CN104298989B - False distinguishing method and its system based on zebra stripes Infrared Image Features - Google Patents
False distinguishing method and its system based on zebra stripes Infrared Image Features Download PDFInfo
- Publication number
- CN104298989B CN104298989B CN201410415338.3A CN201410415338A CN104298989B CN 104298989 B CN104298989 B CN 104298989B CN 201410415338 A CN201410415338 A CN 201410415338A CN 104298989 B CN104298989 B CN 104298989B
- Authority
- CN
- China
- Prior art keywords
- image
- hog
- rmb
- features
- infrared
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
- G06V10/507—Summing image-intensity values; Histogram projection analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- G—PHYSICS
- G07—CHECKING-DEVICES
- G07D—HANDLING OF COINS OR VALUABLE PAPERS, e.g. TESTING, SORTING BY DENOMINATIONS, COUNTING, DISPENSING, CHANGING OR DEPOSITING
- G07D7/00—Testing specially adapted to determine the identity or genuineness of valuable papers or for segregating those which are unacceptable, e.g. banknotes that are alien to a currency
- G07D7/20—Testing patterns thereon
- G07D7/2016—Testing patterns thereon using feature extraction, e.g. segmentation, edge detection or Hough-transformation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06K—GRAPHICAL DATA READING; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
- G06K2207/00—Other aspects
- G06K2207/1012—Special detection of object
Abstract
The present invention relates to RMB authentication detection technology there is provided a kind of false distinguishing method based on zebra stripes Infrared Image Features, comprise the following steps:Step A, the infrared image for gathering RMB, and the infrared image collected is pre-processed;Step B, pretreated infrared image is carried out to HOG feature extractions, and chosen from the HOG features extracted and meet the HOG features of preparatory condition;Step C, according to select come HOG features progress false distinguishing.The described false distinguishing method based on zebra stripes Infrared Image Features is acquired using infrared light to the zebra anti-counterfeiting image of RMB, and zebra anti-counterfeiting image is handled and recognized using the HOG and the method for SVM improved, influence of noise can effectively be eliminated, the performance of infrared image false distinguishing is improved, while the accuracy rate to true and false coin false distinguishing can also be improved.
Description
Technical field
The present invention relates to RMB authentication detection technology, more particularly to false distinguishing method based on zebra stripes Infrared Image Features and its
System.
Background technology
100 yuan of 2005 editions, 50 yuan, 20 yuan, the bank note such as 10 yuan be imaged using infrared transmission mode, safety line region is presented
Zebra stripes security pattern.It is invisible in the sunlight because zebra stripes exist along with opened window safety line, infrared transmission imaging
It is middle just to display, copy difficulty very big.The height of existing report is imitated in counterfeit money, for example, started with HB90 and HD90 sequence numbers
Counterfeit money, possesses watermark true to nature, becomes ink, stealthy denomination, magnetic characteristic and ultraviolet feature, have not seen that infrared zebra stripes are false proof
Point is faked.The feature that can't see under these visible rays so that infrared image false distinguishing has the excellent of uniqueness in various paper money discrimination fields
Gesture.
Characteristics of image is certain expression for containing information in image, and characteristics of image can transform to another from a transform domain
Individual conversion domain representation.In image classification, if the feature space of extraction can find obvious categorised demarcation line, it is possible to preferably
Carry out tagsort.In actual applications, select one accurately character representation be the key solved the problems, such as.In banknote image classification
False distinguishing recognizes field, it is necessary to carry out feature extraction to banknote image first, then complete identification and false distinguishing.
For original target image, feature extraction selection is carried out, the feature with preferable discrimination is selected, i.e., same
Having in class has otherness in similitude, variety classes.Redundancy and correlative character can be removed by feature extraction.
The intrinsic dimensionality of extraction is appropriate, if dimension is too big, influences training effectiveness, and dimension is too small, and difference describes very few influence between class
Recognition effect.Accordingly, it would be desirable to which selection has the feature of preferable classifying quality in numerous characteristics of image.In addition, different applications
Field, the standard of feature selecting is also different.Target image is influenceed by noise and light change, and its shape, size and
The features such as brightness can not be expressed simultaneously, it is necessary to which the problem of being directed to specific is selected and combined.
The content of the invention
The technical problems to be solved by the invention are to provide a kind of RMB mirror based on zebra stripes Infrared Image Features
Fake method and its system, it is intended to solve the problem of existing cash inspecting machine has misrecognition to the imitative counterfeit money of height when differentiating.
The present invention is achieved in that the RMB false distinguishing method based on zebra stripes Infrared Image Features, including following step
Suddenly:
Step A:The infrared image of RMB is gathered, and the infrared image collected is pre-processed;
Step B:Pretreated infrared image is subjected to HOG feature extractions, and chooses full from the HOG features extracted
The HOG features of sufficient preparatory condition;
Step C:According to select come HOG features carry out false distinguishing.
Further, the step A is specifically included:
Step A1:RMB image is acquired and positioned;
Step A2:Interesting image regions are extracted from the image having good positioning.
Further, the step A1 is specifically included:
Step A101:Gather the positive infrared transmission image I of RMBt(x, y) and infrared external reflection image Ir(x,y);
Step A102:By the infrared transmission image It(x, y) and infrared external reflection image Ir(x, y) is carried out as follows
Add operation obtains image I (x, y),
Step A103:Using the edge of hough change detection image I (x, y) most dark areas, according to edge by target image
Positioned, and the target image of collection is entered into line tilt correction.
