CN105389583A - Image classifier generation method, and image classification method and device - Google Patents

Abstract

The invention discloses an image classifier generation method and device, and the method comprises the steps: obtaining a training sample set which comprises N image samples belonging to K classes, wherein N and K are positive integers, and N is greater than K; obtaining a characteristic vector of each image sample, wherein each characteristic vector comprises a hidden variable of the corresponding image sample; and training classifiers for K classes based on the hidden variables of N image samples through a multiple logic regression model. According to the embodiment of the invention, the method trains the classifiers for K classes at the same time in a mode of maximum likelihood through the multiple logic regression model. In other words, the mutual correlation among classifiers of the K classes is maintained in the application of the multiple logic regression model. Compared with a mode of converting a K-class classification problem in the field of object classification through LVSM into a plurality of mutually independent two-class problems, the method is more accurate in training results.

Description

The generation method of Image Classifier, image classification method and device

Technical field

The present invention relates to Images Classification field, and more specifically, relate to a kind of generation method of Image Classifier, image classification method and device.

Background technology

Hidden variable refers to directly to be observed, the comprehensive variable but played an important role in actual applications, as spatial relationship, data structure, inline state etc.Hidden variable is widely used in the fields such as machine vision, natural language processing, speech recognition and public health.Experiment proves, during the process object such as image, voice, the introducing of hidden variable can catch more useful information, and compared with only using the mode of aobvious variable, obvious processing effect improves.

Early stage latent variable model mostly is generation model (generativemodels), as hidden Markov model (HiddenMarkovModel, HMM), gauss hybrid models (GaussianMixtureModel, GMM) etc.More researcher attempts to seek the possibility introducing hidden variable in discrimination model (discriminativemodels) in the recent period.Typical example is as condition random field (ConditionalRandomField, CRF), hidden variable support vector machine (LatentSupportVectorMachine, LSVM) etc., and these models all achieve certain achievement in respective field.It is worth mentioning that, LSVM coordinates local deformable model (DeformablePart-basedModel, DPM), i.e. DPM-LSVM, and the object detecting areas in machine vision has become in recent years comparatively successfully algorithm.DPM detects the feature of classification object for describing, it is made up of three parts: a main body wave filter (rootfilter), multiple local filter (partfilters), and deformation punishment (deformablecosts) of each local correspondence.Main part is for describing the general outline of object, and Part portions is for describing the minutia of inspected object, and deformation punishment is for ensureing that each Local Phase can not have excessive skew for the position of main body.In object detection process, Local Phase can change within the specific limits for the position of main body, can regard hidden variable as, adopts LSVM to train.

The objective function form of LSVM is similar to original SVM, as shown in (1):

$L\left(\mathrm{\β}\right)=\frac{1}{2}{\left|\right|\mathrm{\β}\left|\right|}^2+C\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\max (\mathrm{0,1}-y_i{s}_{\mathrm{\β}}\left(x_i\right))---\left(1\right)$

Wherein, β is the model parameter of sorter, y _irepresent training sample x _ilabel, s (x _i, β) and represent sample x _imark, this mark be likely Local Phase is to mark optimum in position (i.e. hidden variable span), this mark meets formula (2):

$s(x_i,\mathrm{\β})=\underset{z\∈Z\left(x_i\right)}{\max }\mathrm{\β}\·f(x_i,z)---\left(2\right)$

In formula (2), z is hidden variable, and f is feature extracting method, f (x _i, z) be sample x _iproper vector, as used framework histogram of gradients feature in DPM.

Can prove that the objective function (formula (1)) of LSVM has half concavity, when namely fixing the hidden variable value of positive sample, objective function is recessed.Therefore, solving of LSVM can use coordinate gradient to decline (CoordinateGradientDescent), the i.e. model parameter of first fixed cluster device, try to achieve positive sample hidden variable value, fix positive sample hidden variable value again, ask optimization model parameter and negative sample hidden variable value, iteration like this is until convergence.

LSVM and SVM is the same, is mainly applicable to object detecting areas.When being generalized to object classification field, the processing mode of LSVM is the two class problems multi-class problem in object classification field being changed into object detecting areas.Adopt this kind of processing mode, the training process of the multiple sorters for object classification can be made to isolate each other.In reality, may there is certain relevance between multiple object classification, such as, buildings is divided into multiclass architectural style, the buildings in picture to be sorted may have the feature of two or more architectural style simultaneously.Therefore, the training process of multiple sorter is changed into multiple two class problems isolated, either-or each other, classification results can be caused inaccurate.

Summary of the invention

The embodiment of the present invention provides a kind of generation method and apparatus of Image Classifier, to improve the accuracy of classification results.

First aspect, provides a kind of generation method of Image Classifier, comprising: obtain training sample set, described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K; Obtain the proper vector of image pattern described in each, wherein, described proper vector comprises the hidden variable of image pattern; Based on the hidden variable of described N number of image pattern, by multivariate logistic regression model, train the sorter of a described K classification.

