CN103544500B - Multi-user natural scene mark sequencing method - Google Patents

Abstract

The invention provides a multi-user natural scene mark sequencing method. The multi-user natural scene mark sequencing method includes the steps of firstly, extracting feature vectors of natural scene image sets used for conducting training; secondly, obtaining multiple mark sequences of natural scene images from an image marking system on the basis of user interestingness degrees; thirdly, converting the multiple mark sequences into mark distribution; fourthly, obtaining the natural scene images to be marked and sequenced from an input device, and extracting the feature vectors; fifthly, judging whether training is well conducted or not; sixthly, training the optimum parameter vector theta of a natural scene mark distribution model; seventhly, substituting the optimum parameter vector theta and the feature vectors of the natural scene images to be marked and sequenced into the natural scene mark distribution model, and obtaining the mark distribution of the natural scene images to be marked and sequenced; eighthly, treating marks with the description degrees smaller than the description degree of a virtual mark as irrelevant marks, treating the rest of marks as relevant marks, and finally sequencing the relevant marks according to the magnitudes of the description degrees.

Description

Multi-user natural scene tag sort method

Technical field

The present invention relates to the use of the method to multi-user natural scene tag sort for the computer, belong to image processing techniques neck Domain.

Background technology

Currently, the development of Internet information technique and the popularization of digital equipment bring the explosive increase of view data, The appearance of a large amount of digital pictures is brought more convenient with propagation, also enriches the life of people, but the profit to image for the people With not improving with the increase of picture number with selective power, this just give user using bringing new challenge.Therefore, How automatically, quickly and accurately by image to become current according to the automatic classification of wish of people and sequence using computer One urgent task.

One width natural scene image often can be marked with many conception of species marks, and user can be according to oneself to image Understand and relevance ranking is carried out to these marks.At present, the tag sort problem of natural scene image is primarily present two big defects； First, adopt single user mark method, i.e. piece image only one of which annotation results, but due to user itself subjectivity because Element, the annotation results of unique user may be not accurate enough.2nd, to be solved using the method extending specific Multi-label learning algorithm Tag sort problem, does not make full use of data already present tag sort information itself.

Content of the invention

The invention provides a kind of multi-user natural scene tag sort method, solve the method generation of multi-user's mark Inconsistency；On making full use of training data existing tag sort Information base, natural scene image is realized automatically The tag sort changed.

The technical scheme is that a kind of multi-user natural scene tag sort method, comprise the steps:

(1) obtain the natural scene image set for training, and characteristic vector is extracted to every width natural scene image；

(2) multiple tag sorts of natural scene image are obtained from the image labeling system based on user interest degree；

(3) multiple tag sorts are converted into an indicia distribution；

(4) obtain the natural scene image of sequence to be marked from input equipment, extract characteristic vector；

(5) if recognition mechanism does not train, execution step (6), otherwise go to step (7)；

(6) the natural scene image characteristic vector being obtained with step (1) to step (3) and its tag sort information are as instruction Practice collection, using the Algorithm for Training of lbfgs-lld, obtain the optimal parameter vector θ of natural scene indicia distribution model；

(7) natural scene of the sequence to be marked that the optimal parameter obtaining in step (6) vector θ and step (4) are obtained The characteristic vector of image substitutes in natural scene indicia distribution model respectively, obtains the mark of this sequence natural scene image to be marked Score cloth.

(8) described in the indicia distribution that step (7) is obtained, degree regards unrelated mark as less than the mark of virtual tag degree of description Note, and remaining mark is mark of correlation, finally mark of correlation is ranked up according to description degree size.

Wherein, the method for step (1) and step (4) extraction characteristic vector is: coloured image is converted into cieluv colored Space, the method then carrying out piecemeal again extracts the characteristics of image of natural scene.

Step (3) is passed through to build a nonlinear optimal problem, and is solved using interior point method, by multiple tag sorts It is converted into an indicia distribution.

The concrete construction method of described nonlinear optimal problem is: for image m, the tag sort of n position user is respectively Use l₁,l₂...l_nRepresent it is assumed that each user is respectively according to indicia distribution p₁,p₂...p_nMark is ranked up, and comprehensive The indicia distribution of all consumers' opinions represents with q, model with minimum k l distance as object function, that is, makes q and p₁,p₂...p_n Between difference minimum, also should meet some constraint formulas, that is, for distribution p simultaneously₁,p₂...p_nFor, in tag sort, row Should be more than the description degree of the mark coming below in the description degree of mark above；For an indicia distribution, its institute The description degree sum of possible mark should be 1.

