CN107463953A - Image classification method and system based on quality insertion in the case of label is noisy - Google Patents

Abstract

The present invention provides an image classification method and system based on quality embedding in the case of noisy tags, including: a step of collecting network image tags; a tag quality factor embedding step: introducing a tag quality factor into a supervised image classification model for use in Control the prediction value generation of noisy labels and absorb the error return information from wrong labels; use the maximized logarithmic likelihood function to design the optimization objective function after adding the label quality factor; network model construction steps: use deep neural network to optimize Modeling of the objective function; network parameter training step: input training pictures and noisy labels into the network model, use a variant of stochastic gradient descent method to end-to-end linkage training model, and update model parameters at the same time; image classification step. The invention uniformly models the three variables of the real label of the picture, the label provided by the user and the quality of the picture label to form the supervised learning of the noisy label, and can obtain relatively accurate image classification results.

Description

Method and system for image classification based on quality embedding in the case of noisy labels

technical field

The present invention relates to the fields of computer vision and data mining, and in particular, relates to a method and system for learning a picture label when the label contains noise.

Background technique

Image recognition is a basic and important task in the field of artificial intelligence, and its application spans many fields such as natural science, medicine, and industry. With the rapid development of deep learning, image classifiers trained by convolutional neural networks have achieved unprecedented success. However, image classification learning under the deep learning framework relies on large-scale high-quality training data, including clear images and accurate labels. Such training data often comes from manual collection and labeling, which will consume a lot of manpower and material resources, making it relatively expensive and inefficient to deal with image recognition problems in new fields.

Due to the rapid development of network technology, social media and people's love for online self-media, there are countless picture data in the Internet. Image social platforms such as Flickr and NetEase LOFTER have image data and label information provided by nearly millions of users. If these pictures and label data can be used for the training of deep neural network models, the type and quantity of data sets will be greatly improved, and it will help deep neural networks to migrate to image recognition problems in different fields more quickly.

Using pictures and labels uploaded by Internet users as training data can solve the limitations of manual labeling data, but it will also bring corresponding problems and challenges. The image data provided by large manually labeled datasets is of good quality and complete with labels, so the neural network classifier model trained based on such datasets has a high accuracy rate. In contrast, web images and user tags have the characteristics of poor quality and inaccurate labels. If the image label data with a lot of noise is used, the prediction reliability of the model will be greatly reduced. Therefore, research on how to make full use of the inexhaustible data resources of network images and user-provided labels for effective image label learning has received more attention.

Traditional approaches to image label learning using noisy labels include designing robust loss functions, statistical queries, simulating noise characteristics, and so on. Some of these methods require a part of clean label data to assist training image classifiers; others try to build models to simulate the distribution of data noise, the noise caused by the difference between the real label in the picture and the label provided by the user, but does not consider the quality of the picture Good or bad, as well as the accuracy of the labels provided by users, the effect of classification and recognition is not up to expectations.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art, and provide an image classification method and system based on quality embedding in the case of noisy labels, so as to solve the problem of using noisy label pictures to train image classifiers in the prior art. The quality of the image tag itself is a problem.

According to one aspect of the present invention, there is provided an image classification method based on quality embedding in the case of label noise, comprising:

Network picture label collection steps: obtain a large number of pictures and label information provided by users from the network picture sharing platform, filter and organize according to the required types, so as to be used for the training of image classifiers;

Label quality factor embedding step: Introduce the label quality factor in the supervised image classification model, which is used to control the prediction value generation of noisy labels and absorb the error feedback information from wrong labels; using the maximized logarithmic likelihood function, design The optimization objective function after adding the label quality factor;

Network model construction steps: use the deep neural network to model the optimization objective function to obtain the overall network model, which includes four sub-models, namely the encoding model, sampling model, decoding model and classification model;

Network parameter training step: input the training pictures and noisy labels obtained in the network image label collection step into the overall network model obtained in the network model construction step, and use the variant stochastic gradient descent method to end-to-end linkage training of the above four sub-models, and update them at the same time Model parameters of the four sub-models;

Image classification step: For a new picture that requires classification, input it to the trained classification model to obtain the prediction of the real label of the picture.