In conjunction with first aspect, in a kind of implementation of first aspect, the sorter of a described K classification comprises K model parameter respectively, the described hidden variable based on described N number of image pattern, by multivariate logistic regression model, train the sorter of a described K classification, comprising: the initial value obtaining a described K model parameter; Obtain the initial value of the hidden variable of described N number of image pattern; Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, the described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter, comprise: based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, when the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter, when the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, comprise: based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable, when the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter, otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, comprising: according to formula determine the iterative value of a described K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, described according to formula determine the iterative value of a described K model parameter, comprising: according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or described iteration stopping condition is that iterations reaches preset times.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, described according to formula determine the iterative value of a described K model parameter, comprising: according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the described proper vector based on described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, comprising: according to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the described proper vector based on described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, comprising: according to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

In conjunction with any one of first aspect or its above-mentioned implementation, in the another kind of implementation of first aspect, the described initial value based on model parameter described in each, determines the initial value of the hidden variable of image pattern described in each, comprising: according to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

Second aspect, provides a kind of image classification method, comprising: the proper vector obtaining image to be classified; Based on the proper vector of described image to be classified, utilize K sorter, determine the classification of described image to be classified, wherein, a described K sorter is K the sorter utilizing any one implementation of first aspect or first aspect to train; According to formula determine the probability of described image to be classified under a described K classification, wherein, x represents described image to be classified, β _krepresent the model parameter of a kth sorter in a described K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

The third aspect, provides a kind of generating apparatus of Image Classifier, comprising: the first acquiring unit, and for obtaining training sample set, described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K; Second acquisition unit, for obtain that described first acquiring unit obtains each described in the proper vector of image pattern, wherein, described proper vector comprises the hidden variable of image pattern; Training unit, for the hidden variable of described N number of image pattern obtained based on described second acquisition unit, by multivariate logistic regression model, trains the sorter of a described K classification.

In conjunction with second aspect, in a kind of implementation of second aspect, the sorter of a described K classification comprises K model parameter respectively, and described training unit is specifically for obtaining the initial value of a described K model parameter; Obtain the initial value of the hidden variable of described N number of image pattern; Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

In conjunction with the third aspect, in a kind of implementation of the third aspect, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, described training unit is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, when the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter, when the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable, when the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter, otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula determine the iterative value of a described K model parameter, wherein, x _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or described iteration stopping condition is that iterations reaches preset times.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

In conjunction with any one of the third aspect or its above-mentioned implementation, in the another kind of implementation of the third aspect, described training unit is specifically for according to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

Fourth aspect, provides a kind of image classification device, comprising: the first acquiring unit, for obtaining the proper vector of image to be classified; First determining unit, for the proper vector based on described image to be classified, utilize K sorter, determine the classification of described image to be classified, wherein, a described K sorter is K the sorter utilizing any one implementation of the third aspect or the third aspect to train; Second determining unit, for according to formula determine the probability of described image to be classified under a described K classification, wherein, x represents described image to be classified, β _krepresent the model parameter of a kth sorter in a described K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

In the embodiment of the present invention, by multivariate logistic regression model, with form simultaneous training K sorter of maximum likelihood, that is, interrelated between the sorter that the use of multivariate logistic regression model remains K classification, compared with the mode K class classification problem in object classification field being converted to mutually isolated multiple two class problems with LVSM, training result is more accurate.

Accompanying drawing explanation

In order to be illustrated more clearly in the technical scheme of the embodiment of the present invention, be briefly described to the accompanying drawing used required in the embodiment of the present invention below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is the indicative flowchart of the generation method of the Image Classifier of the embodiment of the present invention.

Fig. 2 be utilize the embodiment of the present invention to train classifier parameters to the exemplary plot of Images Classification.

Fig. 3 be utilize the embodiment of the present invention to train classifier parameters to the exemplary plot of Images Classification.

Fig. 4 is the schematic diagram of the generating apparatus of the Image Classifier of the embodiment of the present invention.

Fig. 5 is the schematic diagram of the generating apparatus of the Image Classifier of the embodiment of the present invention.

Fig. 6 is the indicative flowchart of the image classification method of the embodiment of the present invention.

Fig. 7 is the schematic block diagram of the image classification device of the embodiment of the present invention.

Fig. 8 is the schematic block diagram of the image classification device of the embodiment of the present invention.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is a part of embodiment of the present invention, instead of whole embodiment.Based on the embodiment in the present invention, the every other embodiment that those of ordinary skill in the art obtain under the prerequisite of not making creative work, all should belong to the scope of protection of the invention.

Fig. 1 is the indicative flowchart of the generation method of the Image Classifier of the embodiment of the present invention.The method of Fig. 1 comprises:

110, obtain training sample set, training sample set comprises N number of image pattern, and N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K.

Such as, training sample set D={ (x ₁, y ₁) ..., (x _n, y _n), comprise N number of image pattern altogether, wherein, y _ifor image pattern x _ilabel, be used to indicate x _iclassification, this classification is one of above-mentioned K classification.

120, obtain the proper vector of each image pattern, wherein, proper vector comprises the hidden variable of image pattern.

