquired infrared image; B, extracting HOG characteristics of the preprocessed infrared image and selecting HOG characteristics which satisfy preset conditions from the extracted HOG characteristics; and C, carrying out counterfeit identifying according to the selected HOG characteristics. The counterfeit identifying method based on the zero crossing infrared image characteristics acquires a zebra crossing anti-counterfeit image of the RMB by adopting infrared light, and carries out processing and identifying on the zebra crossing anti-counterfeit image by using an improved HOG and a support vector machine, thereby being capable of effectively eliminating noise effects and improving the performance of infrared image counterfeit identifying, and being capable of improving the accuracy in counterfeit identification for authentic and counterfeit banknotes.

Description

Based on false distinguishing method and the system thereof of zebra stripes Infrared Image Features

Technical field

The present invention relates to Renminbi authentication detection technology, particularly relate to the false distinguishing method based on zebra stripes Infrared Image Features and system thereof.

Background technology

The bank note such as 2005 editions 100 yuan, 50 yuan, 20 yuan, 10 yuan adopt the imaging of infrared transmission mode, and safety line region presents zebra stripes security pattern.Because zebra stripes exist along with opened window safety line, invisible in the sunlight, just can display in infrared transmission imaging, imitation difficulty is very big.The height of existing report is imitated in counterfeit money, such as, with the counterfeit money of HB90 and HD90 sequence number beginning, has watermark true to nature, becomes ink, stealthy denomination, magnetic characteristic and ultraviolet feature, do not see the false proof fraud of infrared zebra stripes.The feature can't see under these visible rays, makes infrared image false distinguishing have unique advantage in various paper money discrimination field.

Characteristics of image is certain expression of containing information in image, and characteristics of image can transform to another transform domain from a transform domain and represent.In Images Classification, if the feature space of extraction can find obvious categorised demarcation line, just tagsort can be carried out preferably.In actual applications, select one accurately character representation be the key of dealing with problems.In banknote image classification false distinguishing identification field, need first to carry out feature extraction to banknote image, then complete identification and false distinguishing.

For original target image, carry out feature extraction selection, select the feature with better discrimination, namely in one species, there is similarity, in variety classes, there is otherness.Redundant information and correlative character can be removed by feature extraction.The intrinsic dimensionality extracted is wanted suitably, if dimension is too large, affect training effectiveness, dimension is too small, and between class, difference describes and very fewly affects recognition effect.Therefore, the feature selecting to have better classifying quality in numerous characteristics of image is needed.In addition, different applications, the standard of feature selecting is also different.Target image is subject to the impact of noise and light change, and the features such as its shape, size and brightness can not be expressed simultaneously, need to carry out selecting and combining for concrete problem.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of Renminbi false distinguishing method based on zebra stripes Infrared Image Features and system thereof, is intended to solve existing cash inspecting machine and imitates to high counterfeit money exists identification by mistake problem when differentiating.

The present invention is achieved in that the Renminbi false distinguishing method based on zebra stripes Infrared Image Features, comprises the following steps:

Steps A: the infrared image gathering Renminbi, and the infrared image collected is carried out pre-service;

Step B: pretreated infrared image is carried out HOG feature extraction, and choose from the HOG feature extracted and meet pre-conditioned HOG feature;

Step C: the HOG feature according to choosing out carries out false distinguishing.

Further, described steps A specifically comprises:

Steps A 1: Renminbi image is gathered and locates;

Steps A 2: extract interesting image regions from the image of having good positioning.

Further, described steps A 1 specifically comprises:

Steps A 101: the infrared transmission image I gathering Renminbi front _t(x, y) and infrared external reflection image I _r(x, y);

Steps A 102: by described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x, y) carries out additive operation by following formula and obtains image I (x, y),

Steps A 103: the edge adopting the most dark areas of hough change detection image I (x, y), positions target image according to edge, and the target image of collection is carried out slant correction.