Preferably, the network picture label collection step uses web crawler technology to collect a large number of pictures and labels marked by users on picture social networking sites, and filters and organizes labels and pictures according to the required types, such as keeping Images containing one or more labels from the total m categories.

Preferably, in the step of embedding the tag quality factor, in the existing supervised image classification model, the embedding of the image tag quality factor is added, so that the new optimization objective function is:

where x _m and y _m are the pixel set of the mth image and the noise label provided by the corresponding user, respectively, z _m and s _m are the hidden variables representing the real label and label quality of the image, respectively, and M represents the total number of images used for training. The new optimization objective function can absorb the adverse effects caused by wrong labels in the training data set due to the addition of the label quality factor. At the same time, because the gradient function of the objective function is difficult to calculate, it first turns to optimize its evidence lower bound (ELBO), and at the same time uses the re-parameter technique to simplify the computing resources required for training, and obtains the final optimized objective function formula combination.

Preferably, in the network model building step, a deep neural network is used to model the final optimized objective function formula combination to obtain an overall network model, which includes four sub-models: an encoding model, a sampling model, a decoding model and a classification model;

Wherein, the encoding model adopts a convolutional neural network for generating noise labels from image content X The prior prediction of Combined with the noise label y, the label quality distribution q(S|X,Y) and the real label distribution q(Z|X,Y) are predicted.

Wherein, the sampling model is used to map the label quality distribution q(S|X,Y) generated by the encoding model and the real label distribution q(Z|X,Y) to exact values S and Z.

Wherein, the decoding model adopts a neural network whose input is the output label quality S and the real label Z of the sampling model, and is used to generate the posterior prediction q(Y|Z,S) of the noise label.

Wherein, the method used in the classification model is a convolutional neural network, which uses pictures to generate a prediction of the real label Z.

Preferably, the network parameter training step uses the noise label posterior prediction q(Y|Z,S) restored by the decoding model to perform supervised model training, calculates the return gradient of the encoding model, sampling model, and decoding model, and updates The parameters of these three sub-models, at the same time, use the real label distribution q(Z|X,Y) obtained in the encoding model to perform supervised model training on the classification model, calculate the neural network return gradient, and update the parameters of the classification model.

Preferably, in the image classification step, the pictures required for image classification are input into the trained classification model to obtain the prediction of the real label of the image, and at the same time generate the classification result of the image.

According to a second aspect of the present invention, an image classification system based on quality embedding in the case of noisy labels is provided, including:

Network picture label collection module: obtain a large number of pictures and label information provided by users from the network picture sharing platform, and filter and organize according to the required types;

Label quality factor embedding module: Introduce the label quality factor in the traditional supervised image classification model to control the prediction value generation of noisy labels and absorb the error return information from the wrong label, and calculate the logarithmic likelihood function corresponding to the image classification model as the optimized objective function for training;

Network model construction module: used to utilize the deep neural network to model the optimization objective function, and obtain four sub-models of encoding model, sampling model, decoding model and classification model respectively;

Network parameter training module: input training pictures and noisy labels into the overall network model, use a variant of stochastic gradient descent method to end-to-end linkage training four sub-models, and update the parameters of these four sub-models at the same time;

New image classification task processing module: For new pictures that require classification, input them to the trained classification model to obtain the prediction of the real label of the picture.

Preferably, the label quality factor embedding module, in the existing supervised image classification model, adds the embedding of the picture label quality factor, so that the new optimization objective function is:

Among them, x _m and y _m are the pixel set of the mth picture and the noise label provided by the corresponding user, z _m and s _m are the hidden variables representing the real label and label quality of the picture respectively, and M represents the total number of pictures used for training;

Due to the addition of the label quality factor, the new optimization objective function can absorb the adverse effects caused by wrong labels in the training data set. At the same time, the gradient function of the new optimization objective function is difficult to calculate, so first turn to optimize its evidence lower bound, At the same time, the re-parameter technique is used to simplify the computing resources required for training, and the final optimized objective function formula combination is obtained.