The basis for estimation whether distribution of step (5) judge mark trains is that the indicia distribution of model prediction is marked with real Whether the average kl distance scored between cloth is sufficiently small.

Related definition: (a) example: a width natural scene image.(b) indicia distribution: given example x, all possible mark Note can completely describe x, and the degree of each mark description x is represented with a real number, referred to as degree of description.The markd description of institute One distribution of degree composition is just referred to as indicia distribution.(c) description degree: in the indicia distribution of example x, mark y corresponds to a reality Number p_xY () ∈ [0,1], represents that mark y describes the degree of x, a markd description degree sum of the corresponding institute of example is 1.(d) Image labeling system based on user interest degree: the mark of correlation row that user is every width natural scene image according to the interest of oneself Sequence, then system automatically the result of tag sort is stored in database.(e) kl distance: also known as kullback-leibler divergence, It is a kind of index of two distribution similarities of tolerance.

Beneficial effect:

1. the method that the present invention is marked using multi-user is so that annotation results more tend to objective, and passes through to introduce one Nonlinear optimization, successfully solves the problem of inconsistency that multi-user's mark produces, and has obtained one and has combined all mark use The indicia distribution of family suggestion, the method for expressing of this indicia distribution has included relatively objective complete mark correlation, can be more preferably Expression all mark users suggestions.

2. invention introduces the concept of virtual tag, as dividing one natural " zero of mark of correlation and extraneous markings Point ", allows algorithm automatically select suitable cut-point, than artificial given threshold more science, accurately.

3. the tag sort that the present invention is given Lai comprehensive multiple users using indicia distribution, the lbfgs-lld algorithm of employing Take full advantage of the existing sequencing information of data itself, thus its contained mark of image prediction that not only can not mark for a width These marks can also be ranked up the increment it is achieved that information value by note according to correlation.

4. full automation of the present invention, and speed is fast, and the degree of accuracy is high, can conveniently be applied to image retrieval Most occasions.

Brief description

Fig. 1 is the workflow diagram of multi-user natural scene tag sort method；

Fig. 2 is indicia distribution example；

Fig. 3 is natural scene image example.

Specific embodiment

The present invention is described in detail with instantiation below in conjunction with the accompanying drawings.

As shown in figure 1, multi-user natural scene tag sort method of the present invention comprises the steps:

(1) obtain the natural scene image set for training, and characteristic vector is extracted to every width natural scene image；

(2) multiple tag sorts of natural scene image are obtained from the image labeling system based on user interest degree；Base The mark of correlation row be user being every width natural scene image according to the interest of oneself in the image labeling system of user interest degree Sequence, then system automatically the result of tag sort is stored in database.

(3) pass through to build a nonlinear optimal problem, and solved using interior point method, multiple tag sorts are changed Become an indicia distribution.The concrete construction method of described nonlinear optimal problem is: for image m, the mark row of n position user Sequence uses l respectively₁,l₂...l_nRepresent it is assumed that each user is respectively according to indicia distribution p₁,p₂...p_nMark is ranked up, And the indicia distribution combining all consumers' opinions is represented with q, model with minimum k l distance as object function, that is, makes q and p₁, p₂...p_nBetween difference minimum, also should meet some constraint formulas, that is, for distribution p simultaneously₁,p₂...p_nFor, in mark row In sequence, the description degree coming mark above should be more than the description degree of the mark coming below；One indicia distribution is come Say, the description degree sum of its all possible mark should be 1.This optimization problem is described as follows:

Denotational description:

Label space λ: λ={ λ₀,λ₁,...,λ_m, wherein λ₀It is virtual tag, as mark of correlation and extraneous markings A cut-point, that is, coming before virtual tag is mark of correlation, on the contrary be extraneous markings

l_i: i-th user tag sort to image x, i=1 ..., n

p_i: l_iCorresponding indicia distribution, i=1 ..., n

Q: combine n position consumers' opinions l₁,l₂...l_nIndicia distribution

q_j: mark λ_jDescription degree in indicia distribution q, j=0 ..., m

p_i,j: mark λ_jIn indicia distribution p_iIn description degree, i=1 ..., n, j=0 ..., m

l_i,j: mark λ_jIn tag sort l_iIn position, i=1 ..., n, j=0 ..., m

The number of v: virtual tag, a sub-picture can have multiple virtual tag

D: less than constraints conversion becomes a threshold value less than or equal to constraint