Further, described step B specifically comprises:

Step B1: select suitable HOG characteristic parameter, described interesting image regions is divided into several characteristic blocks, each characteristic block forms by 2 × 2 cell factory, each cell factory is formed by the pixel of 8 × 8;

Step B2: the HOG eigenwert X calculating described each characteristic block _i;

Step B3: by described eigenwert X _ifisher criterion is adopted to calculate Fisher value F _iand the Fisher value F that will calculate _isort;

Step B4: according to ranking results, from all Fisher value F _ithe maximum Fisher value of middle selection is input in SVM classifier as feature, and calculates corresponding classification discrimination R;

Step B5: repeating said steps B4, until R>R _t, wherein, R _tfor the Classification and Identification rate of setting.

Further, described step B2 specifically comprises the following steps:

Step B201: adopt gradient operator [-1,0,1] and [-1,0,1] ^-1calculate gradient direction and the gradient magnitude of each unit, computing formula is: g (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel;

Step B202: the HOG feature of each cell factory is calculated by wherein each pixel Nearest Neighbor with Weighted Voting, utilizes Gauss's weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel;

Step B203: the HOG eigenwert X calculating each characteristic block according to the HOG eigenwert of each cell factory _i, and be normalized.

Further, described step B3 specifically comprises the following steps:

Step B301: according to the within-cluster variance of following formulae discovery i-th class, i=ω _ror ω _c, i=ω _ror ω _c, wherein, ω _rrepresent the kind of genuine note, ω _crepresent the kind of counterfeit money, X represents the HOG feature extracted from each characteristic block, m _irepresentation class ω _ror class ω _csample characteristics mean value;

Step B302: the within-cluster variance sum and the inter _ class relationship that calculate all classes, i=ω _ror ω _c, wherein, S _wrepresent the within-cluster variance sum of all classes, S _brepresent inter _ class relationship, Fisher value F _iequal S _b/ S _w.

Further, described step C specifically comprises:

Step C1: Modling model majorized function, c>=0, (i=1,2 ..., n), restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe relaxation factor, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training;

Step C2: according to the model set up, obtaining decision function is wherein, $b=y_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}y{\mathrm{\α}}_{i}K(X_i\·X),$ α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, $K(X_i,X)=e^{-\mathrm{\γ}{\left|\right|X_i-X\left|\right|}^2};$

Step C3: compare grid data service, genetic algorithm and particle swarm optimization algorithm, draws the parameter of SVM classifier, then from the HOG characteristic value collection of SVM classifier, carries out two quasi-mode identifications according to described parameter.

The present invention also provides a kind of Renminbi identification system based on zebra stripes Infrared Image Features, comprising:

Image pre-processing module, carries out pre-service to the Renminbi infrared image collected;

HOG characteristic extracting module, extracts HOG feature by pretreated infrared image, and chooses from the HOG feature extracted and meet pre-conditioned HOG feature;

Renminbi false distinguishing module, carries out false distinguishing by the HOG feature choosing out.

Further, described image pre-processing module comprises:

Image acquisition and positioning unit, gather the infrared transmission image I in Renminbi front _t(x, y) and infrared external reflection image I _r(x, y); By described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x, y) carries out additive operation by following formula and obtains image I (x, y), with accurate positioning security line position; Adopt the edge of the most dark areas of hough change detection image I (x, y), according to edge, target image is positioned, the target image of collection is carried out Slant Rectify.

Region of interesting extraction unit, the target image according to having good positioning extracts interesting image regions.

Further, described HOG characteristic extracting module comprises:

Area division unit, choose suitable HOG characteristic parameter, described interesting image regions is divided into several characteristic blocks, each characteristic block forms by Unit 2 × 2, and each unit is formed by the pixel of 8 × 8;

HOG eigenwert computing unit, adopts gradient operator [-1,0,1] and [-1,0,1] ^-1calculate gradient direction and the gradient magnitude of each unit, computing formula is: g (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel; The HOG feature of each cell factory is calculated by wherein each pixel Nearest Neighbor with Weighted Voting, utilizes Gauss's weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel; The HOG eigenwert X of each characteristic block is calculated according to the HOG eigenwert of each cell factory _i, and be normalized;

Fisher value F _isequencing unit, by described eigenwert X _ifisher criterion is adopted to calculate Fisher value F _i, and the Fisher value F that will calculate _isort;

HOG eigenwert input block, according to ranking results, from all Fisher value F _ithe Fisher value F that middle selection is maximum _ibe input in SVM classifier as feature, and calculate corresponding classification discrimination R;

END instruction unit, described HOG eigenwert input block often inputs a Fisher value F _i, perform a R and whether be greater than R _tdetection action, until R>R _t, wherein, R _tfor the Classification and Identification rate of setting.

