CN102024152B - Method for recognizing traffic sings based on sparse expression and dictionary study - Google Patents

Abstract

The invention discloses a method for recognizing traffic sings based on sparse expression and dictionary study, comprising the following steps: collecting pictures containing traffic signs and dividing the pictures into C image sample classes by hand, extracting image blocks of each image to form C training image block sets corresponding to the image sample classes, automatically studying a vision dictionary D by utilizing all the training image block sets, calculating the sparse expression of each class of training image block sets on the vision dictionary D, counting the average probability distribution of sparse expression coefficients along dictionary atoms as the feature expression for describing the C different sample classes, counting the probability distribution Pt of the sparse expression coefficients, for a test picture, along the dictionary atoms by using the same method, calculating a distance between the Pt and the average probability distribution, and selecting the sample class with the shortest distance as the recognition result of the traffic sign of the test image. In the invention, the classification of traffic sign pictures is realized by using the sparse expression and the probabilistic method, thereby achieving higher traffic sign recognition rate.

Description

A kind of method of carrying out Traffic Sign Recognition based on sparse expression and dictionary learning

Technical field

The invention belongs to the computer controlled automatic application, a kind of method of carrying out Traffic Sign Recognition based on sparse expression and dictionary learning of specific design.

Background technology

The Traffic Sign Recognition that occupies critical role in intelligent transportation system is more and more paid close attention in recent years.Key in the Traffic Sign Recognition is, how to collect the information of traffic sign, then it showed for classification.The mankind can identify it easily according to the CF of traffic sign.Yet, with the identification of computer realization traffic sign, a lot of challenges are arranged still.At first, because illumination, daytime, the color of traffic sign was different from evening.Secondly, the shape of traffic sign may not be the geometric configuration of standard, such as circle, and triangle or hexagon.At last, traffic sign can such as shade, hide owing to weather, mist etc., and reduce its observability.

Nearest researcher has proposed several methods for Traffic Sign Recognition.They were divided into for two steps to this task of Traffic Sign Recognition: detection-phase and sorting phase.At detection-phase, detect the zone that may have traffic sign with color or shape facility.At sorting phase, the traffic sign that sorter obtains detection-phase is assigned to the class of appointment.Yet, in this traditional method, also there is a subject matter, be exactly before classification, must locate the center of a traffic sign, because the recognition result that the location of mistake meeting of traffic sign candidate region leads to errors.Sparse expression and probabilistic method can well address this problem.

Sparse model is used widely in signal, image and video task, and it is modeled as input signal the linear combination of minority atom in the dictionary.In the research of computer vision, we are to how presentation video is interested.In this model, sparse quilt preferentially is used for the expression of signal.Key challenge in this model is how to choose base or dictionary that image information represents.Nearest studies show that the dictionary that obtains by study is than using predefined dictionary can obtain better effect.

Summary of the invention

The invention provides and a kind ofly carry out the method for Traffic Sign Recognition based on sparse expression and dictionary learning, solved higher-dimension and calculated the difficulty of bringing, the dictionary that upgrades with unceasing study improves accuracy of identification.

A kind ofly carry out the method for Traffic Sign Recognition based on sparse expression and dictionary learning, comprising:

(1) collects the natural scene picture that comprises traffic sign, and be divided into C image pattern class by the classification of traffic sign wherein is manual, then for the every width of cloth image in each image pattern class, extract the image block of some formed objects, form C training image set of blocks of correspondence image sample class;

(2) utilize all C training image set of blocks, automatic learning goes out a vision dictionary D;

(3) training image blocks that calculates each class is integrated into the sparse expression on the vision dictionary D, and statistics sparse expression coefficient is along the Average probability distribution { P of dictionary atom ₁, P ₂..., P _C, as the feature representation of describing C different sample classes;

(4) comprise the test picture of unknown traffic sign for a width of cloth, extract the some image blocks in this picture, consist of the test pattern set of blocks, then calculate the sparse expression of test pattern set of blocks on vision dictionary D, statistics sparse expression coefficient is along the probability distribution P of dictionary atom _t

(5) calculate P _tWith { P ₁, P ₂..., P _CSimilarity, select the most similar sample class as the Traffic Sign Recognition result of test pattern.