Should be understood that characteristics of image and hidden variable can be chosen according to application scenarios or actual needs.Such as, characteristics of image can be chosen (or being defined as) histograms of oriented gradients (HistogramofOrientedGradient, HOG), local binary patterns (LocalBinaryPatterns, LBP), or Haar etc.; Hidden variable can be chosen (or being defined as) object position in the picture, local and intersubjective relative position in image, or the subclass etc. of object.Based on the above-mentioned characteristics of image chosen and hidden variable, obtain the proper vector of each image pattern, now, the proper vector of each image of acquisition is not a fixed value, can change, suppose image x along with the change of hidden variable _ihidden variable be z, the proper vector extracted represents by f (x, z).

130, based on the hidden variable of N number of image pattern, by multivariate logistic regression model, the sorter of training K classification.

In the embodiment of the present invention, by multivariate logistic regression model, with form simultaneous training K sorter of maximum likelihood, that is, interrelated between the sorter that the use of multivariate logistic regression model remains K classification, compared with the mode K class classification problem in object classification field being converted to mutually isolated multiple two class problems with LVSM, training result is more accurate.

Alternatively, as an embodiment, step 130 can comprise: the initial value obtaining K model parameter; Obtain the initial value of the hidden variable of N number of image pattern; Based on the proper vector of N number of image pattern, and the initial value of N number of image pattern hidden variable, by multivariate logistic regression model, the sorter of training K classification, to determine the desired value of K model parameter.

It should be noted that, the hidden variable of an image pattern can comprise K initial value, and that is, the hidden variable of an image pattern has a corresponding initial value under the initial value of a model parameter.By step 130, the initial value of N*K hidden variable can be obtained.

Alternatively, as an embodiment, the initial value of above-mentioned N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, the above-mentioned proper vector based on N number of image pattern, and the initial value of N number of image pattern, by multivariate logistic regression model, the sorter of training K classification, to determine the desired value of K model parameter, can comprise: based on the proper vector of N number of image pattern, and the initial value of the hidden variable of N number of image pattern, by multivariate logistic regression model, the sorter of training K classification, to determine the currency of K model parameter, when the currency of K model parameter meets the default condition of convergence, the currency of K model parameter is defined as the desired value of K model parameter, when the currency of K model parameter does not meet this condition of convergence, based on the proper vector of N number of image pattern, and the currency of K model parameter, determine the currency of positive image pattern hidden variable, and utilize the currency of positive image pattern hidden variable to upgrade the initial value of this positive image pattern hidden variable, repeat this step until the currency of K model parameter meets the condition of convergence.

Specifically, the hidden variable of an image pattern can have different initial values under different model parameter, and that is the hidden variable of an image pattern can comprise K initial value, and the initial value of above-mentioned N number of image pattern hidden variable can comprise: K*N initial value.An image pattern is positive sample under the model parameter that this image pattern classification is corresponding, and the initial value of above-mentioned positive image pattern hidden variable comprises N number of initial value altogether, is the initial value of N number of image pattern under the model parameter that respective classification is corresponding respectively.In K*N initial value, remove remaining K* (N-1) individual initial value outside above-mentioned positive image hidden variable initial value and be the initial value of negative image sample hidden variable.

Can prove, when positive image pattern hidden variable initial value is fixed, multivariate logistic regression model has concavity, and the mode that can be risen by gradient is solved.

Alternatively, as an embodiment, the above-mentioned proper vector based on N number of image pattern, and the initial value of the hidden variable of N number of image pattern, by multivariate logistic regression model, the sorter of training K classification, to determine that the currency of K model parameter can comprise: based on the proper vector of N number of image pattern, and the initial value of N number of image pattern hidden variable, by multivariate logistic regression model, the sorter of training K classification, to determine the iterative value of K model parameter, based on the proper vector of N number of image pattern, and the iterative value of K model parameter, determine the iterative value of negative image sample hidden variable, and utilize the iterative value of negative image sample hidden variable to upgrade the initial value of negative image sample hidden variable, when the iterative value of K model parameter meets default iteration stopping condition, the iterative value of K model parameter is defined as the currency of K model parameter, otherwise, repeat this step until the currency of K model parameter meets iteration stopping condition.

In the embodiment of the present invention, when fixing positive sample hidden variable value, reached the object optimizing K model parameter by the value constantly updating negative sample hidden variable, further increase the accuracy of classification results.

Alternatively, as an embodiment, the above-mentioned proper vector based on N number of image pattern, and the initial value of N number of image pattern hidden variable, by multivariate logistic regression model, the sorter of training K classification, to determine that the iterative value of K model parameter can comprise: according to formula determine the iterative value of K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in N number of image pattern, β _lrepresent l model parameter in K model parameter, θ represents that the K of K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

Alternatively, as an embodiment, above-mentioned according to formula determine the iterative value of K model parameter, can comprise: according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines β _kiterative value.

Alternatively, as an embodiment, above-mentioned iteration stopping condition is that the change of target function value l (θ) is less than predetermined threshold value; Or iteration stopping condition is that iterations reaches preset times.