Further, described Renminbi false distinguishing module comprises:

Unit set up by model, Modling model majorized function, $\underset{\mathrm{\ω},b,\mathrm{\ξ}}{\min }\frac{1}{2}{\left|\right|W\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i,C\≥0,(i=\mathrm{1,2},\·\·\·,n),$ Restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe relaxation factor, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training;

Functional Analysis unit, according to the model set up, obtaining decision function is wherein, $b=y_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}y{\mathrm{\α}}_{i}K(X_i\·X),$ α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, $K(X_i,X)=e^{-\mathrm{\γ}{\left|\right|X_i-X\left|\right|}^2};$

Pattern recognition unit, selects SVM classifier parameter, carries out two quasi-mode identifications according to decision function from the HOG characteristic value collection of SVM classifier.

The present invention compared with prior art, beneficial effect is: the described false distinguishing method based on zebra stripes Infrared Image Features adopts the zebra anti-counterfeiting image of infrared light to Renminbi to gather, and use the HOG method improved process zebra anti-counterfeiting image and identify, the method have chosen suitable HOG characteristic parameter and SVM classifier parameter, can effectively affect by stress release treatment, improve the performance of infrared image false distinguishing, can 99.03% be reached to the accuracy rate of true and false coin.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the false distinguishing method that the present invention is based on zebra stripes Infrared Image Features;

Fig. 2 is the area-of-interest schematic diagram of the false proof point of zebra stripes in the present invention;

Fig. 3 is the process flow diagram of HOG feature extraction in false distinguishing method;

Fig. 4 is the relation schematic diagram that characteristic block chooses number and classification accuracy;

Fig. 5 is the experimental result picture utilizing network technique Selection parameter;

Fig. 6 is the experimental result picture utilizing genetic algorithm Selection parameter;

Fig. 7 is the experimental result picture utilizing particle group optimizing method Selection parameter.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

As shown in Figure 1, be the present invention one preferred embodiment, based on the Renminbi false distinguishing method of zebra stripes Infrared Image Features, comprise the following steps: steps A: the infrared image gathering Renminbi, and the infrared image collected is carried out pre-service; Step B: pretreated infrared image is carried out HOG (Histogram of Orient Gradient, gradient orientation histogram) feature extraction, and choose from the HOG feature extracted and meet pre-conditioned HOG feature; Step C: the HOG feature according to choosing out carries out false distinguishing.

When performing steps A, first Renminbi image gathered and locate.Infrared image capturing system collection is utilized often to open the infrared transmission image I in Renminbi front _t(x, y) and infrared external reflection image I _r(x, y), the position gathering Renminbi for twice remains unchanged, and the image collected is gray level image.By described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x, y) is by formula carry out additive operation and obtain image I (x, y), known from above formula, the gray scale of two width images and more than 255 time, value 255; Gray scale and not more than 255 time, get two figure gray scale and.After this additive operation, region the darkest in former figure retains, and other regions all become white background.Safety line is the darkest region in transmittance and reflectance image, therefore, adopts above-mentioned additive operation can accurately navigate to safety line region.Through the edge of the most dark areas of hough change detection image I (x, y), according to edge, target image positioned, rotate, complete the slant correction of the target image of collection.The imaging under white light of zebra stripes anti-counterfeiting image is invisible, but in infrared transmission imaging, is shown as light and shade alternating rectangular block (zebra stripes namely claimed in the application), and this unique point can as the feature distinguishing true and false coin.Safety line is in infrared imaging, and be a black line, through whole zebra stripes pattern, namely zebra stripes pattern is positioned at around safety line.If not find fixing safety line in the process of above-mentioned process image, so this banknote is directly judged to be counterfeit money.Then, from the image of having good positioning, interesting image regions is extracted.Particularly, after the horizontal level of safety line is determined, the horizontal coordinate of ROI (Region of Interest, interesting image regions) central point also determines.A new zebra stripes pattern is merged into after the wide rectangular area of two 16 pixels splits from the left side of safety line and the right respectively.After completing aforesaid operations, the extracted region of 5H × 32 size out as interesting image regions, as shown in Figure 2, for the feature extraction region in paper money discrimination.