Further, the step B is specifically included:
Step B1:Suitable HOG characteristic parameters are selected, described image area-of-interest is divided into several characteristic blocks,
Each characteristic block is constituted by 2 × 2 cell factories, and each cell factory is constituted by 8 × 8 pixel;
Step B2:Calculate the HOG characteristic values X of each characteristic blocki;
Step B3:By the characteristic value XiFisher values F is calculated using Fisher criterionsiAnd by the Fisher calculated
Value FiIt is ranked up;
Step B4:According to ranking results, from all Fisher values FiThe maximum Fisher values of middle selection are defeated as feature
Enter into SVM classifier, and calculate corresponding classification discrimination R;
Step B5:Repeating said steps B4, until R>Rt, wherein, RtFor the Classification and Identification rate of setting.
Further, the step B2 specifically includes following steps:
Step B201:Using gradient operator [- 1,0,1] and [- 1,0,1]-1Calculate the gradient direction and gradient of each unit
Amplitude, calculation formula is:G (x, y)=[I (and 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 for being located at (x, y) in image, and α (x, y) is represented
The gradient direction of the pixel, G (x, y) represents the gradient magnitude of the pixel;
Step B202:The HOG features of each cell factory are calculated by wherein each pixel Nearest Neighbor with Weighted Voting, using Gauss plus
Weigh gradient magnitude and tri-linear interpolation methods calculate the weights of each pixel;
Step B203:The HOG characteristic values X of each characteristic block is calculated according to the HOG characteristic values of each cell factoryi, go forward side by side
Row normalization.
Further, the step B3 specifically includes following steps:
Step B301:The within-cluster variance of the i-th class is calculated according to equation below,I=ωRorωC,I=ωRorωC, wherein, ωRRepresent the species of genuine note, ωCRepresent the species of counterfeit money, X
Represent the HOG features extracted from each characteristic block, miRepresent class ωROr class ωCSample characteristics average value;
Step B302:The within-cluster variance sum and inter _ class relationship of all classes are calculated,I=ωRorωC,Wherein, SwRepresent the within-cluster variance sum of all classes, SbRepresent inter _ class relationship,
Fisher values FiEqual to Sb/Sw。
Further, the step C is specifically included:
Step C1:Model optimization function is set up,C >=0, (i=1,2 ..., n), restricted condition
For yi[(W·Xi)+b]≥1-ξi,ξi>=0, wherein, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, b presentation classes
The threshold value in face, ξiIt is relaxation factor, C is the penalty factor for wrong point of sample, and n represents that training sample concentrates number of training;
Step C2:According to the model of foundation, obtaining decision function is
Wherein,αiFor Lagrange multiplier, XiFor known sample, yiKnown sample label, X is needs
The sample of classification, K (Xi, X) and it is kernel function,
Step C3:Compare grid data service, genetic algorithm and particle swarm optimization algorithm, draw the parameter of SVM classifier, then
Two quasi-mode identifications are carried out from the HOG characteristic value collections of SVM classifier according to the parameter.
The present invention also provides a kind of RMB identification system based on zebra stripes Infrared Image Features, including:
Image pre-processing module, is pre-processed to the RMB infrared image collected;
HOG characteristic extracting modules, HOG features are extracted by pretreated infrared image, and from the HOG features extracted
Choose the HOG features for meeting preparatory condition;
RMB false distinguishing module, will select next HOG features and carries out false distinguishing.
Further, described image pretreatment module includes:
IMAQ and positioning unit, the positive infrared transmission image I of collection RMBt(x, y) and infrared external reflection image Ir
(x,y);By the infrared transmission image It(x, y) and infrared external reflection image Ir(x, y) carries out add operation and obtained as follows
Image I (x, y),To be accurately positioned safe line position;Adopt
With the edge of hough change detection image I (x, y) most dark areas, target image is positioned according to edge, by the mesh of collection
Logo image carries out Slant Rectify.
Region of interesting extraction unit, interesting image regions are extracted according to the target image having good positioning.
Further, the HOG characteristic extracting modules include:
Area division unit, chooses suitable HOG characteristic parameters, and it is special that described image area-of-interest is divided into several
Block is levied, each characteristic block is constituted by Unit 2 × 2, each unit is constituted by 8 × 8 pixel;
HOG characteristic value computing units, using gradient operator [- 1,0,1] and [- 1,0,1]-1Calculate the gradient side of each unit
To and gradient magnitude, calculation 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 for being located at (x, y) in image, α
(x, y) represents the gradient direction of the pixel, and G (x, y) represents the gradient magnitude of the pixel;The HOG features of each cell factory by
Wherein each pixel Nearest Neighbor with Weighted Voting is calculated, and each pixel is calculated using Gauss weighted gradient amplitude and tri-linear interpolation methods
Weights;The HOG characteristic values X of each characteristic block is calculated according to the HOG characteristic values of each cell factoryi, and be normalized;
Fisher values FiSequencing unit, by the characteristic value XiFisher values F is calculated using Fisher criterionsi, and will calculate
Fisher values F outiIt is ranked up;
HOG characteristic value input blocks, according to ranking results, from all Fisher values FiThe maximum Fisher values of middle selection
FiIt is input to as feature in SVM classifier, and calculates corresponding classification discrimination R;
END instruction unit, the HOG characteristic values input block often inputs a Fisher values Fi, whether R of execution
More than RtDetection operation, until R>Rt, wherein, RtFor the Classification and Identification rate of setting.
