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

Abstract

The invention discloses a method and device for generating an image classifier. The method includes: obtaining a training sample set, the training sample set includes N image samples, the N image samples belong to K categories, N and K are positive integers, and N is greater than K; the feature vector of each image sample is obtained, wherein the feature vector includes hidden variables of the image sample; based on the hidden variables of N image samples, classifiers of K categories are trained through a multiple logistic regression model. In the embodiment of the present invention, the classifiers of K categories are simultaneously trained in the form of maximum likelihood through the multiple logistic regression model, that is to say, the use of the multiple logistic regression model preserves the correlation between the classifiers of K categories , compared with the way that LVSM converts the K-class classification problem in the field of object classification into multiple isolated two-class problems, the training results are more accurate.

Description

Image classifier generation method, image classification method and device

technical field

The present invention relates to the field of image classification, and more specifically, to a method for generating an image classifier, an image classification method and a device.

Background technique

Hidden variables refer to comprehensive variables that cannot be directly observed but play an important role in practical applications, such as spatial relationships, data structures, and inline states. Latent variables are widely used in fields such as machine vision, natural language processing, speech recognition, and public health. Experiments have proved that when processing objects such as images and voices, the introduction of latent variables can capture more useful information, and the processing effect is significantly improved compared with the method of only using explicit variables.

Most of the early hidden variable models were generative models, such as Hidden Markov Model (HMM), Gaussian Mixture Model (GMM) and so on. Recently, more researchers have attempted to explore the possibility of introducing hidden variables into discriminative models. Typical examples include Conditional Random Field (CRF), Latent Support Vector Machine (LSVM), etc. These models have achieved certain results in their respective fields. It is worth mentioning that LSVM combined with a locally deformable model (DeformablePart-basedModel, DPM), that is, DPM-LSVM, has become a relatively successful algorithm in the field of object detection in machine vision in recent years. DPM is used to describe the characteristics of detection category objects. It consists of three parts: a main filter (rootfilter), multiple local filters (partfilters), and each local corresponding deformation penalty (deformablecosts). The main part is used to describe the general outline of the object, the local part is used to describe the detailed features of the detected object, and the deformation penalty is used to ensure that the position of each part relative to the main body does not have too much deviation. In the process of object detection, the position of the part relative to the subject can change within a certain range, which can be regarded as a hidden variable, and LSVM is used for training.

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

$\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; C</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; max</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0,1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, β is the model parameter of the classifier, y _i represents the label of the training sample _xi , s( _xi , β) represents the score of the sample _xi , and this score is in all possible local relative positions (that is, the hidden variable value range ), which satisfies formula (2):

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; Z</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; max</span>}}\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In formula (2), z is a hidden variable, f is a feature extraction method, and f( _xi ,z) is a feature vector of sample _xi , such as the frame gradient histogram feature used in DPM.

It can be proved that the objective function (formula (1)) of LSVM has semi-concave property, that is, when the hidden variable of the fixed positive sample takes a value, the objective function is concave. Therefore, the solution of LSVM can use coordinate gradient descent (CoordinateGradientDescent), that is, first fix the model parameters of the classifier, obtain the value of the hidden variable of the positive sample, and then fix the value of the hidden variable of the positive sample, and find the optimal model parameter and hidden variable of the negative sample. The variable takes a value, and iterates until convergence.

LSVM, like SVM, is mainly applicable to the field of object detection. When extended to the field of object classification, LSVM's processing method is to transform the multi-class problem in the field of object classification into the second-class problem in the field of object detection. With this approach, the training process of multiple classifiers for object classification will be isolated from each other. In practice, there may be a certain correlation between various object categories. For example, buildings are divided into multiple architectural styles, and the buildings in the picture to be classified may have the characteristics of two or more architectural styles at the same time. Therefore, transforming the training process of multiple classifiers into multiple two-class problems that are isolated from each other will lead to inaccurate classification results.

Contents of the invention

Embodiments of the present invention provide a method and device for generating an image classifier, so as to improve the accuracy of classification results.

In a first aspect, a method for generating an image classifier is provided, including: obtaining a training sample set, the training sample set including N image samples, the N image samples belonging to K categories, N and K are positive integers, N is greater than K; obtain the feature vector of each of the image samples, wherein the feature vector includes hidden variables of the image samples; based on the hidden variables of the N image samples, train the K by multiple logistic regression model The classifier for the category.

With reference to the first aspect, in an implementation manner of the first aspect, the classifiers of the K categories respectively include K model parameters, and the latent variables based on the N image samples are passed through a multiple logistic regression model, Training the classifiers of the K categories includes: obtaining the initial values of the K model parameters; obtaining the initial values of the latent variables of the N image samples; based on the feature vectors of the N image samples, and the The initial values of the latent variables of the N image samples are used to train the classifiers of the K categories through the multiple logistic regression model, so as to determine the target values of the K model parameters.

In combination with the first aspect or any of the above-mentioned implementations, in another implementation of the first aspect, the initial values of the hidden variables of the N image samples include: the initial values of the latent variables of the positive image samples and the negative image The initial value of the hidden variable of the sample, the feature vector based on the N image samples, and the initial value of the latent variable of the N image samples, through the multiple logistic regression model, train the classifier of the K categories , to determine the target values of the K model parameters, including: based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, the multiple logistic regression model is used to train the Classifiers of K categories, to determine the current values of the K model parameters, and when the current values of the K model parameters meet a preset convergence condition, determine the current values of the K model parameters as the current values of the K model parameters. The target values of the K model parameters, when the current values of the K model parameters do not satisfy the convergence condition, based on the feature vectors of the N image samples and the current values of the K model parameters, determine The current value of the hidden variable of the positive image sample, and use the current value of the hidden variable of the positive image sample to update the initial value of the hidden variable of the positive image sample, and repeat this step until the current values of the K model parameters satisfy The convergence condition.