A variety of parameter is had to need to select when extracting HOG feature.Wherein have several parameter very important, they respectively: the size of (1) ROI: the size of the area-of-interest extracted from image; (2) size of Block: from the size of ROI region divided block; (3) Block sliding step: utilize the concept of block to carry out overlapping normalization, overcomes illumination variation impact; (4) size of Cell: the cell unit size in Block; (5) size of Bin: the number in statistical gradient direction in a cell unit; (6) the assembled arrangement mode of Block and Cell: multiple cell arrangement mode can be chosen form a Block in ROI interval.

540 samples (500 is genuine note, and 40 is counterfeit money) are adopted to test, wherein 200 genuine note samples and the training of 30 counterfeit money samples.Attempted the arrangement mode of the cell factory (cell) in different parameter sizes and various features block (block), the discrimination obtained the results are shown in Table 1.1.Experiment finds, when ROI region chooses 32 × 128 pixel sizes, choose the cell structure of 2 × 2 in block, when cell chooses 8 × 8 pixel, bin gets 9, and the step-length of sliding shoe chooses the width of a cell factory, and the Classification and Identification rate obtained is the highest.

The relation of each Parameters variation of table 1.1 and discrimination

Shown in Figure 3, when performing step B, first, suitable HOG characteristic parameter is selected according to above-mentioned experimental result, described interesting image regions is divided into several characteristic blocks, and each characteristic block forms by 2 × 2 cell factory, and each cell factory is formed by the pixel of 8 × 8.Then, the HOG eigenwert X of described each characteristic block is calculated _i.Particularly, gradient operator [-1,0,1] and [-1,0,1] is adopted ^-1calculate gradient direction and the gradient magnitude of each cell factory, computing method are: g (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel.The HOG feature of each cell factory is calculated by the Nearest Neighbor with Weighted Voting of each pixel wherein, Gauss's weighted gradient amplitude and tri-linear interpolation methods can be utilized to calculate the weights of each pixel, and calculate the HOG eigenwert X of each characteristic block according to the HOG eigenwert Nearest Neighbor with Weighted Voting of each cell factory _i, and be normalized.Again the normalization of L2 norm is carried out to the histogram vector of each characteristic block (block) wherein, the histogram vectors after the normalization in v representation feature block, || v _k|| represent k-norm calculation, k=1 or 2, ε is a very little constant, prevents infinitely large quantity.The object be normalized each characteristic block is light source change when compensating Gather and input image.Histogram vectors after all normalization being connected into a size is the vector of n × m, and wherein, n represents the dimension of the histogram vectors in each characteristic block, and m represents the number of characteristic block inside interesting image regions.Preferably, m=45, n=36.Vector after series connection is input in SVM (Support Vector Machine, Support Vector Machine) sorter and carries out true and false coin Classification and Identification.But the proper vector of HOG is the vector of higher-dimension.Such as, when bin is set to 9, the Duplication of each piece is set to 0.5, and this HOG intrinsic dimensionality is 45 × 4 × 9=1620.High intrinsic dimensionality make extraction feature, training sample and classification time calculating strength strengthen.Therefore, need to carry out choosing process before HOG feature is input to SVM classifier.Make discovery from observation, the edge direction mainly horizontal and vertical direction of the false proof dot pattern of zebra stripes, can utilize Fisher criterion to remove the HOG feature of redundancy.Survey according to this standard, if this feature is large in the similarity of same group of internal ratio difference group, so this feature has good discrimination.Then carry out feature ordering, feature high for discrimination is picked out as last feature.Particularly, by eigenwert X _ifisher criterion is adopted to calculate, and the Fisher value F that will calculate _isort, computing method are: $m_i=\frac{1}{N_i}\underset{X\∈{\mathrm{\ω}}_i}{\mathrm{\Σ}}X,$ i＝ω _Rorω _C， $S_i=\underset{X\∈{\mathrm{\ω}}_i}{\mathrm{\Σ}}(X-m_i){(X-m_i)}^T,$ i＝ω _Rorω _C， $S_w=\underset{i}{\mathrm{\Σ}}{S}_i,$ I=ω _ror ω _c, wherein, ω _rrepresent the kind of genuine note, ω _crepresent the kind of counterfeit money, X represents the HOG feature extracted from each characteristic block, m _irepresentation class ω _ror class ω _csample characteristics mean value, S _wrepresent the within-cluster variance sum of all classes, S _brepresent inter _ class relationship.F _i=S _b/ S _w, F _iratio larger, the separating capacity of its feature X is stronger.According to ranking results, from all Fisher value F _ithe maximum Fisher value of middle selection is input in SVM classifier as feature, and calculates corresponding classification discrimination R; In the past Fisher value F once remaining again _ithe maximum Fisher value of middle selection is input in SVM classifier as feature, repeats this action, until R>R _t, R _tfor the Classification and Identification rate of setting, stop from remaining Fisher value F _imiddle selection the greater is input to SVM classifier.