Further, the RMB false distinguishing module includes:
Model sets up unit, sets up model optimization function,
Restricted condition is yi[(W·Xi)+b]≥1-ξi,ξi>=0, wherein, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, b
The threshold value in presentation class face, ξiIt is relaxation factor, C is the penalty factor for wrong point of sample, and n represents that training sample concentrates training
Sample number;
Functional Analysis unit, according to the model of foundation, obtaining decision function is
Wherein,αiFor Lagrange multiplier, XiFor known sample, yiKnown sample label, X is needs
The sample of classification, K (Xi, X) and it is kernel function,
Pattern recognition unit, selects SVM classifier parameter, according to decision function from the HOG characteristic value collections of SVM classifier
Two quasi-mode identifications of middle progress.
Compared with prior art, beneficial effect is the present invention:The described false distinguishing based on zebra stripes Infrared Image Features
Method is acquired using infrared light to the zebra anti-counterfeiting image of RMB, and false proof to zebra using the HOG methods improved
Image is handled and recognized that this method have chosen suitable HOG characteristic parameters and SVM classifier parameter, can effectively eliminate and make an uproar
Sound shadow is rung, and improves the performance of infrared image false distinguishing, and the accuracy rate to true and false coin can reach 99.03%.
Brief description of the drawings
Fig. 1 is the flow chart of the false distinguishing method of the invention 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 flow chart of HOG feature extractions 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 using network technique selection parameter;
Fig. 6 is the experimental result picture using genetic algorithm selection parameter;
Fig. 7 is the experimental result picture using particle group optimizing method selection parameter.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
As shown in figure 1, for a preferred embodiment of the invention, the RMB false distinguishing side based on zebra stripes Infrared Image Features
Method, comprises the following steps:Step A:The infrared image of RMB is gathered, and the infrared image collected is pre-processed;Step
Rapid B:Pretreated infrared image is subjected to HOG (Histogram of Orient Gradient, gradient orientation histogram)
Feature extraction, and the HOG features for meeting preparatory condition are chosen from the HOG features extracted;Step C:According to select come
HOG features carry out false distinguishing.
When performing step A, RMB image is acquired and positioned first.It is every using infrared image capturing system collection
Open the positive infrared transmission image I of RMBt(x, y) and infrared external reflection image Ir(x, y), the position that RMB is gathered twice is protected
Hold constant, and the image collected is gray level image.By the infrared transmission image It(x, y) and infrared external reflection image Ir(x,
Y) formula is pressedCarry out add operation and obtain image I (x, y),
It was found from from above formula, the gray scale of two images and during more than 255, value 255;Gray scale takes the ash of two figures with when being not above 255
Degree and.After the add operation, region most dark retains in artwork, and other regions are all changed into white background.Safety line is saturating
Penetrate and be most dark region in reflected image, therefore, safety line region can be pin-pointed to using above-mentioned add operation.
By the edge of hough change detection image I (x, y) most dark areas, target image is positioned according to edge, rotated, it is complete
Into the slant correction of the target image of collection.Zebra stripes anti-counterfeiting image be imaged under white light it is invisible, but infrared transmission into
As in, light and shade alternating rectangular block (zebra stripes claimed in the application) is shown as, this characteristic point can be used as the true and false coin of differentiation
Feature.Safety line is a black line in infrared imaging, is located at peace through whole zebra line pattern, i.e. zebra line pattern
Around completely.If not finding the presence of fixed safety line during above-mentioned processing image, then this banknote is direct
It is determined as counterfeit money.Then, interesting image regions are extracted from the image having good positioning.Specifically, the horizontal level of safety line
It is determined that after, the horizontal coordinate of ROI (Region of Interest, interesting image regions) central point is also determined.Two 16
The wide rectangular area of pixel is merged into a new zebra line pattern after being split respectively from the left side and the right of safety line.It is complete
Into after aforesaid operations, the extracted region of the size of 5H × 32 is out as interesting image regions, as shown in Fig. 2 for bank note mirror
Feature extraction region in puppet.
There are a variety of parameters to need selection when extracting HOG features.Wherein there are several parameters critically important, they are respectively:(1)
ROI size:The size of the area-of-interest extracted from image;(2) Block size:From the big of ROI region divided block
It is small;(3) Block sliding steps:Overlapping normalization is carried out using the concept of block, overcomes illumination variation to influence;(4) Cell's is big
It is small:Cell member size in Block;(5) Bin size:The number in statistical gradient direction in one cell member;(6) Block and
Cell assembled arrangement mode:A variety of cell arrangement modes, which can be chosen, in ROI is interval constitutes a Block.
Tested using 540 samples (500 are genuine notes, and 40 are counterfeit moneys), wherein 200 genuine note samples and 30
Counterfeit money sample is trained.Different parameter size and the cell factory (cell) in various features block (block) are attempted
Arrangement mode, obtained discrimination the results are shown in Table 1.1.Experiment is found, when ROI region chooses 32 × 128 pixel sizes, block
The cell structures of middle selection 2 × 2, when cell chooses 8 × 8 pixel, bin takes 9, step-length one cell factory of selection of sliding shoe
Width, obtained Classification and Identification rate highest.
