CN103927539A - Efficient feature extraction method for off-line recognition of Uyghur handwritten signature - Google Patents
Efficient feature extraction method for off-line recognition of Uyghur handwritten signature Download PDFInfo
- Publication number
- CN103927539A CN103927539A CN201410112023.1A CN201410112023A CN103927539A CN 103927539 A CN103927539 A CN 103927539A CN 201410112023 A CN201410112023 A CN 201410112023A CN 103927539 A CN103927539 A CN 103927539A
- Authority
- CN
- China
- Prior art keywords
- signature
- feature
- refinement
- dimension
- image
- 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.)
- Pending
Links
Landscapes
- Collating Specific Patterns (AREA)
- Character Discrimination (AREA)
Abstract
The invention belongs to the field of mode recognition, and relates to efficient extraction of features of a handwritten signature in the recognition process of the Uyghur handwritten signature. In the feature extraction process, pyramid type resolution subimage segmentation is carried out on each signature image, improvement is carried out according to the features of the Uyghur handwritten signature, 16-dimension direction features of a high-resolution layer are extracted, and extraction of multi-dimension local central point features including 16-dimension local central point features, 32-dimension local central point features and 64-dimension local central point features is carried out on a low-resolution layer. It is the first time that the 16-dimension direction features and the multi-dimension local central point features are fused, and classification and discrimination are carried out through a K-NN classifier and a chi-square distance measurement method. When the 16-dimension direction features are fused with the 64-dimension local central point features, an experiment is carried out on 1500 signatures of 75 persons, and the recognition rate is 100%. The experiment result represents that according to the feature extraction method, the features of the Uyghur handwritten signature can be effectively captured, and the feature extraction method has popularization value in off-line handwritten signature technologies of other similar characters such as Kazakh.
Description
Technical field
The invention belongs to computer patterns identification field, can be used for extracting in Uighur handwritten signature identifying high efficiency handwritten signature feature.
Background technology
Handwritten signature is the biological characteristic that a kind of use is widely used in qualification identity, is the generally accepted a kind of mode of agreeing to or authorizing of social life.In ecommerce banking, process document, the field such as sign a contract is widely used, and has corresponding legal effect.But due to its simple, easy imitation, also become the target of forgery, now handwritten signature qualification will be brought into play its effect.If correct recognition rata is high, will comprise that to social every profession and trade the aspects such as administration, finance, processing legal dispute, security fields can play a key effect, and therefore play a significant role in national economic construction.Signature qualification is also the advanced subject in current computer patterns identification field.In view of signature, qualification has a good application prospect and huge commercial value, and the many scholars in countries in the world and research institution have all shown very big interest to it.Therefore the further investigation to signature technology, has an important significance for the scientific-technical progress that improves actual application level and Relevant Subjects.
Signature generally can obtain by two kinds of approach, that is: online mode and offline mode.Correspondingly, signature is differentiated and also can be divided into that on-line signature is differentiated and two kinds of static signature discriminatings.During on-line signature is differentiated, signature is expressed as one or more time dependent bursts, wherein comprises the multidate information of signature.And in static signature is differentiated, signature is that the form of the two dimensional image feature to be extracted is described, differentiate therefore be also referred to as static signature.In addition,, according to the difference of identity identification system mode of operation, signature can be differentiated and be divided into signature identification (Signature Recognition) and signature verification (Signature Verification) two classes again.Off-line signature recognition is substantially the same in these two steps of pre-service and feature extraction with static signature checking, and their differences are mainly being differentiated in decision-making.Under Validation Mode, system the is write people's identity true and false to signing authenticates, and is two classification problems.Identifying purpose is from reference signature Sample Storehouse, to find out the label writer that given signature is corresponding, is many classification problems.The leading indicator of system performance is correct recognition rata and mates an averaging time that feature is required.
Because signature is differentiated and had a good application prospect and huge commercial value, abroad to existing more than 30 year of its research, and having obtained the discriminating of many achievements, particularly on-line signature, oneself is quite ripe.Because the information that static signature provides is less, difficulty is larger, and on-line signature achievement in research is fewer relatively, does not also have practical static signature identification system to come out.
For many years, the research of domestic scholars mainly concentrate on signature differentiate (checking), and computer handwriting identify do not go up, little for the research of off-line signature recognition.
The research of differentiating for signature, domestic scholars has proposed new method, the document of delivering mainly comprises: " the Application of HMMs for Signature Verification " of " Pattern Recognition ", " Model-based signature verification with rotation invariant features ", " Off-line Signature Verification using Structural Feature Correspondence ".The method that this type of technology utilizes respectively hidden Markov model (Hidden Markov Models, HMM) or static structure feature and pseudo-dynamic structure characteristic to combine is carried out off-line handwritten signature checking." the Off-line Chinese signature verification based on support vector machines " of " Pattern Recognition Letters " stroke width distribute, four kinds of features such as intensity profile combine, use again standard SVM, in the error rate control 5% that off-line type characters signature is differentiated.In " structure of Spline Wavelet Basis and the application in characters signature is differentiated thereof " of " the Wavelet-based Off-line Handwritten Signature Verification " of " Computer Vision and Image Understanding " and " computer engineering ", respectively the method for small echo is used in online Chinese signature identification system, but their method is not suitable for off-line system.In " research of the off line Chinese signature identification system based on neural network " of " computer utility ", use the feature combining based on Gabor and Zernike, for the three layer feedforward neural networks of feature identification, obtain more than 93. 70% discriminations.
" the wavelet packet hidden Markov model of line Handwritten signature identification " of " computer utility " proposed a kind of line Handwritten signature identifying method based on wavelet packet hidden Markov, test in 184 test sample books of Chinese, discrimination and misclassification rate are respectively 92% and 6.25%.But these technology are all differentiated for signature, and research object is English signature or Chinese signature.