Choose after HOG feature is input to SVM classifier, needed to carry out false distinguishing to HOG feature, thus distinguish the true and false of Renminbi.SVM classifier is utilized to carry out the identification of zebra stripes feature.The problem of two quasi-mode identifications in fact to the classification of genuine note and counterfeit money.If given training set be (x1, y1), (x2, y2) ..., (xn, yn) }, wherein xi ∈ Rn is input HOG proper vector, y _i{-1,1} is output vector to ∈.Xi represents the HOG feature of i-th training sample, y _irepresent the class label of i-th training sample, "-1 " represents counterfeit money classification, and " 1 " represents genuine note classification.The target of carrying out true and false coin classification with SVM is, a given training set comprising true and false coin, can find one distinguish two class data and have the lineoid of largest interval with two class data.If this training set can be divided by a lineoid, then this lineoid is WX+b=0, and the parameter W in formula and b determines the position of lineoid, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, and WX is two vectorial inner products.Divide to try to achieve this training set optimization, this problem can be converted into the problem asking optimization lineoid, is thus converted into the problem set up function model and solve.

Concrete false distinguishing method is: first, Modling model majorized function, restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, Subject to y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0

The coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe the relaxation factor considered error in classification and introduce, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training.According to the model set up, can release decision function is wherein, $b=y_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}y{\mathrm{\α}}_{i}K(X_i\·X),$ α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, relatively these three kinds of algorithms of grid data service, genetic algorithm and particle swarm optimization algorithm, show that grid data service is optimum algorithm through comparing, determine the parameter of SVM classifier according to grid data service, then from the HOG characteristic value collection of SVM classifier, carry out two quasi-mode identifications according to described parameter.

As shown in Figures 5 to 7, when adopting grid data service, genetic algorithm and particle swarm optimization algorithm to compare the parameter finding SVM classifier, have employed 230 training samples and test, comparative result is in table 1.2.

Table 1.2 compares based on the parameter of three kinds of algorithms

Above-mentioned three kinds of methods all can obtain the cross validation accuracy rate of 99.5652%, and wherein grid data service is consuming time minimum, and therefore select the method to select the parameter of C-SVM sorter, parametric results is C=1, γ=0.10882.

The present invention also provides a kind of identification system based on zebra stripes Infrared Image Features, comprising: image pre-processing module, carries out pre-service to the Renminbi infrared image collected; HOG characteristic extracting module, extracts HOG feature by pretreated infrared image, and the HOG feature extracted is chosen; Renminbi false distinguishing module, is input in SVM classifier according to the HOG feature choosing out and carries out false distinguishing.

Described image pre-processing module comprises image acquisition and positioning unit and region of interesting extraction unit.Image acquisition and positioning unit are for gathering the infrared transmission image I in Renminbi front _t(x, y) and infrared external reflection image I _r(x, y); And by described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x, y) carries out additive operation by following formula and obtains image I (x, y), with accurate positioning security line position; Also for adopting the edge of the most dark areas of hough change detection image I (x, y), according to edge, target image being positioned, and the target image of collection is carried out slant correction.Region of interesting extraction unit extracts interesting image regions according to the target image of having good positioning.