The relation of each Parameters variation of table 1.1 and discrimination
It is shown in Figure 3, when performing step B, first, suitable HOG characteristic parameters are selected according to above-mentioned experimental result,
Described image area-of-interest is divided into several characteristic blocks, each characteristic block is constituted by 2 × 2 cell factories, Mei Gexi
Born of the same parents' unit is constituted by 8 × 8 pixel.Then, the HOG characteristic values X of each characteristic block is calculatedi.Specifically, using gradient
Operator [- 1,0,1] and [- 1,0,1]-1The gradient direction and gradient magnitude of each cell factory are calculated, computational methods are:G (x, y)=[I (and 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 for being located at (x, y) in image, and α (x, y) represents the gradient side of the pixel
To G (x, y) represents the gradient magnitude of the pixel.The HOG features of each cell factory are thrown by the weighting of each pixel therein
Ticket is calculated, it is possible to use Gauss weighted gradient amplitude and tri-linear interpolation methods calculate the weights of each pixel, and according to every
The HOG characteristic values Nearest Neighbor with Weighted Voting of individual cell factory calculates the HOG characteristic values X of each characteristic blocki, and be normalized.Again to every
The histogram vector of individual characteristic block (block) carries out L2 norm normalizationWherein, v represents characteristic block
In normalization after histogram vectors, | | vk| | k- norm calculations are represented, k=1 or 2, ε are the constants of a very little, are prevented
Produce infinitely large quantity.Light source change when being compensation collection input picture to the purpose that each characteristic block is normalized.By institute
There are the histogram vectors after normalization to be connected into the vector that a size is n × m, wherein, n represents the Nogata in each characteristic block
Scheme the dimension of vector, m represents the number of characteristic block inside interesting image regions.It is preferred that, m=45, n=36.After series connection
Vector is input in SVM (Support Vector Machine, SVM) grader and carries out true and false coin Classification and Identification.So
And, HOG characteristic vector is the vector of higher-dimension.For example, when bin is set to 9, the Duplication of each block is set to 0.5, the HOG features
Dimension is 45 × 4 × 9=1620.High intrinsic dimensionality increases calculating intensity when extraction feature, training sample and classification.Cause
This before HOG features being input to SVM classifier, it is necessary to carry out selection processing.Make discovery from observation, the false proof dot pattern of zebra stripes
Edge direction be mainly both horizontally and vertically, using Fisher criterions remove redundancy HOG features.Surveyed according to the standard,
If this feature is big in the similarity of same group of internal ratio difference group, then this feature has preferable discrimination.Then carry out special
Sequence is levied, the high feature of discrimination is picked out and is used as last feature.Specifically, by characteristic value XiUsing Fisher criterions
Calculated, and by the Fisher values F calculatediIt is ranked up, computational methods are:I=ωRorωC,I=ωRorωC,I=ωRorωC,
Wherein, ωRRepresent the species of genuine note, ωCThe species of counterfeit money is represented, X represents the HOG features extracted from each characteristic block, miRepresent
Class ωROr class ωCSample characteristics average value, SwRepresent the within-cluster variance sum of all classes, SbRepresent inter _ class relationship.Fi=
Sb/Sw, FiRatio it is bigger, its feature X separating capacity is stronger.According to ranking results, from all Fisher values FiMiddle choosing
The Fisher values for selecting maximum are input in SVM classifier as feature, and calculate corresponding classification discrimination R;Again from previous
Secondary remaining Fisher values FiThe maximum Fisher values of middle selection are input in SVM classifier as feature, repeat this action, directly
To R>Rt, RtFor the Classification and Identification rate of setting, stop from remaining Fisher values FiMiddle selection the greater is input to SVM classifier.
HOG features have been chosen to be input to after SVM classifier, it is necessary to false distinguishing be carried out to HOG features, so as to distinguish RMB
The true and false.Zebra stripes feature recognition is carried out using SVM classifier.The classification to genuine note and counterfeit money is a two quasi-modes knowledge in fact
Other problem.If given training set is { (x1, y1), (x2, y2) ..., (xn, yn) }, wherein xi ∈ Rn are input HOG features
Vector, yi∈ { -1,1 } is output vector.Xi represents the HOG features of i-th of training sample, yiRepresent the class of i-th of training sample
Distinguishing label, " -1 " represents counterfeit money classification, and " 1 " represents genuine note classification.The target of true and false coin classification is carried out with SVM is, gives one
The training set of true and false coin is included, one can be found and distinguish two class data and have the hyperplane of largest interval with two class data.
If the training set can be divided by hyperplane, the hyperplane is that parameter W and b in WX+b=0, formula determines super flat
The position in face, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, the threshold value in b presentation classes face, and WX is two vectors
Inner product.Division is optimized in order to try to achieve the training set, the problem can be converted into the problem of seeking optimization hyperplane, thus turn
Turn to and set up the problem of function model is solved.
Specifically false distinguishing method is:First, model optimization function is set up,
Restricted condition is yi[(W·Xi)+b]≥1-ξi,ξi>=0, wherein, Subject to yi[(W·Xi)+b]≥1-ξi,ξi≥0
W represents the coefficient vector of Optimal Separating Hyperplane in feature space, the threshold value in b presentation classes face, ξiIt is to consider that classification is missed
Difference and the relaxation factor introduced, C is the penalty factor for wrong point of sample, and n represents that training sample concentrates number of training.According to
The model of foundation, can release decision function isWherein,αiFor Lagrange multiplier, XiFor known sample, yiKnown sample label, X is to need classification
Sample, K (Xi, X) and it is kernel function,Compare grid data service, genetic algorithm and particle group optimizing
These three algorithms of algorithm, show that grid data service is optimal algorithm by comparing, svm classifier are determined according to grid data service
The parameter of device, two quasi-mode identifications are carried out further according to the parameter from the HOG characteristic value collections of SVM classifier.