For computer handwriting Study of recognition, domestic scholars has proposed new method, the document of delivering mainly comprises: " based on Multichannel Decomposition and the person's handwriting Study of recognition of mating " of " robotization journal ", " based on the identification of person's handwriting ", " a kind of multi-lingual text-independent writing discrimination method based on microstructure features " etc.This type of technology is utilized respectively person's handwriting architectural feature or person's handwriting textural characteristics, and obtains good recognition effect with all kinds of distances (Euclidean distance, weighted euclidean distance, card side's distance etc.) sorter.The sensitivity of such person's handwriting recognition methods is compared classic method and is increased, but research object is the identification of handwriting, different with signature identification.
The external Off-line Handwriting Signature Verification Technology of comparing, in " the Off-line Arabic signature recognition and verification " delivering for 2000 at " Pattern Recognition ", adopt global characteristics and local feature combination, adopted the method for multistratum classification to obtain 98% discrimination to Arab's signature.In " the Support vector machines versus multi-layer perceptrons for efficient off-line signature recognition " delivering for 2006 at " Engineering Applications of Artificial Intelligence ", research support vector machine (Support Vector Machine, SVM) process static signature qualification problem, and obtain than multilayer perceptron (Multi-Layer Perceptron, MLP) higher identification result, its resolution reaches 71.2%.In " Offline signature recognition using neural networks approach " in " Procedia Computer Science " in 2011, utilize the sorting technique based on artificial neural network to test in 1000 English samples, obtain 93.0% discrimination.In " the Hindi and English off-line signature identification and verification " delivering for 2012 at " Advances in Intelligent Systems and Computing ", extract improved Gradient Features, utilize svm classifier device, 2160 Hindi languages and English signature sample of mixing are tested to acquisition 98.05% discrimination.Srikanta Pal in 2012 etc. propose the feature based on background and foreground information, in the signature sample of two kinds of mixing (English and Chinese signature 1680 samples, English and Bengali sign 1800 samples), test the discrimination that obtains respectively 97.70% and 99.41%.
Also be about the technology general character of handwritten signature above: correlative study mostly concentrates on " checking ", about the research of " identification " fewer; Its research object is English signature, Chinese signature, Arabic signature, Hindi language signature and Bengali label, does not almost have for the research report of Uighur signature identification.Compare with other word signatures, the signature of Uighur has singularity (be made up of 32 letters, belong to agglutinative language type on morphosis) and has brought very large difficulty for pre-service and the feature extraction of handwritten signature image.Therefore, need to be at existing English and Chinese signature, particularly, on the basis of the authentication technique of Arabic signature image, further research is for method and the technology of the links such as the pre-service of Uighur signature image, feature extraction, characteristic matching and decision-making judgement.
Uighur, as one of municipal official of Xinjiang of China Uygur word, is using widely in this area.Current Economy in Xinjiang develop rapidly, expanding economy and international association day by day frequent, there are various agreements, contract, regulation and check etc. all to need litigant's signature in financial world and government's session, imitated or the imitation if sign, to cause serious societal consequence and huge economic loss, serious more can destroy whole financial order and social stability.Therefore, to signature carry out effectively, reliably, fast qualification there is important social value and Practical significance, it is studied is very necessary.Because the dispute case that signature causes is more and more, the correct technology of differentiating the signature true and false also seems more and more important.Develop a set of fast, accurately and reliably and signature identification system easy to use this active demand of solving just society.Therefore, studying and develop ethnic group's signature identification system becomes when previous important topic anxious to be resolved, has great political significance and realistic meaning.
The present invention makes up this leak in Chinese Minority Nationalities signature identification identification system just, particularly this leak in Uighur signature identification field, in conjunction with the special features of Uighur signature, propose to meet the efficient feature extracting method of one of Uighur handwritten signature text structure,, the signature image later to refinement extracts direction character and local center feature, calculates the method for similarity by card side's distance, obtains good recognition effect.The present invention, can also be to bank, insurance except public security, procuratorial work, the juridical authorities in an Xinjiang hundreds of counties and cities, area are used, and the other fields such as financial institution are promoted the use of.
Summary of the invention
The object of the invention is in order to solve off-line type Uighur handwritten signature identification problem, and propose a kind of efficiently, for the feature extracting method of off-line Uighur handwritten signature.
The object of the invention is to be achieved through the following technical solutions:
Thinning processing to signature and the attribute combination of feature, then extract the method for feature, its feature is:
1) pretreatment stage signature thinning processing;
2) extraction meets each category feature of the word feature of Uighur handwritten signature;
3) according to the attribute of the feature that will extract, the signature of some feature before refinement, extract, and extract in some feature signature that to be refinement later, determine the validity feature before and after refinement;
4) all kinds of Fusion Features that extract, utilize the Various Classifiers on Regional identification of signing.
The pretreatment stage of signature, uses the method for normalizing that meets Uighur signing structure.When signature refinement, utilize Rosenfeld thinning algorithm.
For the refinement problem of signature, on the basis of the impact of research refinement on signature identification, determine the validity feature before and after refinement:
1) refinement after signature extract after feature, signature discrimination increases, determine this category feature be concerning refinement signature effectively;
2) refinement after signature extract after feature, signature discrimination declines to some extent, determine this category feature be the signature before refinement effectively.
Determine the validity feature before and after refinement by a large amount of signature identification experiments:
1) be best selection for extracting direction character in the signature image after refinement, the obvious height (approximately improve 22.0%) of discrimination before than refinement, for refinement signature, direction character is validity feature;
2) local center feature is to be validity feature for the signature image before refinement, this feature discrimination extracting in the signature image from refinement decline to some extent (approximately declining 5.6%).