Described HOG characteristic extracting module comprises area division unit, HOG eigenwert computing unit, Fisher value S _isequencing unit, HOG eigenwert input block and END instruction unit.Described interesting image regions, for choosing suitable HOG characteristic parameter, is divided into several characteristic blocks by area division unit, and each characteristic block forms by Unit 2 × 2, and each unit is formed by the pixel of 8 × 8.HOG eigenwert computing unit adopts gradient operator [-1,0,1] and [-1,0,1] ^-1calculate gradient direction and the gradient magnitude of each unit, computing formula is: $\mathrm{\α}(x,y)={\tan }^1\left[\frac{I(x,y+1)-I(x,y-1)}{I(x+1,y)-I(x-1,y)}\right],$ G (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel; The HOG feature of each cell factory is calculated by wherein each pixel Nearest Neighbor with Weighted Voting, utilizes Gauss's weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel; The HOG eigenwert X of each characteristic block is calculated according to the HOG eigenwert of each cell factory _i, and be normalized.Fisher value Fi sequencing unit is used for described eigenwert X _ifisher criterion is adopted to calculate, and the Fisher value F that will calculate _isort.HOG eigenwert input block according to ranking results from all Fisher value F _ithe maximum Fisher value of middle selection is input in SVM classifier as feature, and calculates corresponding classification discrimination R.Described HOG eigenwert input block often inputs a Fisher value, and whether execution R is greater than R by END instruction unit _tdetection action, until R>R _t, R _tfor the Classification and Identification rate of setting, stop from remaining Fisher value S _imiddle selection the greater carries out being input to SVM classifier.

Described Renminbi false distinguishing module comprises model and sets up unit, Functional Analysis unit and pattern recognition unit.Model sets up unit for Modling model majorized function, c>=0, (i=1,2 ..., n), restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe the relaxation factor considered error in classification and introduce, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training.Functional Analysis unit is according to the model set up, and can obtain decision function is wherein, α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, pattern recognition unit, for selecting SVM classifier parameter, carries out two quasi-mode identifications according to decision function from the HOG characteristic value collection of SVM classifier.

In the experiment of the method, genuine note sample number is 500, and counterfeit money sample number is 40, is training sample set and test sample book collection by true and false coin sample random division, and such as wherein the sample of 50% is used for training, and the sample of 50% is used for testing.Namely choose 250 genuine notes and 20 counterfeit money samples are used for selecting HOG characteristic sum training SVM classifier, other 250 genuine notes and 20 counterfeit money samples are used for test.Repeat 10 times, carry out training and testing with the sample of different 50% at every turn.The method of calculating Classification and Identification accuracy rate (accuracy), leakage knowledge rate (miss rate or false negative rate), misclassification rate (false positive rate) is: $\mathrm{false\; positive\; rate}=\frac{\mathrm{FP}}{\mathrm{TN}+\mathrm{FP}},\mathrm{Accuracy}=\frac{\mathrm{TN}+\mathrm{TP}}{\mathrm{TN}+\mathrm{TP}+\mathrm{FP}+\mathrm{FN}},$ Wherein, TP (true positive) represents actual and is genuine note and the sample number of also predicted one-tenth genuine note, FN (false negative) represents actual and is genuine note but the sample number of predicted one-tenth counterfeit money, FP (false positive) represents actual and is counterfeit money but the sample number of predicted one-tenth genuine note, and it is the sample number of counterfeit money also predicted one-tenth counterfeit money that TN (true negative) represents actual.