Preferably, the network model building module uses a deep neural network to model the final optimized objective function formula combination respectively to obtain four models: encoding model, sampling model, decoding model and classification model; wherein:

The encoding model, using a convolutional neural network, is used to generate noise labels from image content X The prior prediction of Combined with the noise label y to predict the label quality distribution q(S|X,Y) and the real label distribution q(Z|X,Y);

The sampling model is used to map the label quality distribution q(S|X,Y) generated by the coding model and the real label distribution q(Z|X,Y) to exact values S and Z;

The decoding model adopts a neural network whose input is the output label quality S and the real label Z of the sampling model, and is used to generate a posteriori prediction q(Y|Z,S) of the noise label;

The classification model uses a convolutional neural network, which uses pictures to generate predictions for the true label Z.

Preferably, the network parameter training module uses the noise label posterior prediction q(Y|Z,S) restored by the decoding model to perform supervised model training, calculates the return gradient of the encoding model, sampling model, and decoding model, and updates The parameters of these three sub-models, at the same time, use the real label distribution q(Z|X,Y) obtained in the encoding model to perform supervised model training on the classification model, calculate the neural network return gradient, and update the parameters of the classification model.

Preferably, the network image tag collection module uses web crawler technology to collect a large number of required images and tags marked by users on image social networking sites.

The present invention unifies the modeling of the three variables of the real label of the picture, the label provided by the user, and the quality of the picture label. When training the classifier, it not only predicts the real label of each input picture data, but also infers the quality of the picture label uploaded by the user, and then Form the supervised learning of noisy labels, iterate continuously until the training converges, and obtain the required image classifier for the classification task of new images.

Compared with the prior art, the present invention has the following beneficial effects:

The present invention improves the existing classifier learning mode using noisy labels by digging deep into the picture label data on social media and embedding hidden variables representing label quality in the image classification model. By redesigning the error return gradient formula and constructing the corresponding neural network model, the real label, image label quality and the noisy label provided by the user are simultaneously predicted, so as to effectively absorb the error return in the neural network training caused by the label noise. information, which is conducive to the correct learning of image classifiers.

Using the present invention will help to use a large number of low-cost and available image label data on social media for the training of image classifiers, thereby effectively saving the manpower and material resources required for professional labeling, while avoiding the impact of label noise on classifier training , to obtain more accurate image classification results.

Description of drawings

Other characteristics, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

Fig. 1 is a flow chart of the method in an embodiment of the present invention;

Fig. 2 introduces the supervised image classification model of label quality factor according to an embodiment of the present invention;

Fig. 3 is a structural diagram of each module of the deep neural network model used in an embodiment of the present invention;

Fig. 4 is a system block diagram in an embodiment of the present invention.

detailed description

The present invention will be described in detail below in conjunction with specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

The present invention comprehensively considers the three variables of the real label of the picture, the label provided by the user, and the quality of the picture label, and proposes an image classification technology based on quality embedding when the label is noisy. Divided according to the realization of the overall technology, it is mainly divided into four parts:

(1) Collection of network image tags;

(2) Label quality factor embedding;

(3) Network model construction and parameter training;

(4) New image classification task processing.

The above four parts constitute the image classification method and system of the present invention. In order to better understand the present invention, the implementation of the method and system of the present invention will be introduced below in combination with embodiments.

As shown in Figure 1, it is a flowchart of the classification method provided by this embodiment, wherein:

(1) Collection of network image tags;

Based on the data set YFCC100M released by the photo sharing website Flickr, the pictures with the required classification labels are screened, and the web crawler technology is used to collect, download and organize, and a total of M pictures are obtained.

(2) Label quality factor embedding;

As shown in Figure 2, the label quality factor S is introduced into the existing supervised image classification model, and the relationship between the label quality factor and other variables is established. According to the theory of graphical models, the log likelihood function lnP(Y|X) is rewritten as:

Where X represents the image content set, Y represents the noise label set provided by the user, x _m and y _m are the content of the image m and the noise label provided by the corresponding user, respectively, z _m and s _m represent the real label and label quality of the picture respectively Hidden variable, E stands for expectation.

In this step, the unified modeling of label quality, real labels, and user-provided labels will help to correctly understand the relationship between noise labels and real labels, reduce the impact of noise labels on the training of image classifiers, and improve performance in the case of label noise. Next training to get the accuracy of the image classifier. The objective function to be optimized is calculated by the log-likelihood function:

1. According to the idea of variational inference, and using Jensen's inequality (Jensen's inequality) to calculate the evidence lower bound (ELBO) of the objective function to be optimized, there are three items, namely

as well as This simplifies the amount of computation required to train the classifier;

where x _m and y _m are the content of image m and the noise label provided by the corresponding user, respectively, z _m and s _m are the hidden variables representing the real label and label quality of the image respectively, D _KL [·||·] represents the relative entropy, E is for expectation.