As shown in Figures 5 to 7, searching SVM is compared using grid data service, genetic algorithm and particle swarm optimization algorithm
During the parameter of grader, employ 230 training samples and tested, comparative result is shown in Table 1.2.
Parameter of the table 1.2 based on three kinds of algorithms compares
Above-mentioned three kinds of methods can obtain 99.5652% cross validation accuracy rate, and wherein grid data service is time-consuming minimum,
Therefore selection this method selects the parameter of C-SVM graders, and parametric results are C=1, γ=0.10882.
The present invention also provides a kind of identification system based on zebra stripes Infrared Image Features, including:Image pre-processing module,
The RMB infrared image collected is pre-processed;HOG characteristic extracting modules, pretreated infrared image is extracted
HOG features, and the HOG features extracted are chosen;RMB false distinguishing module, according to select come HOG features input
False distinguishing is carried out into SVM classifier.
Described image pretreatment module includes IMAQ and positioning unit and region of interesting extraction unit.IMAQ
It is used to gather the positive infrared transmission image I of RMB with positioning unitt(x, y) and infrared external reflection image Ir(x,y);And by institute
State infrared transmission image It(x, y) and infrared external reflection image Ir(x, y) as follows carry out add operation obtain image I (x,
Y),To be accurately positioned safe line position;It is additionally operable to use
The edge of hough change detection image I (x, y) most dark areas, is positioned target image according to edge, and by the mesh of collection
Logo image enters line tilt correction.Region of interesting extraction unit extracts interesting image area according to the target image having good positioning
Domain.
The HOG characteristic extracting modules include area division unit, HOG characteristic values computing unit, Fisher values SiSequence
Unit, HOG characteristic values input block and END instruction unit.Area division unit is used to choose suitable HOG characteristic parameters, will
Described image area-of-interest is divided into several characteristic blocks, and each characteristic block is constituted by Unit 2 × 2, and each unit is by 8
× 8 pixel is constituted.HOG characteristic values computing unit uses gradient operator [- 1,0,1] and [- 1,0,1]-1Calculate each unit
Gradient direction and gradient magnitude, calculation 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 pixel grey scale for being located at (x, y) in image
Value, α (x, y) represents the gradient direction of the pixel, and G (x, y) represents the gradient magnitude of the pixel;The HOG of each cell factory is special
Levy by wherein each pixel Nearest Neighbor with Weighted Voting to calculate, each picture is calculated using Gauss weighted gradient amplitude and tri-linear interpolation methods
The weights of element;The HOG characteristic values X of each characteristic block is calculated according to the HOG characteristic values of each cell factoryi, and be normalized.
Fisher value Fi sequencing units are used for the characteristic value XiCalculated using Fisher criterions, and will be calculated
Fisher values FiIt is ranked up.HOG characteristic values input block is according to ranking results from all Fisher values FiMiddle selection maximum
Fisher values are input in SVM classifier as feature, and calculate corresponding classification discrimination R.The HOG characteristic values input
Unit often inputs a Fisher value, and END instruction unit will perform whether a R is more than RtDetection operation, until R>Rt, Rt
For the Classification and Identification rate of setting, stop from remaining Fisher values SiMiddle selection the greater carries out being input to SVM classifier.
The RMB false distinguishing module sets up unit, Functional Analysis unit and pattern recognition unit including model.Model is built
Vertical unit is used to set up model optimization function,C >=0, (i=1,2 ..., n), restricted condition is
yi[(W·Xi)+b]≥1-ξi,ξi>=0, wherein, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, b presentation classes face
Threshold value, ξiIt is the relaxation factor for considering error in classification and introducing, C is the penalty factor for wrong point of sample, and n represents to train sample
This concentration training sample number.Functional Analysis unit is according to the model of foundation, and can obtain decision function isWherein,αiFor Lagrange multiplier, Xi
For known sample, yiKnown sample label, the sample that X classifies for needs, K (Xi, X) and it is kernel function,
Pattern recognition unit is used to select SVM classifier parameter, is entered according to decision function from the HOG characteristic value collections of SVM classifier
The quasi-mode of row two is recognized.
In the experiment of this method, genuine note sample number is 500, and counterfeit money sample number is 40, by true and false coin sample random division
For training sample set and test sample collection, such as wherein 50% sample is used for training, and 50% sample is used for testing.Choose
250 genuine notes and 20 counterfeit money samples are used for selecting HOG features and training SVM classifier, in addition 250 genuine notes and 20 counterfeit moneys
Sample is used for testing.It is repeated 10 times, is trained and tested with 50% different samples every time.Calculate Classification and Identification accuracy rate
(accuracy), leakage knowledge rate (miss rate or false negative rate), misclassification rate (false positive
Rate method) is:
Wherein, TP (true positive) represents really genuine note and is also predicted to the sample number of genuine note, FN (false
Negative) represent really genuine note but be predicted to the sample number of counterfeit money, FP (false positive) represents to be really vacation
Coin but the sample number for being predicted to genuine note, TN (true negative) expressions are really the sample that counterfeit money is also predicted to counterfeit money
Number.