Every width signature image is carried out to the cutting of pyramid resolution subimage, enter according to the feature of Uighur signature
Row improves, and resolution layer has been extracted to 16 dimension direction characters, and low-resolution layer has been extracted to multidimensional (16 dimensions, 32 peacekeepings 64 are tieed up) local center feature.
Multidimensional (16 dimension, 32 peacekeepings 64 tie up) the local center feature of extracting and 16 dimension direction character direction characters merge respectively, with the Various Classifiers on Regional identification of signing.
Utilize K-NN sorter, when distance metric, carry out discriminant classification by Euclidean distance and Ka Fang distance metric method.
beneficial effect
The present invention contrasts prior art and has following innovative point:
1) pretreatment stage (before feature extraction) carries out thinning processing to signature;
2) every width signature image is carried out to the cutting of pyramid resolution subimage, resolution layer has been extracted to 16 dimension direction characters, low-resolution layer has been extracted to totally 32 multidimensional (16 dimensions, 32 peacekeepings 64 are tieed up) dimension local center feature.These features are improved according to the feature of Uighur signature;
3) identification of signing after direction character is combined with local center feature.
The present invention contrasts prior art and has following remarkable advantage:
1), according to the attribute of the feature that will extract, the signature of local center feature before refinement, extract, and extract in the direction character signature that to be refinement later.Then, this two feature is in conjunction with the discrimination that has significantly improved Uighur handwritten signature later;
2), in the feature merging, the dimension that increases local center characteristic direction feature further improves the discrimination of Uighur handwritten signature later.
Brief description of the drawings
Fig. 1 is signature identification based on local center feature and the schematic diagram of the impact of refinement on this recognition methods;
Fig. 2 local center feature extraction schematic diagram and signature image central point;
Fig. 3 local center feature extraction algorithm process flow diagram;
Discrimination correlation curve figure before and after the refinement of Fig. 4 local center feature; Wherein, 1 is the discrimination curve before refinement; 2 is the later discrimination curve of refinement;
Fig. 5 is the schematic diagram based on directional characteristic signature identification and the impact of refinement on this recognition methods;
The directional characteristic feature extraction process flow diagram of Fig. 6;
Fig. 7 is to feature extraction schematic diagram and four kinds of direction of scanning schematic diagram in the signature image before and after refinement:
(a) four of bianry image net regions; (b) four of bianry image kinds of direction of scanning;
(c) four of refined image net regions; (d) four of refined image kinds of direction of scanning;
Identification discrimination curve map before and after the refinement of Fig. 8 direction character; Wherein, 1 is the discrimination curve before refinement; 2 is the later discrimination curve of refinement;
Fig. 9 is the schematic diagram of the embodiment of the present invention.
embodiment
Below in conjunction with drawings and Examples, the invention will be further described.
Basic thought of the present invention is that multidimensional (16 dimensions, 32 peacekeepings 64 are tieed up) the local center feature and the 16 dimension direction characters that extract Uighur signature merge respectively, utilize K-NN sorter, when distance metric, carry out discriminant classification by card side's distance metric method and obtain high-precision discrimination.
embodiment 1
This example and claims 1,2,4,5,7 corresponding, is the signature identification based on local center feature, and the impact of refinement on this signature identification, relevantly implements flow graph as shown in Figure 1.Signature identification mainly comprises pre-service, the three phases compositions such as feature extraction and classification.Each link of simple declaration Fig. 1 is used below method and algorithm.
1) pre-service
Pre-service of the present invention comprises elimination noise, binaryzation, normalization and the refinement of signature.
1. eliminate noise.Signature image is input in computing machine because image can bring noise in digitized process by scanner, due to the inkblot that paper degree of roughness is different or stay because of carelessness while writing, this brings interference to other pre-service and the feature extraction work that will carry out later, if can not dispose well these noises, when feature extraction, just likely think this noise spot.Be an extention of Uighur signature, this can bring huge difficulty to the discrimination of signature.The performance of whole system will be affected.The present invention is the characteristic with the feature that will extract according to the actual conditions of signature sample, must well process these noise spots.Utilize recursive algorithm, in statistics signature image, least part namely has the number of continuity point in the extention in the Uighur of possibility of judging noise spot, judges by selecting certain threshold value.Arrange and judge that the length threshold of discrete noise is 20.If a number for the stain being connected with investigation point is less than 20, thinks that investigating point is noise spot, and all discrete noise points are filled to white.
2. binaryzation.Binarization methods is by setting certain threshold value, the image conversion with gray level is become to only have to the black white image of two gray levels
].Suppose that input picture is
matrix
, output image is
matrix
, given threshold is
, the algorithm of image binaryzation is so:
(1-1)
The key of image binaryzation computing is how to set threshold value.Want first signature image to be converted into gray level image with best threshold value, then obtain the grey level histogram of image, set suitable threshold value according to the statistical information of grey level histogram signature image is converted into bianry image.The present invention adopts and selects the method for threshold value, and by the distribution situation of the gray level of the signature image after gray processing, gray level is greater than 200 pixel and is set to 255, i.e. white point.Gray level is less than 200 pixel and is set to 0, i.e. stain, realizes image binaryzation processing.
3. normalization.Signature image normalization be by image in proportion linearity zoom in or out given size, be convenient to the unified feature of describing and extracting signature.It is very important pre-service part of the present invention that signature image is carried out to size normalized.When the name of different signer label oneself, the size of multiple signatures can difference, and this feature extraction work meeting for next step has a certain impact.If classifying by the size of signing messages is that discrimination is not necessarily fine.This is not typical effective ways.Line number and the columns of a width signature image are very large simultaneously, if first the size of image is not normalized, so whole computation process will be very huge.Therefore be normalized and can reduce unnecessary operand.