In combination with the first aspect or any of the above implementations, in another implementation of the first aspect, the feature vectors based on the N image samples, and the initial hidden variables of the N image samples value, through the multiple logistic regression model, train the classifiers of the K categories to determine the current values of the K model parameters, including: feature vectors based on the N image samples, and the N The initial value of the latent variable of the image sample, through the multiple logistic regression model, train the classifiers of the K categories to determine the iterative values of the K model parameters, based on the feature vectors of the N image samples, and The iterative value of the K model parameters, determine the iterative value of the hidden variable of the negative image sample, and use the iterative value of the hidden variable of the negative image sample to update the initial value of the hidden variable of the negative image sample, when the K When the iteration values of the K model parameters satisfy the preset iteration stop condition, determine the iteration values of the K model parameters as the current values of the K model parameters, otherwise, repeat this step until the K model parameters The current value of satisfies the iteration stopping condition.

In combination with the first aspect or any of the above implementations, in another implementation of the first aspect, the feature vectors based on the N image samples, and the initial hidden variables of the N image samples Value, through the multiple logistic regression model, train the classifiers of the K categories to determine the iterative values of the K model parameters, including: according to the formula Determine the iterative values of the K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&CenterDot;f(x_i,z),$ x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi .

In combination with the first aspect or any of the above-mentioned implementations, in another implementation of the first aspect, according to the formula Determining the iterative values of the K model parameters includes: according to the formula $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&Center\; Dot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&CenterDot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&CenterDot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to the β _k , l(θ) is the objective function , using the gradient ascending algorithm to determine the iterative value of the β _k .

In combination with the first aspect or any of the above implementations, in another implementation of the first aspect, the iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, The iteration stop condition is that the number of iterations reaches a preset number.

In combination with the first aspect or any of the above-mentioned implementations, in another implementation of the first aspect, according to the formula Determining the iterative values of the K model parameters includes: according to the formula $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed by taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

In combination with the first aspect or any of the above-mentioned implementation manners, in another implementation manner of the first aspect, the feature vectors based on the N image samples and the iteration values of the K model parameters, Determining the iterative value of the hidden variable of the negative image sample includes: according to the formula Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K.

In combination with the first aspect or any of the above-mentioned implementation manners, in another implementation manner of the first aspect, the feature vectors based on the N image samples, and the current values of the K model parameters, Determining the current value of the hidden variable of the positive image sample includes: according to the formula Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N.

In combination with the first aspect or any of the above implementations, in another implementation of the first aspect, the determination of the hidden variable of each image sample based on the initial value of each model parameter Initial values, including: According to the formula Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K.

In a second aspect, an image classification method is provided, including: obtaining a feature vector of an image to be classified; based on the feature vector of the image to be classified, using K classifiers to determine the category of the image to be classified, wherein the The K classifiers are K classifiers trained using the first aspect or any implementation of the first aspect; according to the formula Determining the probability of the image to be classified under the K categories, wherein, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x, z) represents the feature vector of x, and Z(x) represents the hidden variable z of x Value range, k is any integer from 1 to K.

In a third aspect, a device for generating an image classifier is provided, including: a first acquisition unit configured to acquire a training sample set, the training sample set includes N image samples, and the N image samples belong to K categories, N and K are positive integers, and N is greater than K; a second acquisition unit, configured to acquire a feature vector of each of the image samples acquired by the first acquisition unit, wherein the feature vector includes a hidden variable of the image sample; The training unit is configured to train the classifiers of the K categories through a multiple logistic regression model based on the latent variables of the N image samples acquired by the second acquisition unit.

With reference to the second aspect, in an implementation manner of the second aspect, the classifiers of the K categories respectively include K model parameters, and the training unit is specifically used to acquire initial values of the K model parameters; The initial values of the hidden variables of the N image samples; based on the feature vectors of the N image samples and the initial values of the hidden variables of the N image samples, through the multiple logistic regression model, train the K classifier to determine target values for the K model parameters.

With reference to the third aspect, in an implementation manner of the third aspect, the initial values of the hidden variables of the N image samples include: the initial values of the hidden variables of the positive image samples and the initial values of the hidden variables of the negative image samples, and the training The unit is specifically configured to train the classifiers of the K categories through the multiple logistic regression model based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples to determine the The current values of the K model parameters, when the current values of the K model parameters meet the preset convergence conditions, the current values of the K model parameters are determined as the target values of the K model parameters, when the K model parameters are determined as target values, When the current values of the K model parameters do not meet the convergence condition, based on the feature vectors of the N image samples and the current values of the K model parameters, determine the current value of the hidden variable of the positive image sample, and update the initial value of the hidden variable of the positive image sample with the current value of the hidden variable of the positive image sample, and repeat this step until the current values of the K model parameters satisfy the convergence condition.

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to be based on the feature vectors of the N image samples, and the N image samples The initial value of the sample hidden variable, through the multiple logistic regression model, train the classifiers of the K categories to determine the iterative values of the K model parameters, based on the feature vectors of the N image samples, and the The iterative value of the K model parameters, determine the iterative value of the hidden variable of the negative image sample, and update the initial value of the hidden variable of the negative image sample by using the iterative value of the hidden variable of the negative image sample, when the K When the iteration value of the model parameter satisfies the preset iteration stop condition, the iteration value of the K model parameters is determined as the current value of the K model parameters, otherwise, this step is repeated until the K model parameters reach The current value satisfies the iteration stop condition.

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to Determine the iterative values of the K model parameters, where, x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi .

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&Center\; Dot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&CenterDot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&CenterDot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&Center\; Dot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to the β _k , l(θ) is the objective function , using the gradient ascending algorithm to determine the iterative value of the β _k .

In combination with the third aspect or any of the above implementations thereof, in another implementation of the third aspect, the iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, The iteration stop condition is that the number of iterations reaches a preset number.

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed by taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K.

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N.

In combination with the third aspect or any of the above-mentioned implementation manners, in another implementation manner of the third aspect, the training unit is specifically configured to Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K.