Have chosen according to above-mentioned feature extraction and Algorithms of Selecting has preferable separating capacity in the false proof dot pattern of zebra stripes
Feature.When choosing the 20th characteristic block, 99.03% classification accuracy can be obtained, Feature Selection process stops.Such as
Shown in Fig. 4, discrimination start selected characteristic block when significantly increase, when selection characteristics of image block quantity more than 20
When, the growth rate of genealogical classification discrimination tends to be steady.Found through overtesting, operator is described using improved HOG features
With C-SVM graders, it is 99.032% to draw discrimination, and leakage discrimination is 1%, and misclassification rate is 0%, and the average detected time is
0.25s/.There is preferable discrimination and algorithm when differentiating RMB using the false distinguishing method of zebra stripes Infrared Image Features
Efficiency high, and detection time can be shortened, this method can be applied in paper money discrimination identifying system.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
Any modifications, equivalent substitutions and improvements made within refreshing and principle etc., should be included in the scope of the protection.
Claims (9)
1. a kind of RMB false distinguishing method based on zebra stripes Infrared Image Features, it is characterised in that comprise the following steps:
Step A:The infrared image of RMB is gathered, and the infrared image collected is pre-processed;
Step B:Pretreated infrared image is subjected to HOG feature extractions, and satisfaction is chosen in advance from the HOG features extracted
If the HOG features of condition;
Step C:According to select come HOG features carry out false distinguishing;
The step B is specifically included:
Step B1:Suitable HOG characteristic parameters are selected, interesting image regions are divided into several characteristic blocks, each feature
Block is constituted by 2 × 2 cell factories, and each cell factory is constituted by 8 × 8 pixel;
Step B2:Calculate the HOG characteristic values X of each characteristic blocki;
Step B3:By the characteristic value XiFisher values F is calculated using Fisher criterionsiAnd by the Fisher values F calculatedi
It is ranked up;
Step B4:According to ranking results, from all Fisher values FiThe maximum Fisher values of middle selection are input to as feature
In SVM classifier, and calculate corresponding classification discrimination R;
Step B5:Repeating said steps B4, until R > Rt, wherein, RtFor the Classification and Identification rate of setting.
2. the RMB false distinguishing method according to claim 1 based on zebra stripes Infrared Image Features, it is characterised in that institute
Step A is stated to specifically include:
Step A1:RMB image is acquired and positioned;
Step A2:Interesting image regions are extracted from the image having good positioning.
3. the RMB false distinguishing method according to claim 2 based on zebra stripes Infrared Image Features, it is characterised in that institute
Step A1 is stated to specifically include:
Step A101:Gather the positive infrared transmission image I of RMBt(x, y) and infrared external reflection image Ir(x, y);
Step A102:By the infrared transmission image It(x, y) and infrared external reflection image Ir(x, y) carries out addition as follows
Computing obtains image I (x, y),
Step A103:Using the edge of hough change detection image I (x, y) most dark areas, target image is carried out according to edge
Positioning, and the target image of collection is entered into line tilt correction.
4. the RMB false distinguishing method according to claim 1 based on zebra stripes Infrared Image Features, it is characterised in that institute
State step B2 and specifically include following steps:
Step B201:Using gradient operator [- 1,0,1] and [- 1,0,1]-1Calculate the gradient direction and gradient of each cell factory
Amplitude, calculation formula is:G (x, y)=[I (and 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 for being located at (x, y) in image, and α (x, y) is represented should
The gradient direction of pixel, G (x, y) represents the gradient magnitude of the pixel;
Step B202:The HOG features of each cell factory are calculated by wherein each pixel Nearest Neighbor with Weighted Voting, and ladder is weighted using Gauss
Spend amplitude and tri-linear interpolation methods calculate the weights of each pixel;
Step B203:The HOG characteristic values X of each characteristic block is calculated according to the HOG characteristic values of each cell factoryi, and carry out normalizing
Change.
5. the RMB false distinguishing method according to claim 1 based on zebra stripes Infrared Image Features, it is characterised in that institute
State step B3 and specifically include following steps:
Step B301:The within-cluster variance of the i-th class is calculated according to equation below, Wherein, ωRRepresent the species of genuine note, ωCRepresent counterfeit money
Species, X represents the HOG features extracted from each characteristic block, miRepresent class ωROr class ωCSample characteristics average value;
Step B302:The within-cluster variance sum and inter _ class relationship of all classes are calculated, Wherein, SwRepresent the within-cluster variance sum of all classes, SbBetween expression class
Dispersion, Fisher values FiEqual to Sb/Sw。
6. the RMB false distinguishing method according to claim 1 based on zebra stripes Infrared Image Features, it is characterised in that institute
Step C is stated to specifically include:
Step C1:Model optimization function is set up,Restricted condition is yi
[(W·Xi)+b]≥1-ξi, ξi>=0, wherein, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, b presentation classes face
Threshold value, ξiIt is relaxation factor, C is the penalty factor for wrong point of sample, and n represents that training sample concentrates number of training;
Step C2:According to the model of foundation, obtaining decision function isIts
In,αiFor Lagrange multiplier, XiFor known sample, yiKnown sample label, X is needs point
The sample of class, K (Xi, X) and it is kernel function,
Step C3:Compare grid data service, genetic algorithm and particle swarm optimization algorithm, draw the parameter of SVM classifier, further according to
The parameter carries out two quasi-mode identifications from the HOG characteristic value collections of SVM classifier.