Input: l₁,l₂...l_n, virtual tag number v, threshold value d

Output: p₁,p₂...p_n, q

Target: $\underset{p_i,q}{\arg \min }\underset{i=1}{\overset{n}{\mathrm{\σ}}}(\underset{j=1}{\overset{m}{\mathrm{\σ}}}{q}_j\log \frac{q_j}{p_{i,j}}+q_0\log \frac{q_0}{p_{i,0}}*v)$

s.t.In // distribution q, the markd description degree sum of institute is 1

// distribution p_iIn, the description degree sum of each mark is 1

3.p_i,k+d≤p_i,jwhenl_i,j＜ l_i,k, j, k=0 ..., m, i=1 ..., n

//l_i,j＜ l_i,kRepresent in tag sort l_iIn, mark λ_jIn mark λ_kAbove

4.q_j≥0,p_i,j>=0, i=1 ..., n, j=1 ..., the description degree of m//mark is not negative

5.p_i,j=0ifl_i,j＞ l_i,0, i=1 ..., n, j=1 ..., m

// i.e. as mark λ_jIn l_iIn when being extraneous markings, then in p_iIn, its degree of description is 0

(4) obtain the natural scene image of sequence to be marked from input equipment, extract characteristic vector；

(5) if recognition mechanism does not train, execution step (6), otherwise go to step (7)；Whether judge mark distribution The basis for estimation training is whether the average kl distance between the indicia distribution of model prediction and real indicia distribution is enough Little.

(6) the natural scene image characteristic vector being obtained with step (1) to step (3) and its tag sort information are as instruction Practice collection, using the Algorithm for Training of lbfgs-lld, obtain the optimal parameter vector θ of natural scene indicia distribution model；lbfgs- Lld algorithm is specific as follows:

The first step: input training set s={ (x₁,p₁(y)),(x₂,p₂(y)),…,(x_n,p_n(y)) }, wherein x_iRepresent one Example, p_iY () is x_iCorresponding distribution.

Second step: assume that the characteristic vector of image and its corresponding indicia distribution meet maximum entropy model, that is, suppose mark Distribution is a distribution of the condition with parameter:Wherein g_kX () is the feature letter of example x Number, z=σ_yexp(σ_kθ_y,kg_k(x)) it is standardizing factor, θ is the parameter vector of model.

3rd step: solution θ makes average between the indicia distribution of model prediction and real indicia distribution in second step Kl distance is minimum, that is, ${\mathrm{\θ}}^{*}=\underset{\mathrm{\θ}}{\arg \min }\underset{i}{\mathrm{\σ}}\underset{y}{\mathrm{\σ}}{p}_i\left(y\right)\log (p_i\left(y\right)/p\left(y\right|x_i,\mathrm{\θ})).$ Using lbfgs algorithm, this optimization is asked Topic is solved, and finds an optimal parameter vector θ^*.Lbfgs algorithm specifically refers to j.nocedal.updating quasi-newton matrices with limited storage (1980),mathematics of Computation35, pp.773-782 and d.c.liu and j.nocedal.on the limited memory method for large scale optimization(1989),mathematical programming b,45,3,pp.503- 528.

(7) natural scene of the sequence to be marked that the optimal parameter obtaining in step (6) vector θ and step (4) are obtained The characteristic vector of image substitutes in natural scene indicia distribution model respectively, obtains the mark of this sequence natural scene image to be marked Score cloth.

(8) described in the indicia distribution that step (7) is obtained, degree regards unrelated mark as less than the mark of virtual tag degree of description Note, and remaining mark is mark of correlation, finally mark of correlation is ranked up according to description degree size.

Wherein, the method for step (1) and step (4) extraction characteristic vector is: coloured image is converted into cieluv colored Space, the method then carrying out piecemeal again extracts the characteristics of image of natural scene.

It is specifically described the present invention below in conjunction with natural scene image example.