The feature in the false proof dot pattern of zebra stripes with better separating capacity is have chosen according to above-mentioned feature extraction and Algorithms of Selecting.When choosing the 20th characteristic block, can obtain the classification accuracy of 99.03%, Feature Selection process stops.As shown in Figure 4, discrimination enlarges markedly when beginning selected characteristic block, when the quantity of the characteristics of image block chosen more than 20 time, the growth rate of genealogical classification discrimination tends to be steady.Find through overtesting, utilize the HOG feature interpretation operator and C-SVM sorter that improve, show that discrimination is 99.032%, leaking discrimination is 1%, and misclassification rate is 0%, and the average detected time is that 0.25s/ opens.The false distinguishing method of employing zebra stripes Infrared Image Features has good discrimination when differentiating Renminbi and efficiency of algorithm is high, and can shorten detection time, and the method can be applicable in paper money discrimination recognition system.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (11)

1., based on the Renminbi false distinguishing method of zebra stripes Infrared Image Features, it is characterized in that, comprise the following steps:

Steps A: the infrared image gathering Renminbi, and the infrared image collected is carried out pre-service;

Step B: pretreated infrared image is carried out HOG feature extraction, and choose from the HOG feature extracted and meet pre-conditioned HOG feature;

Step C: the HOG feature according to choosing out carries out false distinguishing.

2. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 1, it is characterized in that, described steps A specifically comprises:

Steps A 1: Renminbi image is gathered and locates;

Steps A 2: extract interesting image regions from the image of having good positioning.

3. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 2, it is characterized in that, described steps A 1 specifically comprises:

Steps A 101: the infrared transmission image I gathering Renminbi front _t(x, y) and infrared external reflection image I _r(x, y);

Steps A 102: by described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x, y) carries out additive operation by following formula and obtains image I (x, y),

Steps A 103: the edge adopting the most dark areas of hough change detection image I (x, y), positions target image according to edge, and the target image of collection is carried out slant correction.

4. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 2, it is characterized in that, described step B specifically comprises:

Step B1: select suitable HOG characteristic parameter, described interesting image regions is divided into several characteristic blocks, each characteristic block forms by 2 × 2 cell factory, each cell factory is formed by the pixel of 8 × 8;

Step B2: the HOG eigenwert X calculating described each characteristic block _i;

Step B3: by described eigenwert X _ifisher criterion is adopted to calculate Fisher value F _iand the Fisher value F that will calculate _isort;

Step B4: according to ranking results, from all Fisher value F _ithe maximum Fisher value of middle selection is input in SVM classifier as feature, and calculates corresponding classification discrimination R;

Step B5: repeating said steps B4, until R>R _t, wherein, R _tfor the Classification and Identification rate of setting.

5. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 4, it is characterized in that, described step B2 specifically comprises the following steps:

Step B201: adopt gradient operator [-1,0,1] and [-1,0,1] ^-1calculate gradient direction and the gradient magnitude of each cell factory, computing formula is: g (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel;

Step B202: the HOG feature of each cell factory is calculated by wherein each pixel Nearest Neighbor with Weighted Voting, utilizes Gauss's weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel;

Step B203: the HOG eigenwert X calculating each characteristic block according to the HOG eigenwert of each cell factory _i, and be normalized.

6. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 4, it is characterized in that, described step B3 specifically comprises the following steps:

Step B301: according to the within-cluster variance of following formulae discovery i-th class, i=ω _ror ω _c, i=ω _ror ω _c, wherein, ω _rrepresent the kind of genuine note, ω _crepresent the kind of counterfeit money, X represents the HOG feature extracted from each characteristic block, m _irepresentation class ω _ror class ω _csample characteristics mean value;

Step B302: the within-cluster variance sum and the inter _ class relationship that calculate all classes, i=ω _ror ω _c, wherein, S _wrepresent the within-cluster variance sum of all classes, S _brepresent inter _ class relationship, Fisher value F _iequal S _b/ S _w.

7. the Renminbi false distinguishing method based on zebra stripes Infrared Image Features according to claim 1, it is characterized in that, described step C specifically comprises:

Step C1: Modling model majorized function, c>=0, (i=1,2 ..., n), restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe relaxation factor, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training;

Step C2: according to the model set up, obtaining decision function is wherein, $b=y_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}y{\mathrm{\α}}_{i}K(X_i\·X),$ α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, $K(X_i,X)=e^{-\mathrm{\γ}{\left|\right|X_i-X\left|\right|}^2};$

Step C3: compare grid data service, genetic algorithm and particle swarm optimization algorithm, draws the parameter of SVM classifier, then from the HOG characteristic value collection of SVM classifier, carries out two quasi-mode identifications according to described parameter.