Concrete steps in the above-mentioned steps (1) are: collect the natural scene picture comprise traffic sign, and by manual C the image pattern class (then be some STOP traffic indication maps of by the natural conditions such as various shapes, illumination, background under obtaining such as STOP traffic sign class) that be divided into of the classification of traffic sign wherein; Then the every width of cloth image in each image pattern class is processed, centered by each pixel of every width of cloth image, extracting size is the image block of n * n (n is less than or equal to the image size), by row all pixels in every width of cloth image are all carried out the extraction of image block again by row first, thereby each image pattern class has consisted of a training image set of blocks

Wherein

The image block of j class training image, n _jBe such image block number altogether, and the image block here is to allow to cover; At last, C image pattern class all carried out the extraction of image block, form total image block set

Then utilize all training image set of blocks, automatic learning goes out a vision dictionary D, and concrete steps are:

(i) dictionary D is initialized as some image blocks of selecting at random from C training image set of blocks;

(ii) in the sparse coding stage, for dictionary D, by finding the solution Find the sparse expression α of each image block x; Wherein,

Be a sparse coefficient vector, λ is a Regularization coefficient, and P is l ⁰Or l ¹Norm, be one in order to reach the regular of sparse purpose;

(iii) in the dictionary updating stage, with K-SVD algorithm (Aharon, M.and Elad, M.andBruckstein, A, K-SVD:An Algorithm for Designing OvercompleteDictionaries for Sparse Representation, Signal Processing, IEEE Transactionson[see also Acoustics, Speech, and Signal Processing, IEEE Transactions on], 2006) upgrade dictionary

In each element d _l, and l=1,2 ..., k;

(vi) last sparse coding and the dictionary updating of constantly repeating, until convergence or iterations cut-off, obtained a vision dictionary D obtaining from the study of all training image blocks, this dictionary had not only comprised unique visual information of each class but also had comprised the common visual information of all classes.

The below has provided the dictionary updating algorithm based on K-SVD:

Input: the set of C training picture block

Output: dictionary

Arthmetic statement:

Step 1: dictionary D is initialized as some image blocks of from C training image set of blocks, selecting at random;

Step 2: repeating step 3～4, until convergence or iterations stop;

Step 3: in the sparse coding stage, to dictionary D, by finding the solution Find the sparse factor alpha of each image block x;

Step 4: in the dictionary updating stage, for l=1 ..., k constantly updates atom d _l, wherein

Step 4.1: for current d _l, select α ^lThe non-vanishing numbering of coefficient forms set ω _l:

ω _l＝{i∈1，...，N|α ^l[i]≠0}

Here α ^lIt is matrix of coefficients

L capable;

Step 4.2: error of calculation matrix E:

$E=X-\underset{p\≠l}{\mathrm{\Σ}}{d}_p{\mathrm{\α}}^p;$

Step 4.3: take out among the error matrix E corresponding to set ω _lRow, thereby obtain E _l

Step 4.4: by solving following optimization problem, upgrade d _lNonzero coefficient α with correspondence ^l:

$\underset{d_l,{\mathrm{\α}}^l}{\min }{\left|\right|(E_l-d_l{\mathrm{\α}}^l)\left|\right|}_F^2;$

Step 4.5: finish;

Step 5: finish.

Described step (3) is calculated the sparse expression coefficient of i class training image set of blocks along the Average probability distribution P of dictionary atom _iMethod be:

(i) at first calculate the sparse expression of each image block on dictionary D in the training image set of blocks of i class, use

Find the solution and obtain sparse coefficient;

(ii) obtain the sparse expression coefficient of such training image set of blocks along the Average probability distribution P of dictionary atom by the method for averaging _i, method is as follows:

P _i=p _Tr(S _i| D)=[p (S _i| d ₁) ..., p (S _i| d _k)], and

Wherein, for given i class training image set of blocks

Here n _iThe sum of i class image block, set w _lAtom d is used in expression _lThe label of piece, therefore set { w _lIn the number of element can be expressed as atom d _lAccess times, be defined as N _l, S _iThe distribution of dictionary D on the i class.