Alternatively, as an embodiment, above-mentioned according to formula determine the iterative value of K model parameter, can comprise: according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of a parallel computation K model parameter, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

There is logarithm additive function in above-mentioned objective function l (θ), therefore, cannot resolve into K class subproblem superposition form, also just cannot adopt walk abreast or Distributed Calculation searching process is accelerated.

In the embodiment of the present invention, logarithm is utilized to have concavity (Log-concavity), adopt the recessed upper bound of logarithm (Log-concavityBound) objective function l (θ) is converted into K class subproblem add and form, thus can parallel computation be realized, accelerate convergence of algorithm.

Specifically, the form in the recessed upper bound of logarithm is: utilize this formula just l (θ) can be converted into:

$l_{\mathrm{LC}}(\mathrm{\θ},A)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1]$

Adopt above formula as objective function, when utilizing gradient rise method to solve, the form of the gradient of classifier parameters is as follows: $\frac{{\∂l}_{\mathrm{LC}}}{{\∂\mathrm{\β}}_l}=[-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_l\right)+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_l\right))\·(y_{\mathrm{il}}-a_i\exp \left(s(x_i;{\mathrm{\β}}_l)\right))],$ Wherein, auxiliary parameter a _ivalue is: $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

Alternatively, as an embodiment, the above-mentioned proper vector based on N number of image pattern, and the iterative value of K model parameter, determine that the iterative value of negative image sample hidden variable can comprise: according to formula determine the iterative value of negative image sample hidden variable, wherein, x _irepresent the i-th sample in N number of image pattern, β _trepresent t model parameter in K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

Alternatively, as an embodiment, the above-mentioned proper vector based on N number of image pattern, and the currency of K image pattern, determine that the currency of positive image pattern hidden variable can comprise: according to formula determine the currency of positive image pattern hidden variable, wherein, x _irepresent the i-th sample in N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

Alternatively, as an embodiment, the above-mentioned initial value based on each model parameter, determines that the initial value of the hidden variable of each image pattern can comprise: according to formula determine the initial value of the hidden variable of each image pattern, wherein, x _irepresent the i-th sample in N number of image pattern, β _krepresent the kth model parameter in K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

Below in conjunction with concrete example, describe the embodiment of the present invention in detail.It should be noted that these examples are just in order to help those skilled in the art to understand the embodiment of the present invention better, and the scope of the unrestricted embodiment of the present invention.

Embodiment 1:

Input: training sample set { (x ₁, y ₁) ..., (x _n, y _n), initial all hidden variable values.

Export: classifier parameters θ, θ={ β ₁..., β _k.

Embodiment 2:

Input: training sample set { (x ₁, y ₁) ..., (x _n, y _n), initial hidden variable value { h}.

Export: classifier parameters θ.

In specific implementation, the value of constant numOuterLoop and numInnerLoop and application scenarios have comparatively Important Relations, as in numeral identification (digitrecognition), because sample size is many, characteristic dimension is little, can establish numOuterLoop=50, numInnerLoop=1.

In more complicated example, as sample size is little, characteristic dimension is high, can establish numOuterLoop=5, numInnerLoop=1000.

Provide the classifier parameters that trains below to the result of Images Classification.It should be noted that, in the following description, the sorter training patterns of the embodiment of the present invention is called: hidden variable multivariate logistic regression (MultinomialLatentLogisticRegression, MLLR).

Fig. 2 be utilize the embodiment of the present invention to train classifier parameters to the exemplary plot of Images Classification.Be categorized as research object with mammal in the example of Fig. 2, comprise 6 class mammals altogether, every class about 50 pictures.Get 50% picture in experiment as training, another 50% image is as test.Characteristics of image aspect uses HOG feature, and hidden variable is the position of examined object in picture, and specifies that the size of object place frame will at more than 30% of total picture size.Linear SVM, LSVM and MLLR, test result is as follows:

Table 1 mammal classification experiments classification results

Sorting technique	Linear SVM	LSVM	MLLR
				Accuracy rate (%)	64.23	69.59	73.31

Test result shows, the accuracy rate of MLLR is more than LSVM, and the effect of sorter that LSVM and MLLR two kinds of hidden variable modes train all is better than conventional linear SVM method.

In Fig. 2, first is classified as the sorter schematic diagram (adopting HOG feature) that Linear SVM trains, and second is classified as the sorter schematic diagram that MLLR trains.Rectangle frame in Fig. 2 in little picture is the object space that MLLR detects.

Fig. 3 be utilize the embodiment of the present invention to train classifier parameters to the exemplary plot of Images Classification.Fig. 3 for research object, comprises 6 class actions (cricket batting, cricket pitching, volleyball smash, croquet batting, tennis forehand and tennis service) with physical culture figure action altogether.Characteristics of image still uses HOG, and latent variable model uses DPM, and namely object space and local main body relative position are all as hidden variable.Result display classification accuracy MLLR (78.3%) is more than LSVM (74.4%).In Fig. 3, first is classified as the agent model schematic diagram in picture, and second is classified as the partial model schematic diagram in picture, and in Fig. 3, in little picture, dark rectangular frame represents body position, and light rectangle frame represents local location.Should understand, in various embodiments of the present invention, the size of the sequence number of above-mentioned each process does not also mean that the priority of execution sequence, and the execution sequence of each process should be determined with its function and internal logic, and should not form any restriction to the implementation process of the embodiment of the present invention.