In a fourth aspect, an image classification device is provided, comprising: a first acquisition unit, configured to acquire a feature vector of an image to be classified; a first determination unit, configured to use K classifiers based on the feature vector of the image to be classified , to determine the category of the image to be classified, wherein the K classifiers are K classifiers trained by using the third aspect or any implementation of the third aspect; the second determining unit is used for according to the formula Determining the probability of the image to be classified under the K categories, wherein, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x, z) represents the feature vector of x, and Z(x) represents the hidden variable z of x Value range, k is any integer from 1 to K.

In the embodiment of the present invention, K classifiers are simultaneously trained in the form of maximum likelihood through the multiple logistic regression model, that is to say, the use of the multiple logistic regression model retains the correlation between the classifiers of the K categories, and Compared with the method of converting the K-class classification problem in the field of object classification into multiple isolated two-class problems with LVSM, the training results are more accurate.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required in the embodiments of the present invention. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For Those of ordinary skill in the art can also obtain other drawings based on these drawings without making creative efforts.

Fig. 1 is a schematic flowchart of a method for generating an image classifier according to an embodiment of the present invention.

Fig. 2 is an example diagram of classifying images using classifier parameters trained in an embodiment of the present invention.

Fig. 3 is an example diagram of classifying images using classifier parameters trained in an embodiment of the present invention.

Fig. 4 is a schematic structural diagram of a device for generating an image classifier according to an embodiment of the present invention.

Fig. 5 is a schematic structural diagram of a device for generating an image classifier according to an embodiment of the present invention.

Fig. 6 is a schematic flowchart of an image classification method according to an embodiment of the present invention.

Fig. 7 is a schematic block diagram of an image classification device according to an embodiment of the present invention.

Fig. 8 is a schematic block diagram of an image classification device according to an embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Fig. 1 is a schematic flowchart of a method for generating an image classifier according to an embodiment of the present invention. The method of Figure 1 includes:

110. Obtain a training sample set, the training sample set includes N image samples, the N image samples belong to K categories, N and K are positive integers, and N is greater than K.

For example, the training sample set D={(x ₁ ,y ₁ ),...,(x _N ,y _N )} contains N image samples in total, where y _i is the label of the image sample _xi , used for Indicates the category of _xi , which is one of the above K categories.

120. Acquire a feature vector of each image sample, where the feature vector includes hidden variables of the image sample.

It should be understood that image features and latent variables may be selected according to application scenarios or actual needs. For example, image features can be selected (or defined as) Histogram of Oriented Gradient (HOG), local binary pattern (LocalBinaryPatterns, LBP), or Haar, etc.; hidden variables can be selected (or defined as) the position of the object in the image , the relative position between the part and the subject in the image, or the subcategory of the object, etc. Based on the image features and hidden variables selected above, the feature vector of each image sample is obtained. At this time, the feature vector of each image obtained is not a fixed value, but will change with the change of hidden variables. Suppose the image x _i The hidden variable is z, and the extracted feature vector can be represented by f(x, z).

130. Based on hidden variables of N image samples, classifiers of K categories are trained through a multiple logistic regression model.

In the embodiment of the present invention, K classifiers are simultaneously trained in the form of maximum likelihood through the multiple logistic regression model, that is to say, the use of the multiple logistic regression model retains the correlation between the classifiers of the K categories, and Compared with the method of converting the K-class classification problem in the field of object classification into multiple isolated two-class problems with LVSM, the training results are more accurate.

Optionally, as an embodiment, step 130 may include: acquiring initial values of K model parameters; acquiring initial values of latent variables of N image samples; feature vectors based on N image samples, and N image sample hidden variables The initial value of the variable is used to train classifiers of K categories through a multiple logistic regression model to determine the target values of K model parameters.

It should be noted that a hidden variable of an image sample may include K initial values, that is, a hidden variable of an image sample has a corresponding initial value under an initial value of a model parameter. Through step 130, the initial values of N*K hidden variables can be obtained.

Optionally, as an embodiment, the initial values of the hidden variables of the N image samples include: the initial values of the hidden variables of the positive image samples and the initial values of the hidden variables of the negative image samples, the above-mentioned feature vectors based on the N image samples, and The initial value of N image samples, through the multiple logistic regression model, trains classifiers of K categories to determine the target values of K model parameters, which may include: feature vectors based on N image samples, and N image samples The initial value of the hidden variable, through the multiple logistic regression model, trains classifiers of K categories to determine the current values of the K model parameters. When the current values of the K model parameters meet the preset convergence conditions, the K models The current value of the parameter is determined as the target value of the K model parameters. When the current values of the K model parameters do not meet the convergence condition, the positive image is determined based on the eigenvectors of the N image samples and the current values of the K model parameters. The current value of the hidden variable of the sample, and use the current value of the hidden variable of the positive image sample to update the initial value of the hidden variable of the positive image sample, and repeat this step until the current values of the K model parameters meet the convergence condition.

Specifically, the hidden variables of an image sample may have different initial values under different model parameters, that is to say, the hidden variables of an image sample may include K initial values, and the initial values of the hidden variables of the above N image samples may include : K*N initial values. An image sample is a positive sample under the model parameters corresponding to the image sample category, and the initial values of the hidden variables of the above positive image samples include N initial values, which are the initial values of the N image samples under the model parameters corresponding to their respective categories . Among the K*N initial values, the remaining K*(N-1) initial values except the initial value of the hidden variable of the positive image are the initial values of the hidden variable of the negative image sample.

It can be proved that when the initial value of the hidden variable of the positive image sample is fixed, the multiple logistic regression model is concave and can be solved by gradient ascent.