The sparse expression coefficient of described step (4) calculating testing image set of blocks along the method for the Average probability distribution Pt of dictionary atom is:

(i) centered by each pixel of test pattern, extracting size is the image block of n * n, by row all pixels in the image is all carried out the extraction of image block again by row first, thereby has consisted of the set of test pattern piece

(ii) use

Calculate the sparse expression of test pattern set of blocks on vision dictionary D;

(iii) try to achieve the sparse expression coefficient along the probability distribution P of dictionary atom with qualitative modeling at last _t, method is as follows:

P _t=p _Te(I _l| D)=[p (I _l| d ₁) ..., p (I _l| d _k)], and

Wherein, for test picture I _lDictionary distribute the image block set of test pattern

Here M is the sum of image block, w _lBe expressed as and use atom d _lThe index of piece, therefore set { w _lIn the number of element can be expressed as atom d _lAccess times, be defined as N _l

After obtaining the dictionary distribution separately of every class training image blocks and test pattern piece, next be exactly to look for test picture I _lBelong to which class, we can obtain to test classification under the picture by minimizing following optimization problem:

$\underset{i}{\min }{\left|\right|p_{\mathrm{te}}\left(I_t\right|D)-p_{\mathrm{tr}}\left(S_i\right|D)\left|\right|}_2^2$

Wherein, p _Te(I|D) for testing the probability of picture block, p _Tr(S _i| D) for training the probability of picture block.

Beneficial effect of the present invention:

(1) the present invention represents signal with sparse model, so that the test picture solves higher-dimension and calculates the difficulty of bringing only with a few pictures linear dependence in the training picture;

(2) the present invention uses the dictionary that unceasing study upgrades, than obtaining better effect with predefined dictionary;

(3) the present invention's probability of picture fritter, rather than the probability of picture in its entirety is classified, and has reached comparatively ideal effect;

(4) the present invention replaces each class to have separately a dictionary with a large vision dictionary, and this big dictionary has comprised again the common information of all classes namely by each class information separately, so that identification is more accurate.

Description of drawings

Fig. 1 is some the STOP traffic indication maps under the conditions such as the difformity selected among the embodiment, illumination, background;

Fig. 2 serve as reasons sparse expression and the dictionary learning process flow diagram of the training picture that all traffic indication maps form;

Fig. 3 is take dictionary learning and the cognitive phase process flow diagram of STOP traffic sign as the test picture;

Fig. 4 is total schematic flow sheet of recognition methods among the present invention.

Embodiment

As shown in Figure 4, a kind ofly carry out the method for Traffic Sign Recognition based on sparse expression and dictionary learning, detailed process is as follows:

Collect the natural scene picture of traffic sign, and be divided into C image pattern class by the classification of traffic sign wherein is manual, such as some STOP traffic indication maps that contain under the conditions such as difformity, illumination, background, as shown in Figure 1; Then for the every width of cloth image in each image pattern class, centered by each pixel of every width of cloth image, extracting size is the image block of n * n (n is less than or equal to the image size), by row all pixels in every width of cloth image are all carried out the extraction of image block again by row first, thereby each image pattern class has consisted of a training image set of blocks As shown in Figure 2, wherein

The image block of j class training image, n _jBe such image block number altogether, and the image block here is to allow to cover; At last, C image pattern class all carried out the extraction of image block, form total image block set

As shown in Figure 2.

Then utilize and obtain all training image set of blocks Automatic learning goes out a vision dictionary D, and as shown in Figure 2, concrete steps are:

(i) dictionary D is initialized as some image blocks of selecting at random from C training image set of blocks;

(ii) in the sparse coding stage, for dictionary D, by finding the solution Find the sparse expression α of each image block x; Wherein, α ∈ R ^kBe a sparse coefficient vector, λ is a Regularization coefficient, and P is l ⁰Or l ¹Norm, be one in order to reach the regular of sparse purpose;

(iii) in the dictionary updating stage, with K-SVD algorithm (Aharon, M.and Elad, M.andBruckstein, A, K-SVD:An Algorithm for Designing OvercompleteDictionaries for Sparse Representation, Signal Processing, IEEE Transactionson[see also Acoustics, Speech, and Signal Processing, IEEE Transactions on], 2006) upgrade dictionary

In each element d _l, and l=1,2 ..., k;

(vi) last sparse coding and the dictionary updating of constantly repeating, until convergence or iterations cut-off, obtained a vision dictionary D obtaining from the study of all training image blocks, this dictionary had not only comprised unique visual information of each class but also had comprised the common visual information of all classes.