Above composition graphs 1 to Fig. 3, describes the generation method of the Image Classifier according to the embodiment of the present invention in detail, below in conjunction with Fig. 4 to Fig. 5, describes the generating apparatus of the Image Classifier according to the embodiment of the present invention.

Should be understood that each step that can to realize according to the generating apparatus of the Image Classifier of the embodiment of the present invention in Fig. 1, for simplicity, do not repeat them here.

Fig. 4 is the schematic diagram of the generating apparatus of the Image Classifier of the embodiment of the present invention.The device 400 of Fig. 4 comprises:

First acquiring unit 410, for obtaining training sample set, described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K;

Second acquisition unit 420, for obtain that described first acquiring unit 410 obtains each described in the proper vector of image pattern, wherein, described proper vector comprises the hidden variable of image pattern;

Training unit 430, for the hidden variable of the described N number of image pattern based on described second acquisition unit 420 acquisition, by multivariate logistic regression model, trains the sorter of a described K classification.

In the embodiment of the present invention, by multivariate logistic regression model, with form simultaneous training K sorter of maximum likelihood, that is, interrelated between the sorter that the use of multivariate logistic regression model remains K classification, compared with the mode K class classification problem in object classification field being converted to mutually isolated multiple two class problems with LVSM, training result is more accurate.

Alternatively, as an embodiment, the sorter of a described K classification comprises K model parameter respectively, and described training unit 430 is specifically for obtaining the initial value of a described K model parameter; Based on the initial value of model parameter described in each, determine the initial value of the hidden variable of image pattern described in each; Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

Alternatively, as an embodiment, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, described training unit 430 is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, when the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter, when the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

Alternatively, as an embodiment, described training unit 430 is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable, when the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter, otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula determine the iterative value of a described K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

Alternatively, as an embodiment, described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or described iteration stopping condition is that iterations reaches preset times.

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

Alternatively, as an embodiment, described training unit 430 is specifically for according to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

Fig. 5 is the schematic diagram of the generating apparatus of the Image Classifier of the embodiment of the present invention.The device 500 of Fig. 5 comprises:

Storer 510, for storage program;

Processor 520, for performing described program, when described program is performed, described processor 520 is specifically for obtaining training sample set, and described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, N, K are positive integer, and N is greater than K; Obtain the proper vector of image pattern described in each, wherein, described proper vector comprises the hidden variable of image pattern; Based on the hidden variable of described N number of image pattern, by multivariate logistic regression model, train the sorter of a described K classification.

In the embodiment of the present invention, by multivariate logistic regression model, with form simultaneous training K sorter of maximum likelihood, that is, interrelated between the sorter that the use of multivariate logistic regression model remains K classification, compared with the mode K class classification problem in object classification field being converted to mutually isolated multiple two class problems with LVSM, training result is more accurate.

Alternatively, as an embodiment, the sorter of a described K classification comprises K model parameter respectively, and described processor 520 is specifically for obtaining the initial value of a described K model parameter; Based on the initial value of model parameter described in each, determine the initial value of the hidden variable of image pattern described in each; Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

Alternatively, as an embodiment, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, described processor 520 is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, when the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter, when the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

Alternatively, as an embodiment, described processor 520 is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable, when the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter, otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

Alternatively, as an embodiment, described processor 520 is specifically for according to formula determine the iterative value of a described K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

Alternatively, as an embodiment, described processor 520 is specifically for according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

Alternatively, as an embodiment, described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or described iteration stopping condition is that iterations reaches preset times.

Alternatively, as an embodiment, described processor 520 is specifically for according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

Alternatively, as an embodiment, described processor 520 is specifically for according to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

Alternatively, as an embodiment, described processor 520 is specifically for according to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

Alternatively, as an embodiment, described processor 520 is specifically for according to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

Fig. 6 is the indicative flowchart of the image classification method of the embodiment of the present invention.In the method for Fig. 6, K the sorter that Fig. 1 method can be utilized to train is classified to image, and Fig. 6 method comprises:

610, the proper vector of image to be classified is obtained;

620, based on the proper vector of image to be classified, utilize K sorter, determine the classification of image to be classified;

630, according to formula determine the probability of image to be classified under K classification, wherein, x represents image to be classified, β _krepresent the model parameter of a kth sorter in K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

The classification results of existing LSVM only provides image to be classified and which kind of belongs to, but in actual conditions, may there is certain contact between dissimilar, and which kind of a certain image not definitely belongs to.Such as, the style of buildings can be classified, comprise smartness, Middle Ages style etc., in image, a certain buildings style may both have employed some smartness, also used part style in Middle Ages, now, which kind of architectural style the classification results of the existing LSVM only buildings that can show in image to be classified is classified as, obviously not accurate enough.In the present embodiment, except providing the classification belonging to image to be classified, give the probability of this picture in of all categories, compared with prior art, the probability interpretation introducing Images Classification result makes the description of Images Classification result more accurate.