7. a kind of RMB identification system based on zebra stripes Infrared Image Features, it is characterised in that including:
Image pre-processing module, is pre-processed to the RMB infrared image collected;
HOG characteristic extracting modules, extract HOG features, and chosen from the HOG features extracted by pretreated infrared image
Meet the HOG features of preparatory condition;
RMB false distinguishing module, will select next HOG features and carries out false distinguishing;
The HOG characteristic extracting modules include:
Area division unit, chooses suitable HOG characteristic parameters, interesting image regions is divided into several characteristic blocks, often
Individual characteristic block is constituted by Unit 2 × 2, and each unit is constituted by 8 × 8 pixel;
HOG characteristic value computing units, using gradient operator [- 1,0,1] and [- 1,0,1]-1Calculate each unit gradient direction and
Gradient magnitude, calculation formula is:G (x, y)=[I (and 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 for being located at (x, y) in image, α (x, y) table
Show the gradient direction of the pixel, G (x, y) represents the gradient magnitude of the pixel;The HOG features of each cell factory are by wherein each
Pixel Nearest Neighbor with Weighted Voting is calculated, and utilizes Gauss weighted gradient amplitude and tri-linear interpolation methods to calculate the weights of each pixel;Root
The HOG characteristic values X of each characteristic block is calculated according to the HOG characteristic values of each cell factoryi, and be normalized;
Fisher values FiSequencing unit, by the characteristic value XiFisher values F is calculated using Fisher criterionsi, and will calculate
Fisher values FiIt is ranked up;
HOG characteristic value input blocks, according to ranking results, from all Fisher values FiThe maximum Fisher values F of middle selectioniMake
It is characterized and is input in SVM classifier, and calculates corresponding classification discrimination R;
END instruction unit, the HOG characteristic values input block often inputs a Fisher values Fi, perform whether a R is more than Rt
Detection operation, until R > Rt, wherein, RtFor the Classification and Identification rate of setting.
8. the RMB identification system according to claim 7 based on zebra stripes Infrared Image Features, it is characterised in that institute
Stating image pre-processing module includes:
IMAQ and positioning unit, the positive infrared transmission image I of collection RMBt(x, y) and infrared external reflection image Ir(x,
y);By the infrared transmission image It(x, y) and infrared external reflection image Ir(x, y) carries out add operation and obtains figure as follows
As I (x, y),To be accurately positioned safe line position;Using
The edge of hough change detection image I (x, y) most dark areas, is positioned target image according to edge, by the target of collection
Image carries out Slant Rectify;
Region of interesting extraction unit, interesting image regions are extracted according to the target image having good positioning.
9. the RMB identification system according to claim 7 based on zebra stripes Infrared Image Features, it is characterised in that institute
Stating RMB false distinguishing module includes:
Model sets up unit, sets up model optimization function,Restricted condition is
yi[(W·Xi)+b]≥1-ξi, ξi>=0, wherein, W represents the coefficient vector of Optimal Separating Hyperplane in feature space, b presentation classes face
Threshold value, ξiIt is relaxation factor, C is the penalty factor for wrong point of sample, and n represents that training sample concentrates number of training;
Functional Analysis unit, according to the model of foundation, obtaining decision function is
Wherein,αiFor Lagrange multiplier, XiFor known sample, yiKnown sample label, X is needs
The sample of classification, K (Xi, X) and it is kernel function,
Pattern recognition unit, selects SVM classifier parameter, is entered according to decision function from the HOG characteristic value collections of SVM classifier
The quasi-mode of row two is recognized.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410415338.3A CN104298989B (en) | 2014-08-20 | 2014-08-20 | False distinguishing method and its system based on zebra stripes Infrared Image Features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410415338.3A CN104298989B (en) | 2014-08-20 | 2014-08-20 | False distinguishing method and its system based on zebra stripes Infrared Image Features |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104298989A CN104298989A (en) | 2015-01-21 |
CN104298989B true CN104298989B (en) | 2017-08-04 |
Family
ID=52318710
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410415338.3A Active CN104298989B (en) | 2014-08-20 | 2014-08-20 | False distinguishing method and its system based on zebra stripes Infrared Image Features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104298989B (en) |
Families Citing this family (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104809433B (en) * | 2015-04-21 | 2017-12-15 | 电子科技大学 | A kind of zebra line detecting method based on maximum stable region and stochastical sampling |
CN104851184B (en) * | 2015-05-27 | 2017-06-30 | 深圳怡化电脑股份有限公司 | A kind of recognition methods of horizontally-spliced bank note and device |
CN105184950A (en) * | 2015-06-03 | 2015-12-23 | 深圳怡化电脑股份有限公司 | Method and device for analyzing banknote to be old or new |
CN105006058B (en) * | 2015-07-24 | 2017-09-29 | 深圳怡化电脑股份有限公司 | A kind of method and device for recognizing bank note adulterium information |
CN105184955B (en) * | 2015-08-18 | 2018-03-02 | 深圳怡化电脑股份有限公司 | A kind of recognition methods of bank note and device |