Dictionary updating algorithm based on K-SVD is:

Input: the set of C training picture block

Output: dictionary

Arthmetic statement:

Step 1: dictionary D is initialized as some image blocks of from C training image set of blocks, selecting at random;

Step 2: repeating step 3～4, until convergence or iterations stop;

Step 3: in the sparse coding stage, to dictionary D, by finding the solution

Find the sparse factor alpha of each image block x;

Step 4: in the dictionary updating stage, for l=1 ..., k constantly updates atom d _l, wherein

Step 4.1: for current d _l, select α ^lThe non-vanishing numbering of coefficient forms set ω _l:

ω _l＝{i∈1，...，N|α ^l[i]≠0}

Here α ^lIt is matrix of coefficients

L capable;

Step 4.2: error of calculation matrix E:

$E=X-\underset{p\≠l}{\mathrm{\Σ}}{d}_p{\mathrm{\α}}^p;$

Step 4.3: take out among the error matrix E corresponding to set ω _lRow, thereby obtain E _l

Step 4.4: by solving following optimization problem, upgrade d _lNonzero coefficient α with correspondence ^l:

$\underset{d_l,{\mathrm{\α}}^l}{\min }{\left|\right|(E_l-d_l{\mathrm{\α}}^l)\left|\right|}_F^2;$

Step 4.5: finish;

Step 5: finish.

The training image blocks that calculates each class is integrated into the sparse expression on the vision dictionary D, and statistics sparse expression coefficient is along the Average probability distribution { P of dictionary atom ₁, P ₂..., P _C, as the feature representation of describing C different sample classes: (i) at first calculate the sparse expression of each image block on dictionary D in the training image set of blocks of i class, use

Find the solution and obtain sparse coefficient;

(ii) obtain the sparse expression coefficient of such training image set of blocks along the Average probability distribution P of dictionary atom by the method for averaging _i, method is as follows:

P _i=p _Tr(S _i| D)=[p (S _i| d ₁) ..., p (S _i| d _k)], and

Wherein, for given i class training image set of blocks

Here n _iThe sum of i class image block, set w _lAtom d is used in expression _lThe label of piece, therefore set { w _lIn the number of element can be expressed as atom d _lAccess times, be defined as N _l, S _iThe distribution of dictionary D on the i class.

As shown in Figure 3, comprise the test picture of unknown traffic sign for a width of cloth, extract the some image blocks in this picture, consist of the test pattern set of blocks, then calculate the sparse expression of test pattern set of blocks on vision dictionary D, statistics sparse expression coefficient is along the probability distribution P of dictionary atom _t: (i) centered by each pixel of test pattern, extracting size is the image block of n * n, by row all pixels in the image is all carried out the extraction of image block again by row first, thereby has consisted of the set of test pattern piece

(ii) use

Calculate the sparse expression of test pattern set of blocks on vision dictionary D;

(iii) try to achieve the sparse expression coefficient along the probability distribution P of dictionary atom with qualitative modeling at last _t, method is as follows:

P _t=p _Te(I _l| D)=[p (I _l| d ₁) ..., p (I _l| d _k)], and

Wherein, for test picture I _lDictionary distribute the image block set of test pattern Here M is the sum of image block, w _lBe expressed as and use atom d _lThe index of piece, therefore set { w _lIn the number of element can be expressed as atom d _lAccess times, be defined as N _l

As shown in Figure 3, calculate P _tWith { P ₁, P ₂..., P _CSimilarity, select the most similar sample class as the Traffic Sign Recognition result of test pattern, detailed process is: after obtaining every class training image blocks and test pattern piece dictionary separately and distributing, next be exactly to look for test picture I _lBelong to which class, we can obtain to test classification under the picture by minimizing following optimization problem:

$\underset{i}{\min }{\left|\right|p_{\mathrm{te}}\left(I_t\right|D)-p_{\mathrm{tr}}\left(S_i\right|D)\left|\right|}_2^2$

Wherein, p _Te(I|D) for testing the probability of picture block, p _Tr(S _i| D) for training the probability of picture block.