Image is carried out to the normalized detailed process of size as follows:
A) be partitioned into the concrete signature part of image, remove the white space of image periphery;
B) the concrete signature part of image is carried out to translation, size normalization is carried out in the conversion of convergent-divergent equiaffine.
Signature image f (x, y) is carried out to translation, the normalized step of yardstick:
Calculate
barycenter
with scale factor a, will
do translation and change of scale obtains
, be designated as new
.Wherein
,
,
/ T,
deng the moment of the orign that is image, T is the picture size that expectation reaches.Native system is after being normalized the size of concrete signature part, and signature image length is 384 pixels, and width is 96 pixels.
4. refinement.What in off-line type signature identification problem, we mainly considered is the static nature of signature image.So it is features of shape and the unnecessary quantity of information of minimizing to outstanding signature that signature image is carried out to thinning processing, this signature image that scanning is come in carries out the benefit of thinning processing.The object of signature image refinement is to extract the skeleton of original image, is that lines that line thickness in original image is greater than to 1 pixel are refined into and only have a pixel wide, forms on ' skeleton '.
To after complete a slice image thinning, should meet several conditions:
A) signature image after refinement should with original image equivalence;
B) signature image after refinement is the center line of original image as far as possible;
C) signature image after refinement should keep the connectedness of image thin;
D) live width of the signature image after refinement should be a pixel;
E) signature image after refinement is to original image insensitive for noise.
What the present invention adopted is Rosenfeld thinning algorithm.Rosenfeld algorithm is a kind of parallel thinning algorithm, and the result to current pixel process according to the previous round iterative processing of each pixel in this pixel and neighborhood thereof, to beginning to adopting eventually identical refinement criterion.After all having detected the point on man image border, delete and can delete a little again.The matrix morphology of gained is that 8-connects, and is used in 0-1 bianry image.
2) feature extraction
This example is the signature identification based on local center feature, and therefore the extraction of local center feature is described here.
The side-play amount that Uighur is signed in level and vertical direction is larger.Everyone signature has unique structure and writing style, and therefore multiple signatures of a people are mutually more approaching on the whole.In order to reflect better this writing style of signature, the local center feature that Arabic signature extracts used for reference in literary composition of the present invention, in conjunction with Uighur signature ground feature, improve the slit mode of image, signature image after every width normalization is carried out to the cutting of pyramid resolution subimage, low-resolution layer has been extracted to totally 32 dimension local center features.Experimental result shows, this feature can effectively reflect the signature style of Uighur signer.
Under regard to this feature and give at length to introduce.First signature image is divided into 2 × 8 little rectangular windows by (as shown in Figure 2), then the signature section of each rectangular window is carried out respectively to horizontal and vertical projection.Further calculate the central point of each window, using the horizontal ordinate of these 16 central points and ordinate as feature, so just formed 32 dimensional features, local center proper vector corresponding to someone is:
（1-2）
Wherein j=1,2,3 ..., 10 or all training samples of 16(someone), c representation feature dimension, its value is that 32, p is the code name that participates in the someone of training.For the people of all participation training, corresponding low resolution image layer proper vector is:
（1-3）
Wherein, n is the total number of persons that participates in training.If the size of subimage is
(
mfor the width of subimage,
kfor the height of subimage), the horizontal projection of Signature Curve section Z (x, y) in window
and vertical projection
for:
(1-4)
Wherein,
black pixelfor the black pixel of signature image.The center point coordinate of each window is:
(1-5)
Wherein
with
be respectively horizontal stroke and the ordinate of central point;
with
be respectively horizontal projection and the vertical projection of each wicket.Give central point feature extraction schematic diagram and local center below as shown in Figure 2.The feature extraction process flow diagram of this feature, as shown in Figure 3.
3) classification
At sorting phase, extract after effective Uighur handwritten signature feature, the sorter of finding out applicable feature is and is important.Distance between two proper vectors is the fine tolerance of one of their similarities.If other sample of corresponding same class flocks together in feature space, and different classes of sample is mutually from away from must be, and classification is just than being easier to realization.
Sorter, for the difference of model, has multiple branch at present, comprising: Bayes sorter, BP neural network classifier, decision Tree algorithms, support vector machine (Support Vector Machine, SVM) algorithm, distance classification device etc.The several method such as KNN sorter and distance classification (Euclidean distance and Ka Fang distance metric method) that the present invention utilizes is classified.
1. K-NN sorter.The computation process of K-NN similarity is all to start to carry out in the time of classification, and assorting process need to be calculated the distance between sample set to be sorted and training sample set.KNN sorting algorithm has Cover and Hert and proposes for 1986, is one of rudimentary algorithm of instance-based learning, is also one of non-parametric important method of pattern-recognition.Its basic thought is: for a test sample book, calculate it and training sample and concentrate the similarity of each sample, find out k training sample the most similar according to Sample Similarity, then give on this basis the marking of each sample, score value is the Sample Similarity sum belonging in k training sample between such sample and test text.Sort by score value, according to the classification of score value nominative testing sample.For signature classification, all samples of the method retainer of a big family are corresponding to a point in n-dimensional vector space.Treat a point sample set for one, calculate the similarity of it and training sample set.Nearest neighbor method calculates the distance of sample to be sorted to all representative points.Be classified to the affiliated classification of nearest representative point.In order to overcome the defect that nearest neighbor method false determination ratio is higher, arest neighbors is generalized to k neighbour, it chooses k the representative point nearest from sample to be sorted, which kind of belongs to according to k representative point majority, just by test sample book meeting and such.Also can say, a given test sample book to be sorted, system is searched the sample of a most similar k arest neighbors in training set, and carrys out the candidate's classification scoring to this sample according to situation under these neighbours' classification.Can be neighbour's sample class weight.If the part sample in this k neighbour belongs to same class, the class of the each neighbour in such power sum is as the similarity of this classification and test sample book.By the sequence to the scoring of candidate's class, then provide a threshold value, just can judge the classification of test document.
In K-NN, the decision rule of sorter can represent:
（1-6）
Wherein,
represent sample
whether belong to class
(being y=1, no y=0);
Represent test sample book x and training sample
similarity, wherein
one of k arest neighbors of x;
it is the threshold value of decision-making.Generally, similarity function
adopt the cosine value of vector angle to represent:
（1-7）
Wherein, the dimension that m is proper vector,
for the k dimension of vector.
The advantage of KNN method is directly perceived and is convenient to understand and application, very effective in its practical application, is one of sample classification algorithm being most widely used at present.But, the shortcoming of KNN method is also very obvious, and that is exactly that calculation cost is very high, and each test sample book is wanted and all training samples carry out distance calculating, and the time complexity of single test sample book is O(mn) (dimension that m is feature space, the number that n is training sample).In addition, adopt KNN method to need choose reasonable k, the selection of k has determined the quality of classification performance to a great extent, in actual applications, generally adopts the method for cross-certification to select relatively optimum k value.
2. distance classification device.Distance classification device is with respect to other sorter, more intuitively and more effective.The directly distance of more each sample to be identified and training sample, selects wherein and the sample of feature database bee-line, and distance classification device is generally applied in the pattern recognition problem that intrinsic dimensionality is many.Some conventional distance classifications have: Euclidean distance, card side's distance, Weighted distance, mahalanobis distance etc.
Euclidean distance (Euclidean distance) also claims that Euclidean distance is a distance definition conventionally adopting, it is the actual distance between two points in m-dimensional space, the difference between two is the quadratic sum square root again of each variable value difference, and object is that the overall distance of calculating is therebetween dissimilarity.Euclidean distance in two and three dimensions space be exactly the distance between 2, two dimension formula be:
（1-8）
Three-dimensional formula is:
（1-9）
Be generalized to n-dimensional space, the formula of Euclidean distance be for:
（1-10）
Here i=1,2..n,
represent the i dimension coordinate of first point,
the i dimension coordinate n dimension Euclidean space that represents second point is a point set, its each point can be expressed as (x (1), x (2) ..., x (n)), wherein x (i) (i=1,2...n) is real number, is called i the coordinate of x, two some x and y=(y (1), y (2) ..., y (n)) between distance
be defined as formula above;
Euclidean distance is regarded the similarity degree of sample as, and distance is nearer just more similar.The proper vector of supposing certain sample to be identified is:
(1-11)
The proper vector of supposing training sample is:
(1-12)
Wherein the value of N is 3,16,17,32,33,48 and 128 5 kind of situation, and be the intrinsic dimensionality of different characteristic.The Euclidean distance formula calculating between them is:
(1-13)
In formula
with
represent respectively sample to be identified
and training sample
each dimension element.Wherein R representation feature vector dimension;
Card side's distance (Chi-square distance) measure is also one of conventional sample similarity measure.Suppose that having the proper vector of two signature sample is X and Y, the card side between them apart from computing formula is:
（1-14）
In formula
with
represent respectively sample to be identified
and training sample
each dimension element.Wherein R representation feature vector dimension.
4) experimental result
Experimental data selects to utilize the sample (everyone 20 samples) of 50 signers (all ages and classes and sex) from Uighur signature sample storehouse, and totally 1000 signature sample are tested.These signatures are divided into two parts, and a part is for training, and another part is for test.The present invention trains respectively 800 samples and 500 samples to test and contrast.Other sample is for test.
KNN sorter and distance classification (by Euclidean distance and Ka Fang distance metric method) that the present invention utilizes are classified.Because the discrimination of card side's distance metric method is higher than the method based on Euclidean distance, this example only provides the discrimination of card side's distance metric method.Extract after 32 dimension local center features, use the recognition result of card side's distance as shown in following table 1.
Table 1: the recognition result of local center characteristic use card side distance
The discrimination that can find out local center feature from table 1 is not high, reaches as high as 90.5%.In order to determine the validity of this feature to refinement, under identical experiment condition, respectively to extracting 32 dimension local center features in the signature image before refinement and after refinement, use based on KNN sorter and Ka Fang distance metric method 1000 Uygur's signature sample are identified to (accompanying drawing 1 realize and dotted line flow process).The contrast of their discriminations is shown in accompanying drawing 4.Can be found out by accompanying drawing 4, while extraction in the signature image from refinement, this feature discrimination declines to some extent, and portion's central point feature is to be validity feature for high density normalization two-value signature image, and while extracting this feature, pretreatment stage does not need refinement.
embodiment 2
This example and claims 1,2,4,5,7 corresponding, is the signature identification based on local center feature, and the impact of refinement on this signature identification, relevantly implements flow graph as shown in Figure 5.Signature identification mainly comprises pre-service, the three phases compositions such as feature extraction and classification.Each link of simple declaration Fig. 5 is used below method and algorithm.
1) pre-service
The preprocess method of this example is the same with the above embodiments 1, no longer explanation.
2) feature extraction
The present invention is set out by the Uighur own characteristic of signing, and has proposed a kind of 16 dimension direction characters, is a kind of statistical nature.This feature is after signature image is cut apart, to the image pixel dot information of the each direction of each range statistics obtaining.This feature is that the two-value signature image after highly dense bianry image and the refinement from normalization respectively extracts.Prove by experiment to form that in the signature image after skeleton, to extract the discrimination of directional characteristic discrimination before than refinement high.The feature extraction algorithm process flow diagram of this feature as shown in Figure 6.
Under regard to this feature and give at length to introduce.First, respectively signature image before and after every width refinement is carried out to the cutting of pyramid resolution subimage, each resolution layer, by four kinds of directions, is added up to all black pixel number of each direction, describe the signature style of signer by these data.Provide directional characteristic definition below.Press shown in accompanying drawing 7, the signature image before and after refinement is divided into four signature net regions by vertical direction, the size of each grid is all 96*96 mutually, and these form high-resolution subimage layer.Then, to each subimage (also referred to as net region), respectively with 0 °, 90 °, 135 °, 45 ° of four kinds of directions scan, and four kinds of direction of scanning are as accompanying drawing 7(a) and (b) as shown in.Add up respectively the continuous black pixel number of each direction, using the value of these four arrays as four-dimensional direction character.The last like this 16 dimension direction characters that obtain.Direction character vector corresponding to someone is:
（2-1）
Wherein
=1,2,3 ..., 10 or 16(someone's training sample),
value be 16, representation feature dimension, p is someone's the code name that participates in training.For the people of all participation training, corresponding full resolution pricture layer proper vector is:
（2-2）
Wherein, n is the total number of persons that participates in training.
3) classification
This example sorter used is the same with above-mentioned example 1, therefore no longer explanation.
4) experimental result
Experimental data selects to utilize the sample (everyone 20 samples) of 50 signers (all ages and classes and sex) from Uighur signature sample storehouse, and totally 1000 signature sample are tested.These signatures are divided into two parts, and a part is for training, and another part is for test.The present invention trains respectively 800 samples and 500 samples to test and contrast.Other sample is for test.
KNN sorter and distance classification (by Euclidean distance and Ka Fang distance metric method) that the present invention utilizes are classified.Because the discrimination of card side's distance metric method is higher than the method based on Euclidean distance, this example only provides the discrimination of card side's distance metric method.Extract after 16 dimension direction characters, use the recognition result of card side's distance as shown in following table 2.
Table 2: the recognition result of the direction character utilization side of card distance
Can find out that from table 2 directional characteristic discrimination is higher than the discrimination of local center feature, reach as high as 96%, still, this recognition result is to obtain in the signature image after refinement.In order to determine the validity of this feature to refinement, under identical experiment condition, respectively to extracting 16 direction characters in the signature image before refinement and after refinement, use based on KNN sorter and Ka Fang distance metric method 1000 Uygur's signature sample are identified to (accompanying drawing 1 realize and dotted line flow process).The contrast of their discriminations is shown in Fig. 8.As seen from Figure 8, best selection for extracting direction character in the signature image that (forms signature skeleton) after refinement, the obvious height (general 22% left and right) of discrimination before than refinement, therefore direction character is to be validity feature for the signature image after refinement, while extracting this feature, pretreatment stage needs refinement.
embodiment 3
This example and claims 1,4,5,6,7 corresponding, is the signature identification of 16 dimension direction characters and 32 dimension local center feature combinations, be also of the present invention be mainly one of example, it is relevant implements flow graph as shown in Figure 9.Signature identification mainly comprises pre-service, the three phases compositions such as feature extraction and classification.Each link of simple declaration Fig. 5 is used below method and algorithm.
1) pre-service
The preprocess method of this example is the same with the above embodiments 1, no longer explanation.
2) feature extraction
The preprocess method of this example is the same with the above embodiments 1, but extracts after 16 dimension direction characters and 32 dimension local center features combinations, in the identification of signing.Therefore no longer explanation here.
3) classification
This example sorter used is the same with above-mentioned example 1, therefore no longer explanation.
4) experimental result
The sample (everyone 20 samples) of selecting to utilize 50 signers (all ages and classes and sex) from Uighur signature sample storehouse, totally 1000 signature sample are tested.These signatures are divided into two parts, and a part is for training, and another part is for test.The present invention trains respectively 800 samples and 500 samples to test and contrast.Other sample is for test.
When feature extraction, every width signature image is carried out to the cutting of pyramid resolution subimage, feature according to Uighur signature is improved, resolution layer is extracted to 16 dimension direction characters (extracting after refinement), low-resolution layer is extracted to totally 32 dimension local center features (not needing refinement), then these feature combinations.The KNN sorter and the distance classification (by Euclidean distance and Ka Fang distance metric method) that utilize are classified.Because the discrimination of card side's distance metric method is higher than the method based on Euclidean distance, this example only provides the discrimination of card side's distance metric method.After 16 dimension direction characters and 32 dimension local center Fusion Features, use the recognition result of card side's distance as shown in following table 3.
After table 3:16 dimension direction character and 32 dimension local center Fusion Features, use the recognition result of card side's distance
As can be seen from Table 3, after 16 dimension direction characters and 32 dimension local center Fusion Features, discrimination improves, and the highest is that rate can not reach 98.5%.Therefore, the feature extracting method of this direction character and 32 dimension local center feature combinations effectively catches the word feature of Uighur signature, and this is also one of core content of the present invention.
embodiment 4
This example and claims 1,4,5,6,7 corresponding is main embodiment of the present invention.This example is the signature identification of 64 dimension direction characters and 32 dimension local center feature combinations, and its course of work mainly comprises pre-service, the three phases compositions such as feature extraction (containing Fusion Features) and classification.Each link of simple declaration Fig. 5 is used below method and algorithm.
1) pre-service
The preprocess method of this example is the same with the above embodiments 1, no longer explanation.
2) feature extraction
The preprocess method of this example is the same with the above embodiments 1, but extract after 16 dimension direction characters and 64 dimension local center features combinations, in the identification of signing, wherein 64 dimension local center feature extracting methods are substantially the same with 32 local center feature extractions, only here, point 32 rectangular window area such as signature images, then the central point in each region (horizontal, ordinate), as feature, just can extract 64 dimension local center feature (as accompanying drawing 3), therefore no longer explanations here.
3) classification
This example sorter used is the same with above-mentioned example 1, therefore no longer explanation.
4) experimental result
The sample (everyone 20 samples) of selecting to utilize 75 signers (all ages and classes and sex) from Uighur signature sample storehouse, totally 1500 signature sample are tested.These signatures are divided into two parts, and a part is for training, and another part is for test.The present invention trains respectively 1200 samples and 750 samples to test and contrast.Other sample is for test.
When feature extraction, every width signature image is carried out to the cutting of pyramid resolution subimage, feature according to Uighur signature is improved, resolution layer is extracted to 16 dimension direction characters (extracting after refinement), low-resolution layer is extracted to totally 64 dimension local center features (not needing refinement), then these feature combinations.The KNN sorter and the distance classification (by Euclidean distance and Ka Fang distance metric method) that utilize are classified.Because the discrimination of card side's distance metric method is higher than the method based on Euclidean distance, this example only provides the discrimination of card side's distance metric method.After 16 dimension direction characters and 64 dimension local center Fusion Features, use the recognition result of card side's distance as shown in following table 4.
After table 4:16 dimension direction character and 64 dimension local center Fusion Features, use the recognition result of card side's distance
As can be seen from Table 4, after 16 dimension direction characters and 64 dimension local center Fusion Features, discrimination improves, and the highest is that rate can not reach 100%.Therefore, the feature extracting method of this 16 dimension direction characters and 64 dimension local center feature combinations effectively catches the word feature of Uighur signature, and this is also one of core content of the present invention.
Claims (7)
1. the attribute combination of the thinning processing of pair signature and feature, then extract the method for feature, its feature is:
1) pretreatment stage signature thinning processing;
2) extraction meets each category feature of the word feature of Uighur handwritten signature;
3) according to the attribute of the feature that will extract, the signature of some feature before refinement, extract, and extract in some feature signature that to be refinement later, determine the validity feature before and after refinement;
4) all kinds of Fusion Features that extract, utilize the Various Classifiers on Regional identification of signing.
2. according to the pretreatment stage of the signature described in right 1, use the method for normalizing that meets Uighur signing structure; When signature refinement, utilize Rosenfeld thinning algorithm.
3. according to the refinement problem of signature described in right 2, on the basis of the impact of research refinement on signature identification, determine the validity feature before and after refinement:
1) refinement after signature extract after feature, signature discrimination increases, determine this category feature be concerning refinement signature effectively;
2) refinement after signature extract after feature, signature discrimination declines to some extent, determine this category feature be the signature before refinement effectively.
4. according to the validity feature before and after the definite refinement described in right 3:
1) be best selection for extracting direction character in the signature image after refinement, the obvious height (approximately improve 22.0%) of discrimination before than refinement, for refinement signature, direction character is validity feature;
2) local center feature is to be validity feature for the signature image before refinement, this feature discrimination extracting in the signature image from refinement decline to some extent (approximately declining 5.6%).
5. according to the feature extraction described in right 1, every width signature image is carried out to the cutting of pyramid resolution subimage, feature according to Uighur signature is improved, resolution layer is extracted to 16 dimension direction characters, low-resolution layer has been extracted to multidimensional (16 dimensions, 32 peacekeepings 64 are tieed up) local center feature.
6. merge respectively according to the multidimensional described in right 5 (16 dimension, 32 peacekeepings 64 tie up) local center feature and 16 dimension direction characters, with the Various Classifiers on Regional identification of signing.
7. according to the sorter described in right 6, utilize K-NN sorter, when distance metric, carry out discriminant classification by Euclidean distance and Ka Fang distance metric method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410112023.1A CN103927539A (en) | 2014-03-24 | 2014-03-24 | Efficient feature extraction method for off-line recognition of Uyghur handwritten signature |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410112023.1A CN103927539A (en) | 2014-03-24 | 2014-03-24 | Efficient feature extraction method for off-line recognition of Uyghur handwritten signature |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103927539A true CN103927539A (en) | 2014-07-16 |
Family
ID=51145755
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410112023.1A Pending CN103927539A (en) | 2014-03-24 | 2014-03-24 | Efficient feature extraction method for off-line recognition of Uyghur handwritten signature |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103927539A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104134066A (en) * | 2014-08-08 | 2014-11-05 | 科进生物识别(深圳)有限公司 | Recognition method for off-line signatures |
| CN104200239A (en) * | 2014-09-09 | 2014-12-10 | 河海大学常州校区 | Image feature fusion identification based signature authentic identification system and method |
| CN107992508A (en) * | 2017-10-09 | 2018-05-04 | 北京知道未来信息技术有限公司 | A kind of Chinese email signature extracting method and system based on machine learning |
| CN108416355A (en) * | 2018-03-09 | 2018-08-17 | 浙江大学 | A kind of acquisition method of the industry spot creation data based on machine vision |