Optionally, as an embodiment, the above-mentioned feature vectors based on the N image samples and the initial values of the latent variables of the N image samples are used to train classifiers of K categories through a multiple logistic regression model to determine K models The current value of the parameter may include: based on the eigenvectors of N image samples and the initial values of the latent variables of N image samples, through the multiple logistic regression model, classifiers of K categories are trained to determine the iterative values of K model parameters , based on the eigenvectors of N image samples and the iterative values of K model parameters, determine the iterative value of the hidden variable of the negative image sample, and use the iterative value of the hidden variable of the negative image sample to update the initial value of the hidden variable of the negative image sample, when When the iteration values of the K model parameters meet the preset iteration stop condition, determine the iteration values of the K model parameters as the current values of the K model parameters, otherwise, repeat this step until the current values of the K model parameters satisfy the iteration stop condition.

In the embodiment of the present invention, in the case of fixing the value of the hidden variable of the positive sample, the purpose of optimizing K model parameters is achieved by continuously updating the value of the hidden variable of the negative sample, and the accuracy of the classification result is further improved.

Optionally, as an embodiment, the aforementioned feature vectors based on the N image samples and the initial values of the latent variables of the N image samples are used to train classifiers of K categories through a multiple logistic regression model to determine K model parameters The iteration values for can include: According to the formula Determine the iterative values of K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&CenterDot;f(x_i,z),$ x _i represents the i-th sample among the N image samples, β _l represents the l-th model parameter among the K model parameters, and θ represents the K-dimensional variable composed of K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi .

Optionally, as an example, the above formula Determine the iterative values of the K model parameters, which may include: according to the formula $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&Center\; Dot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&Center\; Dot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&CenterDot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial value of the hidden variable of _xi when the model parameter is β _k , f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to β _k , with l(θ) as the objective function, gradient ascent is used Algorithm to determine the iterative value of β _k .

Optionally, as an embodiment, the above iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, the iteration stop condition is that the number of iterations reaches a preset number.

Optionally, as an example, the above formula Determine the iterative values of the K model parameters, which may include: according to the formula $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ Calculate the iterative values of K model parameters in parallel, where l _LC (θ) is converted from the concave upper bound of the logarithm in l(θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

The above objective function l(θ) has a logarithmic sum function, therefore, it cannot be decomposed into the form of superposition of K-type sub-problems, and parallel or distributed computing cannot be used to accelerate the optimization process.

In the embodiment of the present invention, the logarithm has concavity (Log-concavity), and the log-concavity upper bound (Log-concavityBound) is used to convert the objective function l(θ) into the form of the sum of K-type subproblems, so that Parallel computing speeds up the convergence of the algorithm.

Specifically, the logarithmic concave upper bound has the form: Using this formula, l(θ) can be transformed into:

${\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; LC</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}\underset{\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; il</span>}}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; K</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; log</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ]</span>$

When the above formula is used as the objective function and the gradient ascent method is used to solve the problem, the gradient of the classifier parameters has the following form: $\frac{{\&PartialD;l}_{\mathrm{LC}}}{{\&PartialD;\mathrm{\&beta;}}_l}=[-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_l\right)+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_l\right))\&CenterDot;({\mathrm{the\; y}}_{\mathrm{il}}-a_i\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right))],$ Among them, the value of the auxiliary parameter a _i is: $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

Optionally, as an embodiment, based on the feature vectors of the N image samples and the iterative values of the K model parameters, determining the iterative value of the hidden variable of the negative image sample may include: according to the formula Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample in the N image samples, β _t represents the t-th model parameter in the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K.

Optionally, as an embodiment, based on the feature vectors of the N image samples and the current values of the K image samples, determining the current value of the hidden variable of the positive image sample may include: according to the formula Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N.

Optionally, as an embodiment, based on the initial value of each model parameter above, determining the initial value of the hidden variable of each image sample may include: according to the formula Determine the initial value of the latent variable of each image sample, where x _i represents the i-th sample among N image samples, β _k represents the k-th model parameter among the K model parameters, and Z( _xi ) represents x _i The value range of the latent variable z, f( _xi , z) represents the feature vector of _xi , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K.

The embodiments of the present invention will be described in detail below in conjunction with specific examples. It should be noted that these examples are only intended to help those skilled in the art better understand the embodiments of the present invention, rather than limit the scope of the embodiments of the present invention.

Example 1:

Input: training sample set {(x ₁ , y ₁ ),…, (x _N , y _N )}, initial values of all hidden variables.

Output: classifier parameters θ, θ={β ₁ ,...,β _K }.

Example 2:

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

Output: Classifier parameters θ.

In specific implementation, the values of the constants numOuterLoop and numInnerLoop have a great relationship with the application scenario. For example, in digit recognition (digitrecognition), due to the large number of samples and small feature dimension, numOuterLoop=50 and numInnerLoop=1 can be set.

In more complex examples, if the number of samples is small and the feature dimension is high, numOuterLoop=5, numInnerLoop=1000 can be set.

The results of image classification by the trained classifier parameters are given below. It should be noted that, in the following description, the classifier training method in the embodiment of the present invention is called: latent variable multiple logistic regression (MultinomialLatentLogisticRegression, MLLR).

Fig. 2 is an example diagram of classifying images using classifier parameters trained in an embodiment of the present invention. In the example in Figure 2, mammals are classified as the research object, and there are 6 types of mammals in total, each with about 50 pictures. In the experiment, 50% of the images are taken as training, and the other 50% of images are used as testing. In terms of image features, HOG features are used. The hidden variable is the position of the object to be detected in the picture, and the size of the frame where the object is located should be more than 30% of the total picture size. Linear SVM, LSVM and MLLR, the test results are as follows:

Table 1 Classification results of mammal classification experiments

Classification Linear SVM LSVM MLLR Accuracy(%) 64.23 69.59 73.31

The test results show that the accuracy of MLLR exceeds that of LSVM, and the effect of the classifiers trained by the two hidden variable methods of LSVM and MLLR is better than that of the traditional linear SVM method.

In Figure 2, the first column is a schematic diagram of a classifier trained by linear SVM (using HOG features), and the second column is a schematic diagram of a classifier trained by MLLR. The rectangular frame in the small picture in Figure 2 is the object position detected by MLLR.