Claims (2)

1. one kind is carried out the method for Traffic Sign Recognition based on sparse expression and dictionary learning, comprising:

(1) collects the natural scene image that comprises traffic sign, and be divided into C image pattern class by the classification of traffic sign wherein is manual, then for the every width of cloth image in each image pattern class, centered by each pixel of every width of cloth image, extracting size is the image block of n * n, by row all pixels in every width of cloth image are all carried out the extraction of image block again by row first, form C training image set of blocks of correspondence image sample class; Wherein, n is less than or equal to the image size;

(2) utilize all C training image set of blocks, automatic learning goes out a vision dictionary D;

(3) training image blocks that calculates each class is integrated into the sparse expression on the vision dictionary D, and statistics sparse expression coefficient is along the Average probability distribution { P of dictionary atom ₁, P ₂..., P _C, as the feature representation of describing C image pattern class;

(4) comprise the test pattern of unknown traffic sign for a width of cloth, centered by each pixel of test pattern, extracting size is the image block of n * n, by row all pixels in the image is all carried out the extraction of image block again by row first, thereby has consisted of the set of test pattern piece

Wherein, n is less than or equal to the image size, and M is the sum of image block; Then calculate the sparse expression of test pattern set of blocks on vision dictionary D, statistics sparse expression coefficient is along the probability distribution P of dictionary atom _t

(5) calculate P _tWith { P ₁, P ₂..., P _CSimilarity, select the most similar image pattern class as the Traffic Sign Recognition result of test pattern;

The method that described step (2) automatic learning goes out a vision dictionary D is:

(i) dictionary D is initialized as some image blocks of selecting at random from C training image set of blocks;

(ii) in the sparse coding stage, for dictionary D, by finding the solution

Find the sparse expression α of each image block x; Wherein,

Be a sparse coefficient vector, λ is a Regularization coefficient, and P is l ⁰Or l ¹Norm, be one in order to reach the regular of sparse purpose;

(iii) in the dictionary updating stage, upgrade dictionary with the K-SVD algorithm

In each atom d _l, and l=1,2 ..., k;

(vi) last constantly repeating step (ii) and step (iii), until convergence or iterations cut-off have obtained a vision dictionary D who obtains from all training image blocks study;

Described step (3) is calculated the sparse expression of i class training image set of blocks, and statistics sparse expression coefficient is along the Average probability distribution P of dictionary atom _iMethod be:

(i) at first calculate the sparse expression of each image block on dictionary D in the training image set of blocks of i class, use Find the solution and obtain the sparse expression coefficient;

(ii) obtain the sparse expression coefficient of such training image set of blocks along the Average probability distribution P of dictionary atom by averaging method _iDescribed mean value method is:

P _i=p _Tr(S _i| D)=[p (S _i| d ₁) ..., p (S _i| d _k)], and

Wherein, for given i class training image set of blocks

Here n _iIt is the sum of i class image block; Set { w _lExpression use atom d _lThe label of piece,

Therefore gather { w _lIn the number of element can be expressed as atom d _lAccess times, be defined as N _l, S _iThe distribution of dictionary D on the i class;

Described step (4) is calculated the sparse expression of test pattern set of blocks, and statistics sparse expression coefficient is along the probability distribution P of dictionary atom _tMethod be:

(i) use

Calculate the sparse expression of test pattern set of blocks on vision dictionary D;

(ii) at last try to achieve the sparse expression coefficient along the probability distribution P of dictionary atom with mean value method _tBe specially:

P _t=p _Te(I _l| D)=[p (I _l| d ₁) ..., p (I _l| d _k)], and

Wherein, for test pattern I _lProbability distribution, the set of test pattern piece Here M is the sum of test pattern piece.

2. according to claim 1ly carry out the method for Traffic Sign Recognition based on sparse expression and dictionary learning, it is characterized in that P in the described step (5) _tWith { P ₁, P ₂..., P _CThe computing method of similarity be:

$\underset{i}{\min }{\left|\right|p_{\mathrm{te}}\left(I_l\right|D)-p_{\mathrm{tr}}\left(S_i\right|D)\left|\right|}_2^2$

Wherein, p _Te(I _l| D) be the probability distribution of test pattern piece, p _Tr(S _i| D) be the probability distribution of training image blocks.

Images