System first passes through the gray level image that Digital Image Input Device obtains natural scene, such as Fig. 3.Subsequently enter Computer processing procedure, this process includes extracting the characteristic vector of input natural scene image, in the present invention, first colour Image is converted into cieluv color space, and the method then carrying out piecemeal again extracts natural scene image feature.The method by Matthew r.boutell and jiebo luo is in 2004 in learning multi-label scene Propose in classification, pattern recognition37 (9) (2004) 1757-1771, the present invention shares 294 Individual feature.

Obtain multiple tag sorts of Fig. 3 from the image labeling system based on user interest degree, be shown in Table 1.

The tag sort to Fig. 3 for the table 1:10 position user

plant	sky	cloud	snow	building	desert	mountain	water	sun	virtual
										3	10	10	2	4	10	1	10	10	9
4	10	10	2	3	10	1	10	10	9
										4	10	10	2	3	10	1	10	10	9
3	10	10	2	4	10	1	10	10	9
										10	2	10	10	10	10	1	10	10	9
4	10	10	2	3	10	1	10	10	9
										3	10	10	2	4	10	1	10	10	9
2	3	10	10	10	10	1	10	10	9
										2	10	10	10	3	10	1	10	10	9
4	10	10	2	3	10	1	10	10	9

As shown in figure 3, this figure has mark mountain, it is sequential between snow, plant etc., and these marks, its In 1 expression interest-degree highest, 2 expression interest-degrees take second place, 10 represent unrelated, 9 expression virtual tag, interest-degree passing with sequence number Increase and successively decrease.Can be clearly seen that 10 users are not quite similar to the annotation results of Fig. 3 from table 1, this exactly multi-user marks Problem of inconsistency.

Build nonlinear optimal problem, then solved using interior point method, multiple tag sorts of Fig. 3 are converted into one Individual indicia distribution q, this indicia distribution combines the suggestion of all mark users, and so we have just obtained the indicia distribution of Fig. 3. In conjunction with table 1, table 2 can draw, distribution q is shown the positional information being marked in each tag sort in the form of degree of description Come, unified the suggestion of each user well.Also introduce the concept of virtual tag in the middle of this, be inserted in mark of correlation nothing to do with Between mark, as a natural zero point dividing mark of correlation nothing to do with mark.

Corresponding indicia distribution q of table 2: Fig. 3

plant	sky	cloud	snow	building	desert	mountain	water	sun	virtual
										0.079	0.0597	0	0.0826	0.0758	0	0.1054	0	0	0.0664

Indicia distribution is as shown in Figure 2.In fig. 2, abscissa represents the mark y that sample may contain_i, wherein i= 1 ..., 5, ordinate p (y) represents the degree of each mark description sample, and σ_ip(y_i)=1.

In natural scene data set, virtual tag is to divide the cut-point of mark of correlation nothing to do with mark, wherein, in void Intending before mark is mark of correlation, and is to be that sequential (centre can between extraneous markings, and mark of correlation after virtual tag There to be draw, constitute a bucket, the mark in bucket is parallel), may be considered out-of-order between extraneous markings, form one Individual group.For example, a secondary natural scene image can so mark (p_x,{λ_0,1...λ_0,v},n_x), p_xIt is mark of correlation subset, {λ_0,1...λ_0,vIt is virtual tag subset, n_xIt is extraneous markings subset, if it is considered that implicit draw relation in tag set, (b can so be represented₁....b_i-1,b_i, c), wherein p_x=(b₁...b_i-1), b_i={λ_0,1...λ_0,v, c=n_x, wherein v is virtual Mark number, b is bucket(bucket) the meaning, bucket internal labeling is out-of-order, i.e. draw relation, but is complete ordering between bucket 's.C is collective(group) the meaning, because being all extraneous markings, we are indifferent to their order to the mark in c, but b But be complete ordering and c between, that is, in c markd priority will be less than mark in the b before c.V is simultaneously It is a parameter controlling false appearance to close with false unrelated Type Ⅰ Ⅱ error punishment degree, v > 1 effectively represents, and cut-point is marked relative to other Note has bigger importance most, does not more allow to make a mistake in its position, and virtual tag is as dividing mark of correlation and extraneous markings One natural " zero point ", allows algorithm automatically select appropriate division points, than artificial given threshold method more science, accurately.

Indicia distribution q that the characteristic vector of Fig. 3 is corresponding constitutes a complete sample together.Continue from image input Equipment obtains natural scene image, equally carries out front three step process to it, until obtaining appropriate sample.Then train to whole Collection, obtains the optimal parameter vector θ of natural scene indicia distribution model with lbfgs-lld Algorithm for Training.