Fig. 7 is the schematic block diagram of the device of the Images Classification of the embodiment of the present invention.K the sorter that device 700 in Fig. 7 can utilize the device 400 of Fig. 4 to train is classified to image, and device 700 comprises:

First acquiring unit 710, for obtaining the proper vector of image to be classified;

First determining unit 720, for the proper vector based on image to be classified, utilizes K sorter, determines the classification of image to be classified;

Second determining unit 730, for according to formula determine the probability of image to be classified under K classification, wherein, x represents image to be classified, β _krepresent the model parameter of a kth sorter in K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

The classification results of existing LSVM only provides image to be classified and which kind of belongs to, but in actual conditions, may there is certain contact between dissimilar, and which kind of a certain image not definitely belongs to.Such as, the style of buildings can be classified, comprise smartness, Middle Ages style etc., in image, a certain buildings style may both have employed some smartness, also used part style in Middle Ages, now, which kind of architectural style the classification results of the existing LSVM only buildings that can show in image to be classified is classified as, obviously not accurate enough.In the present embodiment, except providing the classification belonging to image to be classified, give the probability of this picture in of all categories, compared with prior art, the probability interpretation introducing Images Classification result makes the description of Images Classification result more accurate.

Fig. 8 is the schematic block diagram of the device of the Images Classification of the embodiment of the present invention.K the sorter that image classification device 800 in Fig. 8 can utilize the device 500 of Fig. 5 to train is classified to image, and Fig. 8 method comprises:

Storer 810, for storage program;

Processor 820, for executive routine, when described program is performed, described program is for obtaining the proper vector of image to be classified; Based on the proper vector of image to be classified, utilize K sorter, determine the classification of image to be classified; According to formula determine the probability of image to be classified under K classification, wherein, x represents image to be classified, β _krepresent the model parameter of a kth sorter in K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

The classification results of existing LSVM only provides image to be classified and which kind of belongs to, but in actual conditions, may there is certain contact between dissimilar, and which kind of a certain image not definitely belongs to.Such as, the style of buildings can be classified, comprise smartness, Middle Ages style etc., in image, a certain buildings style may both have employed some smartness, also used part style in Middle Ages, now, which kind of architectural style the classification results of the existing LSVM only buildings that can show in image to be classified is classified as, obviously not accurate enough.In the present embodiment, except providing the classification belonging to image to be classified, give the probability of this picture in of all categories, compared with prior art, the probability interpretation introducing Images Classification result makes the description of Images Classification result more accurate.

Should be understood that in embodiments of the present invention, term "and/or" is only a kind of incidence relation describing affiliated partner, and expression can exist three kinds of relations.Such as, A and/or B, can represent: individualism A, exists A and B simultaneously, these three kinds of situations of individualism B.In addition, character "/" herein, general expression forward-backward correlation is to the relation liking a kind of "or".

Those of ordinary skill in the art can recognize, in conjunction with unit and the algorithm steps of each example of embodiment disclosed herein description, can realize with electronic hardware, computer software or the combination of the two, in order to the interchangeability of hardware and software is clearly described, generally describe composition and the step of each example in the above description according to function.These functions perform with hardware or software mode actually, depend on application-specific and the design constraint of technical scheme.Professional and technical personnel can use distinct methods to realize described function to each specifically should being used for, but this realization should not thought and exceeds scope of the present invention.

Those skilled in the art can be well understood to, and for convenience of description and succinctly, the specific works process of the system of foregoing description, device and unit, with reference to the corresponding process in preceding method embodiment, can not repeat them here.

In several embodiments that the application provides, should be understood that disclosed system, apparatus and method can realize by another way.Such as, device embodiment described above is only schematic, such as, the division of described unit, be only a kind of logic function to divide, actual can have other dividing mode when realizing, such as multiple unit or assembly can in conjunction with or another system can be integrated into, or some features can be ignored, or do not perform.In addition, shown or discussed coupling each other or direct-coupling or communication connection can be indirect coupling by some interfaces, device or unit or communication connection, also can be electric, machinery or other form connect.

The described unit illustrated as separating component or can may not be and physically separates, and the parts as unit display can be or may not be physical location, namely can be positioned at a place, or also can be distributed in multiple network element.Some or all of unit wherein can be selected according to the actual needs to realize the object of embodiment of the present invention scheme.

In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, and also can be that the independent physics of unit exists, also can be that two or more unit are in a unit integrated.Above-mentioned integrated unit both can adopt the form of hardware to realize, and the form of SFU software functional unit also can be adopted to realize.