CN105632015B (en) * | 2015-12-24 | 2019-03-01 | 中钞实业有限公司 | A kind of bank money fingerprint characteristic anti false authentication method based on cloud platform |
CN106504409B (en) * | 2016-11-16 | 2019-07-05 | 深圳怡化电脑股份有限公司 | A kind of detection method and device of bank note magnetic security line |
CN106875546B (en) * | 2017-02-10 | 2019-02-05 | 大连海事大学 | A kind of recognition methods of VAT invoice |
CN106780966B (en) * | 2017-02-17 | 2019-10-11 | 深圳怡化电脑股份有限公司 | A kind of paper money discrimination method and device |
CN107330472A (en) * | 2017-07-06 | 2017-11-07 | 南开大学 | A kind of automatic identifying method of unmarked model animal individual |
CN107705416B (en) * | 2017-09-11 | 2019-12-06 | 深圳怡化电脑股份有限公司 | paper money image processing method, paper money image processing device and electronic equipment |
CN108711213B (en) * | 2018-03-30 | 2020-01-14 | 深圳怡化电脑股份有限公司 | Method and device for identifying black and white blocks of paper money zebra stripes |
CN110135491A (en) * | 2019-05-13 | 2019-08-16 | 四川中新华搜信息技术有限公司 | A kind of exterior of commodity method for anti-counterfeit of image or video identification based on SVM |
CN111428748B (en) * | 2020-02-20 | 2023-06-27 | 重庆大学 | HOG feature and SVM-based infrared image insulator identification detection method |
CN111915792B (en) * | 2020-05-19 | 2022-06-07 | 武汉卓目科技有限公司 | Method and device for identifying zebra crossing image-text |
CN112735021B (en) * | 2020-12-31 | 2022-08-16 | 沈阳中钞信达金融设备有限公司 | Banknote zebra crossing identification and counterfeit detection method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102289672A (en) * | 2011-06-03 | 2011-12-21 | 天津大学 | Infrared gait identification method adopting double-channel feature fusion |
CN202171834U (en) * | 2011-06-28 | 2012-03-21 | 浙江金利电子有限公司 | Currency detector provided with counterfeit-checking sensor with currency counting and checking functions |
CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8750573B2 (en) * | 2010-08-02 | 2014-06-10 | Sony Corporation | Hand gesture detection |
-
2014
- 2014-08-20 CN CN201410415338.3A patent/CN104298989B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102289672A (en) * | 2011-06-03 | 2011-12-21 | 天津大学 | Infrared gait identification method adopting double-channel feature fusion |
CN202171834U (en) * | 2011-06-28 | 2012-03-21 | 浙江金利电子有限公司 | Currency detector provided with counterfeit-checking sensor with currency counting and checking functions |
CN103902976A (en) * | 2014-03-31 | 2014-07-02 | 浙江大学 | Pedestrian detection method based on infrared image |
Non-Patent Citations (6)
Title |
---|
HOG在人脸识别中的性能研究;向征 等;《计算机工程》;20120805;第38卷(第15期);194-196 * |
Pedestrian Detection using Infrared images and Histograms of Oriented Gradients;F. Suard 等;《Intelligent Vehicles Symposium 2006》;20060615;摘要,第207-209页第II、III节 * |
一种基于 PSO 的 RBF-SVM 模型优化新方法;徐海龙 等;《控制与决策》;20100315;第25卷(第3期);367-370、377 * |
基于卷积运算的纸币红外图像斑马线特征提取;胡雪娟 等;《激光与红外》;20121129;第42卷(第10期);第1197页第3段,第1197-1198页第2节 * |
基于改进的网格搜索法的 SVM 参数优化;王健峰 等;《应用科技》;20120615;第39卷(第3期);28-31 * |
基于模拟谐振子的优化 K-means聚类算法;于海涛 等;《计算机工程与应用》;20121021;第48卷(第30期);第123页第1节 * |
Also Published As
Publication number | Publication date |
---|---|
CN104298989A (en) | 2015-01-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104298989B (en) | False distinguishing method and its system based on zebra stripes Infrared Image Features | |
He et al. | Multi-patch convolution neural network for iris liveness detection | |
Yilmaz et al. | Offline signature verification using classifier combination of HOG and LBP features | |
Laavanya et al. | Real time fake currency note detection using deep learning | |
CN102722708B (en) | Method and device for classifying sheet media | |
Tessfaw et al. | Ethiopian banknote recognition and fake detection using support vector machine | |
Kruthi et al. | Offline signature verification using support vector machine | |
CN106056752A (en) | Banknote authentication method based on random forest | |
Jadhav et al. | Currency identification and forged banknote detection using deep learning | |
Dhar et al. | Paper currency detection system based on combined SURF and LBP features | |
Dittimi et al. | Multi-class SVM based gradient feature for banknote recognition | |
Desai et al. | Cnn based counterfeit Indian currency recognition using generative adversarial network | |
Padmaja et al. | Indian Currency Denomination Recognition and Fake Currency Identification | |
Kumar | Signature verification using neural network | |
CN103035061A (en) | Anti-counterfeit characteristic generation method of valuable file and identification method and device thereof | |
Muhamad et al. | Image-Based Processing of Paper Currency Recognition and Fake Identification: A Review | |
Al-Maqaleh et al. | An efficient offline signature verification system using local features | |
Hajare et al. | Face Anti-Spoofing Techniques and Challenges: A short survey | |
Zainuddin et al. | Malaysian coins recognition using machine learning methods | |
Yu et al. | Two strategies to optimize the decisions in signature verification with the presence of spoofing attacks | |
Shinde et al. | Identification of fake currency using soft computing | |
Spencer et al. | Presentation attack detection using convolutional neural networks and local binary patterns | |
Antonius et al. | DeepCyberDetect: Hybrid AI for Counterfeit Currency Detection with GAN-CNN-RNN using African Buffalo Optimization | |
KR102394894B1 (en) | Banknote and coin detection device and method based on deep learning | |
Naseem et al. | Counterfeit Recognition of Pakistani Currency |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |