CN102034094A - Digital ball identification method based on sparse representation and discriminant analysis - Google Patents

Abstract

The invention discloses a digital ball identification method based on sparse representation and discriminant analysis, which comprises the following steps of: putting each digital ball in a digital ball set under a monocolor background, continuously acquiring a single image or a plurality of images by using a single camera, automatically positioning the digital ball in each image, extracting visual features, and establishing sparse representation for all features to form a training sample feature set; putting one or more digital balls to be identified in the same scene, acquiring a single image or a plurality of images, automatically positioning all digital balls in each image, extracting sub-images, extracting visual features of the sub-images, and establishing the sparse representation of the sub-images by utilizing the training sample feature set; and identifying by adopting a discriminant analysis method, wherein for the condition of a plurality of images, joint posterior discrimination is adopted to improve identification precision. By fully utilizing the sparse representation and the discriminant analysis, the method is applied to identification of a single ball or a plurality of balls under the monocolor background and has good identification effect.

Description

A kind of digital ball recognition methods based on rarefaction representation and judgment analysis

Technical field

The invention belongs to image sparse and represent and the probability calculation field, specifically relate to a kind of digital ball recognition methods based on rarefaction representation and judgment analysis.

Background technology

Traditional character recognition is all operated in the plane, if but the method for these plane operations is used for curved surface, a lot of problems will appear.But a lot of things in our life not merely are identification on the plane, for curved surface not even the identification on the regular figure be a highly significant very practical again method.Digital ball is a kind of ball that is printed on unique numeral on its surface, so a kind of new method of our needs is discerned the numerical information on the sphere.We realize the identification of still picture earlier, then the digital ball of high-speed motion are taken continuously, discern again, and this is the problem that a very challenging property also has commercial value very much.

Owing to be to operate on the sphere, so will detect circle earlier, the location numeral then shows numerical information topmost this numerical information of still discerning then.Yet because digital ball exists three-dimensional rotation, problems such as different visual angles are so caused different observed results.In addition, by single camera, we can only obtain the partial information of digital ball, can not obtain complete sphere information and be used for identification.At last, when having a plurality of digital ball in the picture, it is very important how we accurately locate and remove the circle of those locations of mistake.So we proposed with sparse expression and repeatedly the measurement of probability distribution realize the identification of sphere numeral.

What in recent years, sparse expression was used in machine learning and pattern-recognition is more and more.Especially for handling high dimensional data, sparse method is very effective.Based on this technology of sparse expression, each sample can be expressed as the sparse linear combination of training data.When this optimizes expression when enough sparse, can be effective to address this problem based on the algorithm of protruding optimization.More famous sparse expression method has lasso, elastic network(s) (elastic net) and non-losing side method (nonnegative garrote).In this invention, the information of expressing digital ball with these three kinds of methods is used to classify.

After the information of having represented digital ball, be to have classified with that.All the time, classification all is the emphasis in the machine learning, and the method for classification also is diversified.Before machine learning put forward, main sorting technique was relevant.Extensive rise along with machine learning, increasing sorting technique has been arranged, as PCA (Shlens Jonathon, A Tutorial on Principal Component Analysis.Systems Neurobiology Laboratory, Salk Insitute for Biological Studies, 2009), Fisher judgement (Fisher, Ronald A., The use of multiple measurements intaxonomic problems.Annals Eugen., 1936.), linear discriminant analysis (LDA) (R.Duda, P.Hart, and D.Stork, Pattern classification, 2rd ed.Wiley-Interscience, 2000).Utilization of the present invention is repeatedly measured with probabilistic method and is classified, and reaches good digital ball discrimination by single even posteriority decision method repeatedly.

Summary of the invention

The invention provides a kind of digital ball recognition methods based on rarefaction representation and judgment analysis, this method recognition capability is strong, and recognition effect is good.

A kind of digital ball recognition methods based on rarefaction representation and judgment analysis comprises:

(1) be placed on each the digital ball in the digital ball set under the color background separately, utilize single camera continuous acquisition single width or multiple image, automatically locate the digital ball in every width of cloth image and extract visual signature, and all visual signatures are set up sparse expression, form the training sample characteristic set;

(2) one or more digital ball to be identified is placed in the same scene, gathers single width or multiple image, subimage is located and extracted to all the digital balls in every width of cloth image automatically; Subimage to corresponding same digital ball in single width or the multiple image extracts visual signature, and utilizes the training sample characteristic set to set up the sparse expression of this subimage;

(3) adopt the judgment analysis method to discern, obtain the affiliated classification of test pattern,, adopt the method for associating posteriority judgement to realize wherein for the situation of multiple image.

The method that forms the training sample characteristic set in the described step (1) is with single camera continuous acquisition single width or multiple image under color background, has only a digital ball among every width of cloth figure, the set of formation training sample, ball in the positioning image also extracts visual signature, set up the sparse expression of training sample set: each the digital ball in the digital ball set, be placed on separately under the color background (as black), utilize single camera continuous acquisition single width or multiple image to gather as training sample, each image that obtains all is a single-view, and concrete steps are as follows:

(a) image of gathering is done pre-service, use the Canny operator to carry out rim detection, obtain binary image, then provide round roughly radius, the method for using Hough transformation or circumscribed circle to construct coupling is located the digital ball position among every width of cloth figure;

(b) on the digital ball picture of every width of cloth, find the information of interest zone, detect and extract this region-of-interest (Huang Tongtong, the fast detecting of digital ball and identification.Computing machine institute of Zhejiang University: Computer Software and Theory, 2010), at last this region-of-interest is carried out coordinate axis transform, carry out feature extraction and represent with polar form, its flow process is as follows: detect the ellipse on the digital ball picture, then ellipse is moved on to the center, rotation and this ellipse of re-projection, ellipse is carried out binaryzation, be converted to polar coordinate image from binary image;

(c) at last will be from the n of i class _iWidth of cloth training picture constitutes matrix

A wherein _{I, j}, j=1,2 ... n _iBe the column vector by every width of cloth image construction, each element is the l of the unit of being standardized as all ²Norm, the training picture of all K class is combined into a training sample matrix A=[A ₁, A ₂..., A _K], be the training sample characteristic set.

Process to automatic location of all the digital balls in every width of cloth image and extraction subimage in the described step (2) is:

(a) the test pattern A to input carries out pre-service, uses the Canny operator to carry out rim detection, obtains binary image;

(b) in the image after the binaryzation have a few and all preserve, be the center with all non-zero points, the radius of ball is a radius, in this zone have a few and all add 1, this edge histogram of standardization again;

(c) all points in the traversal standardization edge histogram, also preserve at the center that estimates all circles;

(d) being the center with each center of circle information, is radius with the radius of ball, extracts all area-of-interests of test picture and preserves to scheme sheet mode, and output obtains the subimage A={A of all area-of-interests ₁, A ₂... A _n, i.e. the subimage of test pattern.

The enforcement algorithm that many balls detect is as follows:

Input: single width test pattern A

Output: all area-of-interest subimage A={A ₁, A ₂... A _n}

Step 1: the test pattern to input carries out pre-service, uses the Canny operator to carry out rim detection, obtains binary image;

Step 2: calculating the edge histogram of binary image, is the center with all non-zero points promptly, and the radius of ball is a radius, in this zone have a few and all add 1, this edge histogram of standardization again;

Step 3: all point in the traversing graph, get that (greater than certain preset threshold) peaked points are the center of circle and preserve in those certain zones;

Step 4: with each center of circle information is the center, is radius with the radius of ball, extracts all area-of-interests of test picture and preserves to scheme sheet mode.

The process that described subimage to corresponding same digital ball in identification single width or the multiple image extracts visual signature is: take out an area-of-interest subimage in the test picture, this region-of-interest is piled a column vector

Obtain the approximate sparse expression of y with all training sample characteristic sets, that is:

Wherein, β is sparse coefficient vector.Ideally, among the β except with y under the relevant coefficient of class i non-vanishing, other all coefficients all are zero.But normally most of nonzero value of β all concentrates on the i class, and the fraction nonzero value is dispersed on other class.Present problem is how to try to achieve β, thereby obtains the approximate expression of y.We can realize with following three kinds of methods:

First method is with lasso algorithm (R.Tibshirani, Regression Shrinkage and Selection via the Lasso.Journal of the Royal Statistical Society, Series B, 1996.) try to achieve sparse coefficient vector β, that is:

$\arg \underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1$

Wherein, || || ₁Be l ¹Norm, λ || β || ₁Be l ¹Penalty term.

Second method is with elastic network(s) algorithm (H.Zou and T.Hastie, Regularization and variable selection via the Elastic Net.Journal of the Royal Statistical Society, Series B, 2005.) try to achieve sparse coefficient vector β, that is:

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|\mathrm{\β}\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|\mathrm{\β}\left|\right|}_2^2$

Wherein, || || ₂Be expressed as the l of a vector ²Norm, and Elastic network(s) can use LARS (B.Efron, T.Hastie, I.Johnstone and R.Tibshirani, Least angle regression.Annals of Statistics, 2004.) method to solve.

The third method is to ask sparse coefficient vector β with non-losing side method (nonnegative garrote) (L.Breiman, Better subset regression using the nonnegative garrote.Technometrics, 1995.), that is:

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1,s.t.{\mathrm{\β}}_i>0$ For i=1 ..., n.

Non-losing side method can effectively be tried to achieve with the numerical method of classics, these classical numerical methods comprise that minimum angular convolution returns (LARS) and pathwise coordinate (J.Friedman, T.Hastie, H.Hofling and R.Tibshirani, Pathwise coordinate optimization.Annals OfApplied Statistics, 2007.) method.

After obtaining testing the sparse expression β of picture y, next be exactly to judge which class y belongs to.The feature of this extraction should be different as much as possible between different classes.According to the sparse expression of front, known that most of nonzero value of coefficient vector β mainly concentrates on the affiliated class of test sample book, promptly

Wherein, γ _iBe the coefficient vector relevant with the i class.In order to estimate the class under the test sample book, the posteriority below we have defined is estimated (posterior estimate):

$p(y\∈C_i|\mathrm{\β})=\frac{\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\mathrm{ij}}}{\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k}$

The γ here _IjThe coefficient vector that to be the i class calculating for the j time.

Then, we select the class of probable value maximum as the classification under the test picture y, that is:

$\underset{i}{\max }p(y\∈C_i|\mathrm{\β}).$

For single image, the recognition methods that we can adopt posteriority to estimate is discerned.For the identification of multiple image, in order to improve precision, we can take the method for associating posteriority judgement to realize.Calculate m ordered coefficients vector

Judge that y belongs to C _iThe repeatedly posterior probability of class, the computing formula of associating posteriority judgement is as follows:

$p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)=\frac{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\left(\mathrm{ij}\right)l}}{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{jl}}}$

γ _IjBe the i class at the coefficient vector that calculates for the j time, γ _{(ij) l}Be vectorial γ _IjL element, then, we select the class of probable value maximum as the classification under the test picture y, that is:

$\underset{i}{\max }p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)$

Following algorithm has provided the sparse expression and the judgment analysis process of image:

Input: the matrix that the training sample of k class constitutes

By testing the column vector that picture constitutes

Class label m under output: the y.

Step 1: find the solution problem with lasso:

$\arg \underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1;$

Or: (elastic net) finds the solution problem with elastic network(s):

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|\mathrm{\β}\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|\mathrm{\β}\left|\right|}_2^2;$

Or: (nonnegative garrote) finds the solution problem with non-negative method:

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1,s.t.{\mathrm{\β}}_i>0$ For i=1 ..., n

Step 2: the posteriority that calculates single is estimated:

$p(y\∈C_i|\mathrm{\β})=\frac{\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\mathrm{ij}}}{\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k}$

Or: posteriority is repeatedly estimated:

$p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)=\frac{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\left(\mathrm{ij}\right)l}}{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{jl}}}$

Step 3: the class label m under the output test sample y.

$\underset{i}{\max }p(y\∈C_i|\mathrm{\β})$ Or $\underset{i}{\max }p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)$

Description of drawings

Fig. 1 is the step of digital ball feature extraction among the embodiment;

Fig. 2 is the test picture of many ball identifications among the embodiment;

Fig. 3 is that synoptic diagram is discerned in the digital ball recognition methods based on rarefaction representation among the embodiment;

Fig. 4 is the identification process synoptic diagram that the present invention is based on the digital ball recognition methods of rarefaction representation.

Embodiment

As shown in Figure 4, a kind of digital ball recognition methods based on rarefaction representation and judgment analysis comprises:

(1) be placed on each the digital ball in the digital ball set under the color background separately, with single camera continuous acquisition single width or multiple image, as shown in Figure 1, has only a digital ball among every width of cloth figure, the set of formation training sample, ball in the positioning image also extracts visual signature, sets up the sparse expression of training sample set:

Each digital ball in the digital ball set is placed on separately under the black background, utilizes single camera continuous acquisition single width or multiple image to gather as training sample, and each image that obtains all is a single-view; The image of gathering is done pre-service, use the Canny operator to carry out rim detection, obtain binary image, then provide round roughly radius, the method for using Hough transformation or circumscribed circle to construct coupling is located the digital ball position among every width of cloth figure; On every width of cloth training picture, find information of interest zone (Huang Tongtong, the fast detecting of digital ball and identification.Computing machine institute of Zhejiang University: Computer Software and Theory, 2010), detect and extract this zone, at last this region-of-interest is carried out coordinate axis transform, carry out feature extraction and represent its flow process such as Fig. 1 with polar form: promptly detect the ellipse on the picture, then ellipse is moved on to the center, rotation and this ellipse of re-projection carry out binaryzation to ellipse, are converted to polar coordinate image from binary image; Then will be from the n of i class _iWidth of cloth training picture constitutes matrix

A wherein _{I, j}, j=1,2 ... n _iBe the column vector by every width of cloth image construction, each element is the l of the unit of being standardized as all ²Norm, therefore, the training picture of all K class is combined into a training sample matrix

(2) a plurality of digital ball that will test is placed under the identical color background, the collecting test image, and to the automatic location and extract subimage of all the digital balls in every width of cloth image: the one or more digital ball that detect is placed under as above the color background, gather single width or several test patterns, as Fig. 2; The test picture is carried out pre-service, use the Canny operator to carry out rim detection, obtain binary image; In the image after the binaryzation have a few and all preserve, calculate its edge histogram and standardization then; The institute of traversal in the histogram have a few, estimates all centers of justifying and leaves; Extracting the center of circle information preserves, is the center with each center of circle information, is radius with the radius of ball, extracts all area-of-interests of testing in the picture and preserves to scheme sheet mode.

The enforcement algorithm that many balls detect is as follows:

Input: single width test pattern A

Output: all area-of-interest subimage A={A ₁, A ₂... A _n}

Step 1: the test pattern to input carries out pre-service, uses the Canny operator to carry out rim detection, obtains binary image;

Step 2: calculating the edge histogram of binary image, is the center with all non-zero points promptly, and the radius of ball is a radius, in this zone have a few and all add 1, this edge histogram of standardization again;

Step 3: all point in the traversing graph, get that (greater than certain preset threshold) peaked points are the center of circle and preserve in those certain zones;

Step 4: with each center of circle information is the center, is radius with the radius of ball, extracts all area-of-interests of test picture and preserves to scheme sheet mode.

(3) process that the subimage of corresponding same digital ball in the single width of identification or the multiple image is extracted visual signature is: take out an area-of-interest subimage in the test picture, this region-of-interest is piled a column vector

Obtain the approximate sparse expression of y with all training sample characteristic sets, that is:

Wherein, β is sparse coefficient vector.Ideally, among the β except with y under the relevant coefficient of class i non-vanishing, other all coefficients all are zero.But normally most of nonzero value of β all concentrates on the i class, and the fraction nonzero value is dispersed on other class.Present problem is how to try to achieve β, thereby obtains the approximate expression of y.We can realize with following lasso algorithm:

Try to achieve sparse coefficient vector β with lasso (R.Tibshirani, Regression Shrinkage and Selection via the Lasso.Journal of the Royal Statistical Society, Series B, 1996.) algorithm, that is:

$\arg \underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1$

Wherein, || || ₁Be l ¹Norm, λ || β || ₁Be l ¹Penalty term.

(4) obtain testing the sparse expression β of picture y after, next be exactly to judge which class y belongs to.The feature of this extraction should be different as much as possible between different classes.According to the sparse expression of front, known that most of nonzero value of coefficient vector β mainly concentrates on the affiliated class of test sample book, promptly

Wherein, γ _iBe the coefficient vector relevant with the i class.In order to estimate the class under the test sample book, discern for multiple image, we take the method for associating posteriority judgement to realize.Calculate m ordered coefficients vector

Judge that y belongs to C _iThe repeatedly posterior probability of class, following calculating:

$p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)=\frac{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\left(\mathrm{ij}\right)l}}{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{jl}}}$

γ _IjBe the i class at the coefficient vector that calculates for the j time, γ _{(ij) l}Be l the element of vectorial γ ij, then, we select the class of probable value maximum as the classification under the test picture y, that is:

$\underset{i}{\max }p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)$

Following algorithm has provided the sparse expression and the judgment analysis process of image, as shown in Figure 3:

Input: the matrix that the training sample of k class constitutes

By testing the column vector that picture constitutes

Class label m under output: the y.

Step 1: find the solution problem with lasso:

$\arg \underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1;$

Step 2: the posteriority that calculates is repeatedly estimated:

$p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)=\frac{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\left(\mathrm{ij}\right)l}}{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{jl}}}$

Step 3: the class label m under the output test sample y.

$\underset{i}{\max }p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)$

Claims (9)

1. digital ball recognition methods based on rarefaction representation and judgment analysis comprises:

(1) be placed on each the digital ball in the digital ball set under the color background separately, utilize single camera continuous acquisition single width or multiple image, automatically locate the digital ball in every width of cloth image and extract visual signature, and all visual signatures are set up sparse expression, form the training sample characteristic set;

(2) one or more digital ball to be identified is placed in the same scene, gathers single width or multiple image, subimage is located and extracted to all the digital balls in every width of cloth image automatically; Subimage to corresponding same digital ball in single width or the multiple image extracts visual signature, and utilizes the training sample characteristic set to set up the sparse expression of this subimage;

(3) adopt the judgment analysis method to discern, obtain testing the affiliated classification of picture,, adopt the method for associating posteriority judgement to realize wherein for the situation of multiple image.

2. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 1 is characterized in that, the method that forms the training sample characteristic set in the described step (1) is:

(a) image of gathering is done pre-service, use the Canny operator to carry out rim detection, obtain binary image, then provide round roughly radius, the method for using Hough transformation or circumscribed circle to construct coupling is located the digital ball position among every width of cloth figure;

(b) on the digital ball picture of every width of cloth, find the information of interest zone, detect and extract this zone, at last this region-of-interest is carried out coordinate axis transform, carry out feature extraction and represent with polar form, its flow process is as follows: detect the ellipse on the digital ball picture, then ellipse is moved on to the center, rotation and this ellipse of re-projection, ellipse is carried out binaryzation, be converted to polar coordinate image from binary image;

(c) at last will be from the n of i class _iWidth of cloth training picture constitutes matrix

A wherein _{I, j}, j=1,2 ... n _iBe the column vector by every width of cloth image construction, each element is the l of the unit of being standardized as all ²Norm, the training picture of all K class is combined into a training sample matrix A=[A ₁, A ₂..., A _K], be the training sample characteristic set.

3. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 1 is characterized in that, the process to automatic location of all the digital balls in every width of cloth image and extraction subimage in the described step (2) is:

(a) the test pattern A to input carries out pre-service, uses the Canny operator to carry out rim detection, obtains binary image;

(b) in the image after the binaryzation have a few and all preserve, be the center with all non-zero points, the radius of ball is a radius, in this zone have a few and all add 1, this edge histogram of standardization again;

(c) all points in the traversal standardization edge histogram, also preserve at the center that estimates all circles;

(d) being the center with each center of circle information, is radius with the radius of ball, extracts all area-of-interests of test picture and preserves to scheme sheet mode, and output obtains the subimage A={A of all area-of-interests ₁, A ₂... A _n, i.e. the subimage of test pattern.

4. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 1, it is characterized in that, described to identification single width or multiple image in the subimage of the corresponding same digital ball process of extracting visual signature be: take out an area-of-interest subimage in the test picture, this region-of-interest piled a column vector

Obtain the approximate sparse expression of y with all training sample characteristic sets, that is:

Wherein, β is sparse coefficient vector.

5. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 4 is characterized in that described sparse coefficient vector β is found the solution by the lasso algorithm and obtains, that is:

$\arg \underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1$

Wherein, || || ₁Be l ¹Norm, λ || β || ₁Be l ¹Penalty term.

6. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 4 is characterized in that described sparse coefficient vector β is found the solution by elastic network(s) and obtains, that is:

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+{\mathrm{\λ}}_1{\left|\right|\mathrm{\β}\left|\right|}_1+{\mathrm{\λ}}_2{\left|\right|\mathrm{\β}\left|\right|}_2^2$

Wherein, || || ₂Be expressed as the l of a vector ²Norm, and

7. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 4 is characterized in that, described sparse coefficient vector β is found the solution by non-losing side method and obtains, that is:

$\underset{\mathrm{\β}}{\min }\frac{1}{2}{\left|\right|y-\mathrm{A\β}\left|\right|}_2^2+\mathrm{\λ}{\left|\right|\mathrm{\β}\left|\right|}_1,s.t.{\mathrm{\β}}_i>0$ For i=1 ..., n.

8. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 1 is characterized in that, in the described step (3), the recognition methods of adopting for single image is the single posteriority estimation technique, and its expression formula is:

$p(y\∈C_i|\mathrm{\β})=\frac{\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\mathrm{ij}}}{\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_k}$

Wherein, γ _IjThe coefficient vector that to be the i class calculating for the j time;

The class of the probable value maximum that calculates is as the classification under the test picture y, that is:

$\underset{i}{\max }p(y\∈C_i|\mathrm{\β}).$

9. the digital ball recognition methods based on rarefaction representation and judgment analysis according to claim 1 is characterized in that, in the described step (3), the recognition methods of adopting for multiple image is for repeatedly uniting the posteriority decision method, and its expression formula is:

$p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m)=\frac{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n_i}{\mathrm{\Σ}}}{\mathrm{\γ}}_{\left(\mathrm{ij}\right)l}}{\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{jl}}}$

Wherein, γ _IjBe the i class at the coefficient vector that calculates for the j time, γ _{(ij) l}Be vectorial γ _IjL element;

The class of the probable value maximum that calculates is as the classification under the test picture y, that is:

$\underset{i}{\max }p(y\∈C_i|{\mathrm{\β}}_1,{\mathrm{\β}}_2,...,{\mathrm{\β}}_m).$

Images