Make x=r^dRepresent the input space, y={ y₁,y₂,...,y_cExpress possibility mark constitute finite aggregate, then Problem concerning study based on indicia distribution can be described as follows: given training set s={ (x₁,p₁(y)),(x₂,p₂(y)),…,(x_n,p_n (y)) }, wherein x_i∈ x represents an example, p_iY () is and x_iThe distribution of related stochastic variable y ∈ y, the target of study is To condition distribution p { y | x } it is assumed that p the parameter model of { y | x } be p y | x;θ }, whereinIt is the parameter vector of this model. The target of lbfgs-lld algorithm is in given training set s and example x_iIn the case of ∈ x, study obtains optimal parameter vector, Pass throughOne and p can be generated_iY () most like distribution, to weigh the similarity of two distributions used here as kl distance.

The basic thought of lbfgs-lld algorithm is that the indicia distribution form to example is assumed it is assumed that this distribution is one Individual with parameter and the distribution of relevant with example x condition:Wherein z=σ_yexp(σ_k θ_y,kg_k(x)) it is standardizing factor, g_kX () is the characteristic function of example x, such as given example x, then g_kX () is certain width figure of x As feature, θ_y,kFor unknown parameter to be solved, once θ_y,kDetermine, the indicia distribution of this example can determine that, according to description degree Size mark is ranked up to obtain the corresponding tag sort of this example.This is substantially a maximum entropy model vacation If.

The parameter vector θ of wherein model is obtained by solving the such unconstrained optimization problem of (1) formula.

$\begin{array}{c}{\mathrm{\θ}}^{*}=\underset{\mathrm{\θ}}{\arg \min }\underset{i}{\mathrm{\σ}}\underset{y}{\mathrm{\σ}}\left(p_i\left(y\right)\log \frac{p_i\left(y\right)}{p\left(y\right|x_i,\mathrm{\θ})}\right)\&doubleleftrightarrow;\underset{\mathrm{\θ}}{\arg \max }\underset{i}{\mathrm{\σ}}\underset{y}{\mathrm{\σ}}{p}_i\left(y\right)\log p\left(y\right|x_i,\mathrm{\θ})\\ \&doublerightarrow;\underset{\mathrm{\θ}}{\arg \max }t\left(\mathrm{\θ}\right)=\underset{i,y}{\mathrm{\σ}}{p}_i\left(y\right)\underset{k}{\mathrm{\σ}}{\mathrm{\θ}}_{y,k}{g}_k\left(x_i\right)-\underset{i}{\mathrm{\σ}}\log \underset{y}{\mathrm{\σ}}\exp \left(\underset{k}{\mathrm{\σ}}{\mathrm{\θ}}_{y,k}{g}_k\left(x_i\right)\right)\end{array}---\left(1\right)$

Present invention employs lbfgs algorithm to solve this unconstrained optimization problem, lbfgs algorithm is the limited of bfgs algorithm Internal memory version is it is adaptable to process (more than 1000) situation of known variables large number.Solving unconstrained optimization problem at present should It is most widely pseudo-Newtonian algorithm, wherein bfgs algorithm is considered as current maximally effective Quasi-Newton algorithm.Lbfgs simultaneously Algorithm or the general training algorithm of maximum entropy model, its memory consumption is little, and fast convergence rate can be quickly obtained near-optimization Solution.Lbfgs algorithm specifically may be referred to j.nocedal.updating quasi-newton matrices with limited Storage (1980), mathematics of computation35, pp.773-782 and d.c.liu and j.nocedal.on the limited memory method for large scale optimization(1989), mathematical programming b,45,3,pp.503-528.

Input test natural scene image, and the characteristic vector of this image is extracted with the feature extracting method of the first step.Will The characteristic vector of optimal parameter vector θ and test natural scene image substitutes in natural scene indicia distribution model respectively, you can Obtain the indicia distribution of this natural scene image.

Degree described in the indicia distribution obtaining is regarded as extraneous markings less than the mark of virtual tag degree of description, and remaining Mark is mark of correlation, finally mark of correlation is ranked up according to description degree size.Thus obtain test nature field The tag sort of scape image.