8., based on the Renminbi identification system of zebra stripes Infrared Image Features, it is characterized in that, comprising:

Image pre-processing module, carries out pre-service to the Renminbi infrared image collected;

HOG characteristic extracting module, extracts HOG feature by pretreated infrared image, and chooses from the HOG feature extracted and meet pre-conditioned HOG feature;

Renminbi false distinguishing module, carries out false distinguishing by the HOG feature choosing out.

9. the Renminbi identification system based on zebra stripes Infrared Image Features according to claim 8, it is characterized in that, described image pre-processing module comprises:

Image acquisition and positioning unit, gather the infrared transmission image I in Renminbi front _t(x, y) and infrared external reflection image I _r(x, y); By described infrared transmission image I _t(x, y) and infrared external reflection image I _r(x _,y) carry out additive operation by following formula and obtain image I (x, y), with accurate positioning security line position; Adopt the edge of the most dark areas of hough change detection image I (x, y), according to edge, target image is positioned, the target image of collection is carried out Slant Rectify.

Region of interesting extraction unit, the target image according to having good positioning extracts interesting image regions.

10. the Renminbi identification system based on zebra stripes Infrared Image Features according to claim 9, is characterized in that, described HOG characteristic extracting module comprises:

Area division unit, choose suitable HOG characteristic parameter, described interesting image regions is divided into several characteristic blocks, each characteristic block forms by Unit 2 × 2, and each unit is formed by the pixel of 8 × 8;

HOG eigenwert computing unit, adopts gradient operator [-1,0,1] and [-1,0,1] ^-1calculate gradient direction and the gradient magnitude of each unit, computing formula is: g (x, y)={ [I (x+1, y)-I (x-1, y)] ²+ [I (x, y+1)-I (x, y-1)] ²} ^1/2, wherein, I (x, y) is the grey scale pixel value being positioned at (x, y) in image, and α (x, y) represents the gradient direction of this pixel, and G (x, y) represents the gradient magnitude of this pixel; The HOG feature of each cell factory is calculated by wherein each pixel Nearest Neighbor with Weighted Voting, utilizes Gauss's weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel; The HOG eigenwert X of each characteristic block is calculated according to the HOG eigenwert of each cell factory _i, and be normalized;

Fisher value F _isequencing unit, by described eigenwert X _ifisher criterion is adopted to calculate Fisher value F _i, and the Fisher value F that will calculate _isort;

HOG eigenwert input block, according to ranking results, from all Fisher value F _ithe Fisher value F that middle selection is maximum _ibe input in SVM classifier as feature, and calculate corresponding classification discrimination R;

END instruction unit, described HOG eigenwert input block often inputs a Fisher value F _i, perform a R and whether be greater than R _tdetection action, until R>R _t, wherein, R _tfor the Classification and Identification rate of setting.

The 11. Renminbi identification systems based on zebra stripes Infrared Image Features according to claim 8, is characterized in that, described Renminbi false distinguishing module comprises:

Unit set up by model, Modling model majorized function, $\underset{\mathrm{\ω},b,\mathrm{\ξ}}{\min }\frac{1}{2}{\left|\right|W\left|\right|}^2+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i,C\≥0,(i=\mathrm{1,2},\·\·\·,n),$ Restricted condition is y _i[(WX _i)+b]>=1-ξ _i, ξ _i>=0, wherein, the coefficient vector of Optimal Separating Hyperplane in W representation feature space, the threshold value in b presentation class face, ξ _ibe relaxation factor, C is penalty factor mistake being divided to sample, and n represents that training sample concentrates number of training;

Functional Analysis unit, according to the model set up, obtaining decision function is wherein, $b=y_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}y{\mathrm{\α}}_{i}K(X_i\·X),$ α _ifor Lagrange multiplier, X _ifor known sample, y _iknown sample label, X is the sample needing classification, K (X _i, X) and be kernel function, $K(X_i,X)=e^{-\mathrm{\γ}{\left|\right|X_i-X\left|\right|}^2};$

Pattern recognition unit, selects SVM classifier parameter, carries out two quasi-mode identifications according to decision function from the HOG characteristic value collection of SVM classifier.