Fig. 3 is an example diagram of classifying images using classifier parameters trained in an embodiment of the present invention. Figure 3 takes sports figures as the research object, including six types of actions (cricket hitting, cricket pitching, volleyball smashing, croquet hitting, tennis forehand and tennis serve). The image feature still uses HOG, and the hidden variable model uses DPM, that is, the object position and the relative position of the local subject are both used as hidden variables. The results show that the classification accuracy of MLLR (78.3%) exceeds that of LSVM (74.4%). In Figure 3, the first column is the schematic diagram of the main body model in the picture, and the second column is the schematic diagram of the local model in the picture. In the small picture in Figure 3, the dark rectangle represents the position of the subject, and the light rectangle represents the local position. It should be understood that in various embodiments of the present invention, the sequence numbers of the above-mentioned processes do not mean the order of execution, and the execution order of each process should be determined by its functions and internal logic, rather than by the embodiment of the present invention. The implementation process constitutes any limitation.

The method for generating an image classifier according to an embodiment of the present invention is described in detail above with reference to FIG. 1 to FIG. 3 , and the apparatus for generating an image classifier according to an embodiment of the present invention will be described below in conjunction with FIGS. 4 to 5 .

It should be understood that the apparatus for generating an image classifier according to the embodiment of the present invention can implement each step in FIG. 1 , and details are not repeated here for brevity.

Fig. 4 is a schematic structural diagram of a device for generating an image classifier according to an embodiment of the present invention. The device 400 of Fig. 4 comprises:

The first acquiring unit 410 is configured to acquire a training sample set, the training sample set includes N image samples, the N image samples belong to K categories, N and K are positive integers, and N is greater than K;

The second acquiring unit 420 is configured to acquire a feature vector of each image sample acquired by the first acquiring unit 410, wherein the feature vector includes a latent variable of the image sample;

The training unit 430 is configured to train the classifiers of the K categories based on the latent variables of the N image samples acquired by the second acquiring unit 420 through a multiple logistic regression model.

In the embodiment of the present invention, K classifiers are simultaneously trained in the form of maximum likelihood through the multiple logistic regression model, that is to say, the use of the multiple logistic regression model retains the correlation between the classifiers of the K categories, and Compared with the method of converting the K-class classification problem in the field of object classification into multiple isolated two-class problems with LVSM, the training results are more accurate.

Optionally, as an embodiment, the classifiers of the K categories respectively include K model parameters, and the training unit 430 is specifically configured to obtain initial values of the K model parameters; based on each of the model parameters The initial value of the initial value, determine the initial value of the hidden variable of each image sample; Based on the feature vector of the N image samples, and the initial value of the hidden variable of the N image samples, through the multiple logistic regression model, Classifiers of the K categories are trained to determine target values of the K model parameters.

Optionally, as an embodiment, the initial values of the hidden variables of the N image samples include: the initial values of the hidden variables of the positive image samples and the initial values of the hidden variables of the negative image samples, and the training unit 430 is specifically configured to The feature vectors of the N image samples, and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the current K model parameters value, when the current values of the K model parameters meet the preset convergence conditions, determine the current values of the K model parameters as the target values of the K model parameters, and when the K model parameters When the current value does not satisfy the convergence condition, based on the eigenvectors of the N image samples and the current values of the K model parameters, determine the current value of the hidden variable of the positive image sample, and use the positive image The current value of the sample latent variable updates the initial value of the positive image sample latent variable, and this step is repeated until the current values of the K model parameters satisfy the convergence condition.

Optionally, as an embodiment, the training unit 430 is specifically configured to train the multivariate logistic regression model based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples. The classifiers of the K categories are used to determine the iteration values of the K model parameters, based on the feature vectors of the N image samples and the iteration values of the K model parameters, determine the negative image sample hidden variable, and use the iterative value of the hidden variable of the negative image sample to update the initial value of the hidden variable of the negative image sample, when the iterative value of the K model parameters satisfies the preset iteration stop condition, the The iteration values of the K model parameters are determined as the current values of the K model parameters, otherwise, this step is repeated until the current values of the K model parameters satisfy the iteration stop condition.

Optionally, as an embodiment, the training unit 430 is specifically configured to Determine the iterative values of the K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&Center\; Dot;f(x_i,z),$ x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi .

Optionally, as an embodiment, the training unit 430 is specifically configured to $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&CenterDot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&CenterDot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&Center\; Dot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to the β _k , l(θ) is the objective function , using the gradient ascending algorithm to determine the iterative value of the β _k .

Optionally, as an embodiment, the iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, the iteration stop condition is that the number of iterations reaches a preset number.

Optionally, as an embodiment, the training unit 430 is specifically configured to $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed by taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

Optionally, as an embodiment, the training unit 430 is specifically configured to Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K.

Optionally, as an embodiment, the training unit 430 is specifically configured to Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N.

Optionally, as an embodiment, the training unit 430 is specifically configured to Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K.

Fig. 5 is a schematic structural diagram of a device for generating an image classifier according to an embodiment of the present invention. The device 500 of Fig. 5 comprises:

memory 510, for storing programs;

The processor 520 is configured to execute the program. When the program is executed, the processor 520 is specifically configured to obtain a training sample set, the training sample set includes N image samples, and the N image samples belong to K categories, N and K are positive integers, and N is greater than K; obtain a feature vector of each of the image samples, wherein the feature vector includes hidden variables of the image samples; based on the hidden variables of the N image samples, Classifiers for the K categories are trained through a multiple logistic regression model.

In the embodiment of the present invention, the multiple logistic regression model is used to simultaneously train K classifiers in the form of maximum likelihood, that is to say, the use of the multiple logistic regression model retains the correlation between the classifiers of the K categories, and Compared with the method of converting the K-class classification problem in the field of object classification into multiple isolated two-class problems with LVSM, the training results are more accurate.

Optionally, as an embodiment, the classifiers of the K categories respectively include K model parameters, and the processor 520 is specifically configured to obtain initial values of the K model parameters; based on each of the model parameters The initial value of the initial value, determine the initial value of the hidden variable of each image sample; Based on the feature vector of the N image samples, and the initial value of the hidden variable of the N image samples, through the multiple logistic regression model, Classifiers of the K categories are trained to determine target values of the K model parameters.

Optionally, as an embodiment, the initial values of the N hidden variables of image samples include: initial values of hidden variables of positive image samples and initial values of hidden variables of negative image samples, and the processor 520 is specifically configured to The feature vectors of the N image samples, and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the current K model parameters value, when the current values of the K model parameters meet the preset convergence conditions, determine the current values of the K model parameters as the target values of the K model parameters, and when the K model parameters When the current value does not satisfy the convergence condition, based on the eigenvectors of the N image samples and the current values of the K model parameters, determine the current value of the hidden variable of the positive image sample, and use the positive image The current value of the sample latent variable updates the initial value of the positive image sample latent variable, and repeats this step until the current values of the K model parameters satisfy the convergence condition.

Optionally, as an embodiment, the processor 520 is specifically configured to train the multivariate logistic regression model based on the feature vectors of the N image samples and the initial values of the hidden variables of the N image samples. The classifiers of the K categories are used to determine the iteration values of the K model parameters, based on the feature vectors of the N image samples and the iteration values of the K model parameters, determine the negative image sample hidden variable, and use the iterative value of the hidden variable of the negative image sample to update the initial value of the hidden variable of the negative image sample, when the iterative value of the K model parameters satisfies the preset iteration stop condition, the The iteration values of the K model parameters are determined as the current values of the K model parameters, otherwise, this step is repeated until the current values of the K model parameters satisfy the iteration stop condition.

Optionally, as an embodiment, the processor 520 is specifically configured to Determine the iterative values of the K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&Center\; Dot;f(x_i,z),$ x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi .

Optionally, as an embodiment, the processor 520 is specifically configured to $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&CenterDot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&Center\; Dot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&CenterDot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to the β _k , l(θ) is the objective function , using the gradient ascending algorithm to determine the iterative value of the β _k .

Optionally, as an embodiment, the iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, the iteration stop condition is that the number of iterations reaches a preset number.

Optionally, as an embodiment, the processor 520 is specifically configured to $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed by taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$

Optionally, as an embodiment, the processor 520 is specifically configured to Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K.

Optionally, as an embodiment, the processor 520 is specifically configured to Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N.

Optionally, as an embodiment, the processor 520 is specifically configured to Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K.

Fig. 6 is a schematic flowchart of an image classification method according to an embodiment of the present invention. In the method of Fig. 6, the K classifiers trained by the method of Fig. 1 can be used to classify images, and the method of Fig. 6 includes:

610. Obtain the feature vector of the image to be classified;

620. Based on the feature vector of the image to be classified, use K classifiers to determine the category of the image to be classified;

630. According to the formula Determine the probability of the image to be classified under K categories, where, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x,z) represents the feature vector of x, Z(x) represents the value range of the latent variable z of x, k is any integer from 1 to K.

The classification result of the existing LSVM only gives which category the image to be classified belongs to, but in actual situations, there may be a certain relationship between different types, and a certain image does not absolutely belong to which category. For example, the style of the building can be classified, including modern style, medieval style, etc. A certain building style in the image may adopt some modern styles and part of the medieval style. At this time, the classification results of the existing LSVM It only shows which architectural style the buildings in the image to be classified belong to, which is obviously not accurate enough. In this embodiment, in addition to the category to which the image to be classified belongs, the probability of the picture in each category is also given. Compared with the prior art, the introduction of the probability interpretation of the image classification result makes the description of the image classification result more accurate .

Fig. 7 is a schematic block diagram of an image classification device according to an embodiment of the present invention. The device 700 in FIG. 7 can use the K classifiers trained by the device 400 of FIG. 4 to classify images, and the device 700 includes:

A first acquiring unit 710, configured to acquire a feature vector of an image to be classified;

The first determination unit 720 is configured to determine the category of the image to be classified by using K classifiers based on the feature vector of the image to be classified;

The second determining unit 730 is used for according to the formula Determine the probability of the image to be classified under K categories, where, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x,z) represents the feature vector of x, Z(x) represents the value range of the latent variable z of x, k is any integer from 1 to K.

The classification result of the existing LSVM only gives which category the image to be classified belongs to, but in actual situations, there may be a certain relationship between different types, and a certain image does not absolutely belong to which category. For example, the style of the building can be classified, including modern style, medieval style, etc. A certain building style in the image may adopt some modern styles and part of the medieval style. At this time, the classification results of the existing LSVM It only shows which architectural style the buildings in the image to be classified belong to, which is obviously not accurate enough. In this embodiment, in addition to the category to which the image to be classified belongs, the probability of the picture in each category is also given. Compared with the prior art, the introduction of the probability interpretation of the image classification result makes the description of the image classification result more accurate .

Fig. 8 is a schematic block diagram of an image classification device according to an embodiment of the present invention. The image classification device 800 in Fig. 8 can utilize the K classifiers trained by the device 500 in Fig. 5 to classify the images, and the method in Fig. 8 includes:

memory 810, for storing programs;

The processor 820 is used to execute a program. When the program is executed, the program is used to obtain the feature vector of the image to be classified; based on the feature vector of the image to be classified, K classifiers are used to determine the category of the image to be classified ;according to the formula Determine the probability of the image to be classified under K categories, where, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x,z) represents the feature vector of x, Z(x) represents the value range of the latent variable z of x, k is any integer from 1 to K.

The classification result of the existing LSVM only gives which category the image to be classified belongs to, but in actual situations, there may be a certain relationship between different types, and a certain image does not absolutely belong to which category. For example, the style of the building can be classified, including modern style, medieval style, etc. A certain building style in the image may adopt some modern styles and part of the medieval style. At this time, the classification results of the existing LSVM It only shows which architectural style the buildings in the image to be classified belong to, which is obviously not accurate enough. In this embodiment, in addition to the category to which the image to be classified belongs, the probability of the picture in each category is also given. Compared with the prior art, the introduction of the probability interpretation of the image classification result makes the description of the image classification result more accurate .