| CN108564064A (en) * | 2018-04-28 | 2018-09-21 | 北京宙心科技有限公司 | A kind of efficient OCR recognizers of view-based access control model |
| CN108664975A (en) * | 2018-04-24 | 2018-10-16 | 新疆大学 | A kind of hand-written Letter Identification Method of Uighur, system and electronic equipment |
| CN108764155A (en) * | 2018-05-30 | 2018-11-06 | 新疆大学 | A kind of handwriting Uighur words cutting recognition methods |
| CN111553336A (en) * | 2020-04-27 | 2020-08-18 | 西安电子科技大学 | Print Uyghur document image recognition system and method based on link segment |
| CN111753804A (en) * | 2020-07-07 | 2020-10-09 | 深兰盛视科技（苏州）有限公司 | Recognition device, acquisition device, fee deduction device and non-contact biological recognition system for public transport |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7620244B1 (en) * | 2004-01-06 | 2009-11-17 | Motion Computing, Inc. | Methods and systems for slant compensation in handwriting and signature recognition |
| CN101866417A (en) * | 2010-06-18 | 2010-10-20 | 西安电子科技大学 | Method for identifying handwritten Uigur characters |
-
2014
- 2014-03-24 CN CN201410112023.1A patent/CN103927539A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7620244B1 (en) * | 2004-01-06 | 2009-11-17 | Motion Computing, Inc. | Methods and systems for slant compensation in handwriting and signature recognition |
| CN101866417A (en) * | 2010-06-18 | 2010-10-20 | 西安电子科技大学 | Method for identifying handwritten Uigur characters |
Non-Patent Citations (3)
| Title |
|---|
| VU NGUYEN ET AL.: "Off-line Signature Verification Using Enhanced Modified Direction Features in Conjunction with Neural Classifiers and Support Vector Machines", 《NINTH INTERNATIONAL CONFERENCE ON DOCUMENT ANALYSIS AND RECOGNITION》 * |
| 古丽热娜•阿布里孜: "维吾尔文脱机签名识别技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 古孜丽塔吉·乃拜: "基于多方向特征融合的维吾尔文笔迹鉴别技术", 《计算机工程与应用》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104134066A (en) * | 2014-08-08 | 2014-11-05 | 科进生物识别(深圳)有限公司 | Recognition method for off-line signatures |
| CN104134066B (en) * | 2014-08-08 | 2017-06-16 | 科进生物识别(深圳)有限公司 | For the recognition methods of static signature |
| CN104200239A (en) * | 2014-09-09 | 2014-12-10 | 河海大学常州校区 | Image feature fusion identification based signature authentic identification system and method |
| CN107992508A (en) * | 2017-10-09 | 2018-05-04 | 北京知道未来信息技术有限公司 | A kind of Chinese email signature extracting method and system based on machine learning |
| CN108416355A (en) * | 2018-03-09 | 2018-08-17 | 浙江大学 | A kind of acquisition method of the industry spot creation data based on machine vision |
| CN108664975A (en) * | 2018-04-24 | 2018-10-16 | 新疆大学 | A kind of hand-written Letter Identification Method of Uighur, system and electronic equipment |
| CN108664975B (en) * | 2018-04-24 | 2022-03-25 | 新疆大学 | Uyghur handwritten letter recognition method and system and electronic equipment |
| CN108564064A (en) * | 2018-04-28 | 2018-09-21 | 北京宙心科技有限公司 | A kind of efficient OCR recognizers of view-based access control model |
| CN108764155A (en) * | 2018-05-30 | 2018-11-06 | 新疆大学 | A kind of handwriting Uighur words cutting recognition methods |
| CN108764155B (en) * | 2018-05-30 | 2021-10-12 | 新疆大学 | Handwritten Uyghur word segmentation recognition method |
| CN111553336A (en) * | 2020-04-27 | 2020-08-18 | 西安电子科技大学 | Print Uyghur document image recognition system and method based on link segment |
| CN111553336B (en) * | 2020-04-27 | 2023-03-24 | 西安电子科技大学 | Print Uyghur document image recognition system and method based on link segment |
| CN111753804A (en) * | 2020-07-07 | 2020-10-09 | 深兰盛视科技（苏州）有限公司 | Recognition device, acquisition device, fee deduction device and non-contact biological recognition system for public transport |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103927539A (en) | Efficient feature extraction method for off-line recognition of Uyghur handwritten signature | |
| Tian et al. | Multilingual scene character recognition with co-occurrence of histogram of oriented gradients | |
| Liu et al. | A new benchmark on the recognition of handwritten Bangla and Farsi numeral characters | |
| WO2010092952A1 (en) | Pattern recognition device | |
| Zhang | Off‐line signature verification and identification by pyramid histogram of oriented gradients | |
| CN104200239A (en) | Image feature fusion identification based signature authentic identification system and method | |
| Samuel et al. | Novel feature extraction technique for off-line signature verification system | |
| Pal et al. | Off-line Bangla signature verification | |
| Chaabouni et al. | Multi-fractal modeling for on-line text-independent writer identification | |
| Raja et al. | Off-line signature verification based on fusion of grid and global features using neural networks | |
| Cruz et al. | Feature representation selection based on classifier projection space and oracle analysis | |
| Kaur et al. | Newspaper text recognition of Gurumukhi script using random forest classifier | |
| Tan et al. | A stroke shape and structure based approach for off-line chinese handwriting identification | |
| CN103942572A (en) | Method and device for extracting facial expression features based on bidirectional compressed data space dimension reduction | |
| Ubul et al. | Off-line Uyghur signature recognition based on modified grid information features | |
| Diaz-Cabrera et al. | Emerging issues for static handwritten signature biometrics | |
| Luo et al. | Alphanumeric character recognition based on BP neural network classification and combined features | |
| Roy et al. | Offline signature verification using grid based and centroid based approach | |
| Verma et al. | Offline signature verification and identification using angle feature and pixel density feature and both method together | |
| CN104899551A (en) | Form image classification method | |
| Kennard et al. | Offline signature verification and forgery detection using a 2-D geometric warping approach | |
| Zinjore et al. | Recognition of handwritten bilingual Characters-Numerals using shape context | |
| Rusinol et al. | CVC-UAB's Participation in the Flowchart Recognition Task of CLEF-IP 2012. | |
| CN102332088B (en) | Vote symbolic machine visual identification method based on run length feature | |
| Singh et al. | Survey on offline signature recognition and verification schemes |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20140716 |