If described integrated unit using the form of SFU software functional unit realize and as independently production marketing or use time, can be stored in a computer read/write memory medium.Based on such understanding, technical scheme of the present invention is in essence in other words to the part that prior art contributes, or all or part of of this technical scheme can embody with the form of software product, this computer software product is stored in a storage medium, comprising some instructions in order to make a computer equipment (can be personal computer, server, or the network equipment etc.) perform all or part of step of method described in each embodiment of the present invention.And aforesaid storage medium comprises: USB flash disk, portable hard drive, ROM (read-only memory) (ROM, Read-OnlyMemory), random access memory (RAM, RandomAccessMemory), magnetic disc or CD etc. various can be program code stored medium.

The above; be only the specific embodiment of the present invention; but protection scope of the present invention is not limited thereto; anyly be familiar with those skilled in the art in the technical scope that the present invention discloses; can expect amendment or the replacement of various equivalence easily, these amendments or replacement all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (24)

1. a generation method for Image Classifier, is characterized in that, comprising:

Obtain training sample set, described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K;

Obtain the proper vector of image pattern described in each, wherein, described proper vector comprises the hidden variable of image pattern;

Based on the hidden variable of described N number of image pattern, by multivariate logistic regression model, train the sorter of a described K classification.

2. the method for claim 1, is characterized in that, the sorter of a described K classification comprises K model parameter respectively,

The described hidden variable based on described N number of image pattern, by multivariate logistic regression model, train the sorter of a described K classification, comprising:

Obtain the initial value of a described K model parameter;

Obtain the initial value of the hidden variable of described N number of image pattern;

Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

3. method as claimed in claim 2, it is characterized in that, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable,

The described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter, comprising:

Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter,

When the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter,

When the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

4. method as claimed in claim 3, it is characterized in that, the described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, comprising:

Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter,

Based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable,

When the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter,

Otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

5. method as claimed in claim 4, it is characterized in that, the described proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, comprising:

According to formula determine the iterative value of a described K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

6. method as claimed in claim 5, is characterized in that, described according to formula determine the iterative value of a described K model parameter, comprising:

According to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector;

Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

7. method as claimed in claim 6, is characterized in that,

Described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or,

Described iteration stopping condition is that iterations reaches preset times.

8. method as claimed in claim 5, is characterized in that, described according to formula determine the iterative value of a described K model parameter, comprising:

According to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

9. the method according to any one of claim 4-8, is characterized in that, the described proper vector based on described N number of image pattern, and the iterative value of a described K model parameter, determines the iterative value of described negative image sample hidden variable, comprising:

According to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

10. the method according to any one of claim 3-9, is characterized in that, the described proper vector based on described N number of image pattern, and the currency of a described K model parameter, determines the currency of described positive image pattern hidden variable, comprising:

According to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

11. methods according to any one of claim 2-10, it is characterized in that, the initial value of the hidden variable of the described N number of image pattern of described acquisition, comprising:

According to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

12. 1 kinds of image classification methods, is characterized in that, comprising:

Obtain the proper vector of image to be classified;

Based on the proper vector of described image to be classified, utilize K sorter, determine the classification of described image to be classified, wherein, a described K sorter is K the sorter utilizing the method according to any one of claim 1 to claim 11 to train;

According to formula determine the probability of described image to be classified under a described K classification, wherein, x represents described image to be classified, β _krepresent the model parameter of a kth sorter in a described K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.

The generating apparatus of 13. 1 kinds of Image Classifiers, is characterized in that, comprising:

First acquiring unit, for obtaining training sample set, described training sample set comprises N number of image pattern, and described N number of image pattern belongs to K classification, and N, K are positive integer, and N is greater than K;

Second acquisition unit, for obtain that described first acquiring unit obtains each described in the proper vector of image pattern, wherein, described proper vector comprises the hidden variable of image pattern;

Training unit, for the hidden variable of described N number of image pattern obtained based on described second acquisition unit, by multivariate logistic regression model, trains the sorter of a described K classification.

14. devices as claimed in claim 13, is characterized in that, the sorter of a described K classification comprises K model parameter respectively, and described training unit is specifically for obtaining the initial value of a described K model parameter; Obtain the initial value of the hidden variable of described N number of image pattern; Based on the proper vector of described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the desired value of a described K model parameter.

15. devices as claimed in claim 14, it is characterized in that, the initial value of described N number of image pattern hidden variable comprises: the positive initial value of image pattern hidden variable and the initial value of negative image sample hidden variable, described training unit is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the currency of a described K model parameter, when the currency of a described K model parameter meets the default condition of convergence, the currency of a described K model parameter is defined as the desired value of a described K model parameter, when the currency of a described K model parameter does not meet the described condition of convergence, based on the proper vector of described N number of image pattern, and the currency of a described K model parameter, determine the currency of described positive image pattern hidden variable, and utilize the currency of described positive image pattern hidden variable to upgrade the initial value of described positive image pattern hidden variable, repeat this step until the currency of a described K model parameter meets the described condition of convergence.