What the present invention took is 10 times of cross validation methods, experimental result employ common several for tolerance two have The index of sequence grade variables similitude, has gamma coefficient, spearman coefficient, sag coefficient and kendall coefficient to be used for processing Two class mutation, tau_b coefficient and the tau_c coefficient of draw.Control methods is included currently used for process multiple labeling sequencing problem effect Really best cknn+median algorithm (is published in ijcai's with reference to klaus brink and eyke hullermeier2007 Paper " cased-based multilabel ranking "), crpc algorithm and clrt algorithm.

Table 3: the result of invention

	gamma	spearman	sag	tau_b	tau_c
						lbfgs_lld	0.6241	0.5660	0.2534	0.3461	0.3405
cknn+median	0.6070	0.5550	0.2442	0.3344	0.3293
						crpc	0.5567	0.5286	0.2240	0.3065	0.3020
clrt	0.5002	0.5025	0.2057	0.2807	0.2750

Intuitively can see from table 3, the present invention is effective on processing natural scene tag sort problem.This Multi-user's mask method is taken in invention first, then effectively solves multi-user's mark using mathematical Optimized model and brings Inconsistence problems, meanwhile, the introducing of virtual tag be also divide mark of correlation and extraneous markings provide more scientific, more accurately Method, finally, by the learning algorithm based on indicia distribution, take full advantage of the existing sequencing information of training set it is achieved that The increment of information value, these measures both contribute to improve the accuracy predicting the outcome.

Claims (3)

1. a kind of multi-user natural scene tag sort method is it is characterised in that comprise the steps:

(1) obtain the natural scene image set for training, and characteristic vector is extracted to every width natural scene image；

(2) multiple tag sorts of natural scene image are obtained from the image labeling system based on user interest degree；

(3) multiple tag sorts are converted into an indicia distribution；

(4) obtain the natural scene image of sequence to be marked from input equipment, extract characteristic vector；

(5) if recognition mechanism does not train, execution step (6), otherwise go to step (7)；

(6) the natural scene image characteristic vector being obtained with step (1) to step (3) and its tag sort information are as training Collection, using the Algorithm for Training of lbfgs-lld, obtains the optimal parameter vector θ of natural scene indicia distribution model；

(7) natural scene image of the sequence to be marked that the optimal parameter obtaining in step (6) vector θ and step (4) are obtained Characteristic vector substitute into respectively in natural scene indicia distribution model, obtain this to be marked sequence natural scene image mark divide Cloth；

(8) described in the indicia distribution that step (7) is obtained, degree regards extraneous markings as less than the mark of virtual tag degree of description, and Remaining mark is mark of correlation, finally mark of correlation is ranked up according to description degree size；

Wherein, the method for step (1) and step (4) extraction characteristic vector is: coloured image is converted into cieluv color space, Then the method carrying out piecemeal again extracts the characteristics of image of natural scene；

Step (3) is passed through to build a nonlinear optimal problem, and the concrete construction method of described nonlinear optimal problem is: to figure As, for m, l is used in the tag sort of n position user respectively₁,l₂...l_nRepresent it is assumed that each user is respectively according to indicia distribution p₁, p₂...p_nMark is ranked up, and the indicia distribution combining all consumers' opinions is represented with q, model is with minimum k l distance For object function, that is, make q and p₁,p₂...p_nBetween difference minimum；And solved using interior point method, by multiple mark rows Sequence is converted into an indicia distribution.

2. multi-user natural scene tag sort method as claimed in claim 1 is it is characterised in that described nonlinear optimization is asked The concrete construction method of topic is: for image m, l is used in the tag sort of n position user respectively₁,l₂...l_nRepresent it is assumed that each User is respectively according to indicia distribution p₁,p₂...p_nMark is ranked up, and combine the indicia distribution of all consumers' opinions Represented with q, model with minimum k l distance as object function, that is, makes q and p₁,p₂...p_nBetween difference minimum, also should simultaneously Meet some constraint formulas, that is, for distribution p₁,p₂...p_nFor, in tag sort, the description degree coming mark above should When the description degree more than the mark coming below；For an indicia distribution, the description degree of its all possible mark it With should be 1.

3. multi-user natural scene tag sort method as claimed in claim 1 is it is characterised in that step (5) judge mark It is distributed the average kl distance that the basis for estimation whether training is between the indicia distribution of model prediction and real indicia distribution Whether sufficiently small.