It should be understood that in the embodiments of the present invention, the term "and/or" is only an association relationship describing associated objects, indicating that there may be three relationships. For example, A and/or B may mean that A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character "/" in this article generally indicates that the contextual objects are an "or" relationship.

Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the relationship between hardware and software Interchangeability. In the above description, the composition and steps of each example have been generally described according to their functions. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present invention.

Those skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices or units, and may also be electrical, mechanical or other forms of connection.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment of the present invention.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium In, several instructions are included to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-OnlyMemory), random access memory (RAM, RandomAccessMemory), magnetic disk or optical disk, and various media that can store program codes.

The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of various equivalents within the technical scope disclosed in the present invention. Modifications or replacements shall all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (24)

1. A method for generating an image classifier, comprising: Obtain a training sample set, the training sample set includes N image samples, the N image samples belong to K categories, N and K are positive integers, and N is greater than K; Obtaining a feature vector of each of the image samples, wherein the feature vector includes hidden variables of the image sample; Based on the latent variables of the N image samples, the classifiers of the K categories are trained through a multiple logistic regression model. 2. The method according to claim 1, wherein the classifiers of the K categories comprise K model parameters respectively, The hidden variables based on the N image samples are trained by multiple logistic regression models to classifiers of the K categories, including: Obtain the initial values of the K model parameters; Acquiring initial values of hidden variables of the N image samples; Based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the K model parameters target value. 3. The method according to claim 2, wherein the initial values of the N image sample hidden variables comprise: the initial values of the positive image sample hidden variables and the initial values of the negative image sample hidden variables, The feature vectors based on the N image samples and the initial values of the latent variables of the N image samples are used to train the classifiers of the K categories through the multiple logistic regression model to determine the K Target values for model parameters, including: Based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the K model parameters the current value of When the current values of the K model parameters satisfy a preset convergence condition, determining the current values of the K model parameters as the target values of the K model parameters, When the current values of the K model parameters do not satisfy the convergence condition, based on the feature vectors of the N image samples and the current values of the K model parameters, determine the current values of the hidden variables of the positive image samples value, and use the current value of the positive image sample latent variable to update the initial value of the positive image sample latent variable, and repeat this step until the current values of the K model parameters meet the convergence condition. 4. the method for claim 3, is characterized in that, described feature vector based on described N image samples, and the initial value of described N image sample latent variables, by described multiple logistic regression model, training The classifiers of the K categories, to determine the current values of the K model parameters, include: Based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the K model parameters the iteration value of Based on the eigenvectors of the N image samples and the iterative values of the K model parameters, determine the iterative value of the hidden variable of the negative image sample, and use the iterative value of the hidden variable of the negative image sample to update the negative The initial value of the hidden variable of the image sample, When the iteration values of the K model parameters meet the preset iteration stop condition, determining the iteration values of the K model parameters as the current values of the K model parameters, Otherwise, repeat this step until the current values of the K model parameters satisfy the iteration stop condition. 5. the method for claim 4, is characterized in that, described feature vector based on described N image samples, and the initial value of described N image sample latent variables, by described multiple logistic regression model, training The classifiers of the K categories, to determine the iteration values of the K model parameters, include: According to the formula Determine the iterative values of the K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&CenterDot;f(x_i,z),$ x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi . 6. The method according to claim 5, wherein the formula Determine the iterative values of the K model parameters, including: According to the formula $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&Center\; Dot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&CenterDot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&Center\; Dot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f(x _i , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); Based on the gradient corresponding to the β _k , with l(θ) as the objective function, a gradient ascending algorithm is used to determine the iterative value of the β _k . 7. The method of claim 6, wherein, The iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, The iteration stop condition is that the number of iterations reaches a preset number. 8. The method according to claim 5, wherein the formula Determine the iterative values of the K model parameters, including: According to the formula $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed by taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$ 9. The method according to any one of claims 4-8, wherein the feature vectors based on the N image samples and the iteration values of the K model parameters determine the negative image Iterative values of sample hidden variables, including: According to the formula Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K. 10. The method according to any one of claims 3-9, wherein the positive image is determined based on the feature vectors of the N image samples and the current values of the K model parameters Current values of sample latent variables, including: According to the formula Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N. 11. The method according to any one of claims 2-10, wherein said obtaining the initial values of hidden variables of said N image samples comprises: According to the formula Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K. 12. An image classification method, characterized in that, comprising: Obtain the feature vector of the image to be classified; Based on the feature vector of the image to be classified, using K classifiers to determine the category of the image to be classified, wherein the K classifiers use the method described in any one of claims 1 to 11 K classifiers trained; According to the formula Determining the probability of the image to be classified under the K categories, wherein, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x, z) represents the feature vector of x, and Z(x) represents the hidden variable z of x Value range, k is any integer from 1 to K. 13. A generating device for an image classifier, comprising: The first acquisition unit is used to acquire a training sample set, the training sample set includes N image samples, the N image samples belong to K categories, N and K are positive integers, and N is greater than K; A second acquiring unit, configured to acquire a feature vector of each image sample acquired by the first acquiring unit, wherein the feature vector includes a latent variable of the image sample; The training unit is configured to train the classifiers of the K categories through a multiple logistic regression model based on the latent variables of the N image samples acquired by the second acquisition unit. 14. The device according to claim 13, wherein the classifiers of the K categories respectively include K model parameters, and the training unit is specifically used to obtain initial values of the K model parameters; The initial values of the hidden variables of the N image samples; based on the feature vectors of the N image samples and the initial values of the hidden variables of the N image samples, the K categories are trained through the multiple logistic regression model classifier to determine the target values for the K model parameters. 