16. devices as claimed in claim 15, it is characterized in that, described training unit is specifically for the proper vector based on described N number of image pattern, and the initial value of described N number of image pattern hidden variable, by described multivariate logistic regression model, train the sorter of a described K classification, to determine the iterative value of a described K model parameter, based on the proper vector of described N number of image pattern, and the iterative value of a described K model parameter, determine the iterative value of described negative image sample hidden variable, and utilize the iterative value of described negative image sample hidden variable to upgrade the initial value of described negative image sample hidden variable, when the iterative value of a described K model parameter meets default iteration stopping condition, the iterative value of a described K model parameter is defined as the currency of a described K model parameter, otherwise, repeat this step until the currency of a described K model parameter meets described iteration stopping condition.

17. devices as claimed in claim 16, is characterized in that, described training unit is specifically for according to formula determine the iterative value of a described K model parameter, wherein, $p_{y_i}(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_{y_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_{y_i})=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_{y_i}\·f(x_i,z),$ X _irepresent the i-th sample in described N number of image pattern, β _lrepresent l model parameter in a described K model parameter, θ represents that the K of described K model parameter composition ties up variable, represent x _imodel parameter corresponding to classification, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector.

18. devices as claimed in claim 17, is characterized in that, described training unit is specifically for according to formula $\▿l\left({\mathrm{\β}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}f(x_i,z_i\left({\mathrm{\β}}_k\right))\·h(x_i,{\mathrm{\β}}_k)-\mathrm{\λ}\·\mathrm{sgn}\left({\mathrm{\β}}_k\right),$ Determine β _kcorresponding gradient, wherein, $p_k(x_i;\mathrm{\θ})=\frac{\exp \left(s(x_i,{\mathrm{\β}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i,{\mathrm{\β}}_l)\right)},$ $s(x_i,{\mathrm{\β}}_k)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_k\·f(x_i,z),$ $s(x_i,{\mathrm{\β}}_l)=\underset{z\∈Z\left(x_i\right)}{\max }{\mathrm{\β}}_l\·f(x_i,z),$ represent that l (θ) is about β _kpartial derivative, β _krepresent the kth model parameter in a described K model parameter, z _i(β _k) expression model parameter is β _ktime x _ithe initial value of hidden variable, f (x _i, z _i(β _k)) represent hidden variable z value z _i(β _k) time x _iproper vector; Based on described β _kcorresponding gradient, with l (θ) for objective function, adopts gradient ascent algorithm, determines described β _kiterative value.

19. devices as claimed in claim 18, is characterized in that, described iteration stopping condition is that the change of described target function value l (θ) is less than predetermined threshold value; Or described iteration stopping condition is that iterations reaches preset times.

20. devices as claimed in claim 17, is characterized in that, described training unit is specifically for according to formula $l_{\mathrm{LC}}\left(\mathrm{\θ}\right)=-\mathrm{\λ}\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\left|{\mathrm{\β}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}[\underset{l=1}{\overset{K}{\mathrm{\Σ}}}{y}_{\mathrm{il}}s(x_i;{\mathrm{\β}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)+\log \left(a_i\right)+1],$ The iterative value of the model parameter of K described in parallel computation, wherein, l _lC(θ) be get the recessed upper bound to the logarithm in l (θ) to be transformed, $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\Σ}}}\exp \left(s(x_i;{\mathrm{\β}}_l)\right)}.$

21. devices according to any one of claim 16-20, is characterized in that, described training unit is specifically for according to formula determine the iterative value of described negative image sample hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, β _trepresent t model parameter in a described K model parameter, and represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ttime x _ithe iterative value of hidden variable, i is the arbitrary integer in 1 to N, and t is the arbitrary integer in 1 to K.

22. devices according to any one of claim 15-21, is characterized in that, described training unit is specifically for according to formula determine the currency of described positive image pattern hidden variable, wherein, x _irepresent the i-th sample in described N number of image pattern, represent x _ithe model parameter that classification is corresponding, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, represent that model parameter is time x _ithe currency of hidden variable, i is the arbitrary integer in 1 to N.

23. devices according to any one of claim 14-22, is characterized in that, described training unit is specifically for according to formula determine the initial value of the hidden variable of image pattern described in each, wherein, x _irepresent the i-th sample in described N number of image pattern, β _krepresent the kth model parameter in a described K model parameter, Z (x _i) represent x _ithe span of hidden variable z, f (x _i, z) represent x _iproper vector, expression model parameter is β _ktime x _ithe initial value of hidden variable z, i is the arbitrary integer in 1 to N, and k is the arbitrary integer in 1 to K.

24. 1 kinds of image classification devices, is characterized in that, comprising:

First acquiring unit, for obtaining the proper vector of image to be classified;

First determining unit, for the proper vector based on described image to be classified, utilize K sorter, determine the classification of described image to be classified, wherein, a described K sorter is K the sorter utilizing the device according to any one of claim 13 to claim 23 to train;

Second determining unit, for according to formula determine the probability of described image to be classified under a described K classification, wherein, x represents described image to be classified, β _krepresent the model parameter of a kth sorter in a described K sorter, f (x, z) represents the proper vector of x, and Z (x) represents the span of the hidden variable z of x, and k is the arbitrary integer in 1 to K.