15. The device according to claim 14, wherein the initial values of the N image sample hidden variables include: the initial values of the positive image sample hidden variables and the initial values of the negative image sample hidden variables, and the training unit Specifically, based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, through the multiple logistic regression model, train the classifiers of the K categories to determine the K The current values of the K model parameters, when the current values of the K model parameters meet the preset convergence conditions, the current values of the K model parameters are determined as the target values of the K model parameters, when the When the current values of the K model parameters do not meet the convergence condition, based on the feature vectors of the N image samples and the current values of the K model parameters, determine the current value of the hidden variable of the positive image sample, and Using the current value of the hidden variable of the positive image sample to update the initial value of the hidden variable of the positive image sample, and repeat this step until the current values of the K model parameters meet the convergence condition. 16. The device according to claim 15, wherein the training unit is specifically configured to, based on the feature vectors of the N image samples and the initial values of the latent variables of the N image samples, through the multivariate Logistic regression model, training the classifiers of the K categories to determine the iterative values of the K model parameters, based on the feature vectors of the N image samples and the iterative values of the K model parameters, determine the The iterative value of the negative image sample hidden variable, and use the iterative value of the negative image sample hidden variable to update the initial value of the negative image sample hidden variable, when the iterative values of the K model parameters meet the preset iteration stop condition, determine the iteration values of the K model parameters as the current values of the K model parameters, otherwise, repeat this step until the current values of the K model parameters meet the iteration stop condition. 17. The device according to claim 16, characterized in that the training unit is specifically used according to the formula Determine the iterative values of the K model parameters, where, $p_{{\mathrm{the\; y}}_i}(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i})=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_{{\mathrm{the\; y}}_i}\&Center\; Dot;f(x_i,z),$ x _i represents the i-th sample in the N image samples, β _l represents the l-th model parameter in the K model parameters, and θ represents the K-dimensional variable composed of the K model parameters, Indicates the model parameters corresponding to the category of _xi , Z( _xi ) indicates the value range of hidden variable z of _xi , and f( _xi ,z) indicates the feature vector of _xi . 18. The device according to claim 17, wherein the training unit is specifically configured to $\&dtri;l\left({\mathrm{\&beta;}}_k\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}f(x_i,z_i\left({\mathrm{\&beta;}}_k\right))\&CenterDot;h(x_i,{\mathrm{\&beta;}}_k)-\mathrm{\&lambda;}\&Center\; Dot;\mathrm{sgn}\left({\mathrm{\&beta;}}_k\right),$ Determine the gradient corresponding to β _k , where, $p_k(x_i;\mathrm{\&theta;})=\frac{\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)\right)}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)\right)},$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_k)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_k\&Center\; Dot;f(x_i,z),$ $\mathrm{the\; s}(x_i,{\mathrm{\&beta;}}_l)=\underset{z\&Element;Z\left(x_i\right)}{\max }{\mathrm{\&beta;}}_l\&CenterDot;f(x_i,z),$ Represents the partial derivative function of l(θ) with respect to β _k , β _k represents the kth model parameter among the K model parameters, z _i (β _k ) represents the initial hidden variable of _xi when the model parameter is β _k value, f( _xi , z _i (β _k )) represents the eigenvector of x _i when the hidden variable z takes the value z _i (β _k ); based on the gradient corresponding to the β _k , l(θ) is the objective function , using the gradient ascending algorithm to determine the iterative value of the β _k . 19. The device according to claim 18, wherein the iteration stop condition is that the change of the objective function value l(θ) is less than a preset threshold; or, the iteration stop condition is that the number of iterations reaches a preset frequency. 20. The device according to claim 17, characterized in that the training unit is specifically used according to the formula $l_{\mathrm{LC}}\left(\mathrm{\&theta;}\right)=-\mathrm{\&lambda;}\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|+\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}[\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{the\; y}}_{\mathrm{il}}\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)-a_i\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)+\log \left(a_i\right)+1],$ The iterative values of the K model parameters are calculated in parallel, wherein l _LC (θ) is transformed from taking the concave upper bound of the logarithm in l (θ), $a_i=\frac{1}{\underset{l=1}{\overset{K}{\mathrm{\&Sigma;}}}\exp \left(\mathrm{the\; s}(x_i;{\mathrm{\&beta;}}_l)\right)}.$ 21. The device according to any one of claims 16-20, wherein the training unit is specifically configured to Determine the iterative value of the hidden variable of the negative image sample, where x _i represents the i-th sample among the N image samples, and βt represents the _t -th model parameter among the K model parameters, and Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates the iterative value of hidden variable x _i when the model parameter is β _t , i is any integer from 1 to N, and t is any integer from 1 to K. 22. The device according to any one of claims 15-21, wherein the training unit is specifically configured to Determine the current value of the hidden variable of the positive image sample, where x _i represents the i-th sample in the N image samples, Indicates the model parameters corresponding to the _xi category, Z( _xi ) indicates the value range of the hidden variable z of _xi , f( _xi ,z) indicates the feature vector of _xi , Indicates that the model parameters are When is the current value of hidden variable x _i , i is any integer from 1 to N. 23. The device according to any one of claims 14-22, wherein the training unit is specifically configured to Determine the initial value of the hidden variable of each of the image samples, wherein x _i represents the i-th sample in the N image samples, β _k represents the k-th model parameter in the K model parameters, and Z( x _i ) represents the value range of the latent variable z of x _i , f( _xi , z) represents the feature vector of x _i , Indicates the initial value of hidden variable z of x _i when the model parameter is β _k , i is any integer from 1 to N, and k is any integer from 1 to K. 24. An image classification device, comprising: The first acquisition unit is used to acquire the feature vector of the image to be classified; The first determination unit is configured to determine the category of the image to be classified by using K classifiers based on the feature vector of the image to be classified, wherein the K classifiers are used in claims 13 to 23 K classifiers trained by the device described in any one; The second determination unit is used according to the formula Determining the probability of the image to be classified under the K categories, wherein, x represents the image to be classified, β _k represents the model parameters of the kth classifier among the K classifiers, f(x, z) represents the feature vector of x, and Z(x) represents the hidden variable z of x Value range, k is any integer from 1 to K.