2. Calculate the explicit expression of two relative entropy expressions, assuming that the real label variable obeys the K-dimensional polynomial distribution of q(z _m |x _m ,y _m ) and P(z _m |x _m ), the label quality variable probability P(s _m ), obeys the multivariate Gaussian distribution with mean value μ(x _m ,y _m ) and covariance diag(σ(x _m ,y _m ));

3. Using the re-parameter technique to construct the mapping of the real label z _m and label quality s _m to the supplementary random variables γ _m and ζ _m can solve the problem of excessive sampling variance caused by the traditional Monte Carlo sampling method and improve the performance of the training classifier. stability.

Utilize the (Gumbel-Normalized Exponential Function, Gumbel-SoftMax Function) function to construct the mapping from the real label z _m to the Gumbel variable γ _m , z _m = g(γ _m );

Construct the mapping of label quality s _m to standard normal variable ζ _m ～N(0,1), s _m = μ(x _m ,y _m )+σ ² (x _m ,y _m )eζ _m ,

where μ(x _m , y _m ) is the mean of the label quality s _m , and σ ² (x _m , y _m ) is the variance of the label quality s _m .

Accordingly, the expected term in the calculation of the evidence lower bound An explicit expression of , using this as the objective function to be optimized.

(3) Network model construction and parameter training;

1. As shown in Figure 3, according to the evidence lower bound expression of the objective function to be optimized, the neural network is used for overall modeling, and the obtained network model includes four sub-models, which are divided into four sub-models: coding model, sampling model, Decoding model, classification model. When a picture and its noisy label are input into the overall network model, the overall network model first starts forward propagation:

a) The encoding model makes predictions on the distribution q(z|x,y) and q(s|x,y) of the true label z and label quality s based on the image content x and the noisy label y.

b) Sampling model According to the probability distribution q(z|x,y) and q(s|x,y) given by the encoding model, the specific values of the real label z and the label quality s are sampled.

c) The decoding model inputs the specific values of the real label z and label quality s obtained by the sampling model into the neural network to obtain the prediction of the noise label, and then calculates the cross-entropy loss with the noisy label given by the picture.

d) Independently train a neural network classification model to predict the real label P(z) of the picture, and calculate the relative entropy between the real label z distribution q(z|x,y) given by the encoding model, thereby coupling The remaining three sub-models above.

Wherein, the encoding model uses a convolutional neural network to generate a priori prediction P(y) of the label from the image content x, and combines the noise label y to the label quality distribution q(s|x,y) and the real label Distribution q(z|x,y) for prediction. Among them, the distribution of generated label quality is implemented by the contrast layer, and the distribution of generated real labels is implemented by the addition layer.

Wherein, the sampling model is used to map the label quality distribution q(s|x,y) generated by the encoding model and the real label distribution q(z|x,y) to exact values, including the label quality s and the real label z . The re-referencing technique is adopted in the sampling model to reduce the variance of the sampling results and make the model training more stable.

Wherein, the decoding model adopts a neural network whose input is the output label quality s and the real label z of the sampling model, and is used to restore the prediction of the noise label by the noise layer

Wherein, the neural network classification model adopts a convolutional neural network, its input is the picture content x, and the output is the prediction P(z) of the real label of the picture, and the real label obtained during the operation of the encoding model and the decoding model is used The distribution q(z|x,y) is calculated to obtain the relative entropy, and supervised training is performed to obtain the required image classifier.

2. From the output of the decoding model, that is, the noisy label prediction Provide the noisy label y with the user, input the loss layer together to calculate the loss, and use the stochastic gradient descent method to update the network parameters of each sub-model.

3. After multiple rounds of feedback iterations, until the neural network converges, the training is completed.

(4) New image classification task processing.

Using the neural network image classification model trained in (3), when there is a new unlabeled picture that needs to be classified, input it to the trained classification model to obtain the prediction of the real label of the picture.

As shown in Figure 4, in another embodiment, corresponding to the above method, an embodiment of an image classification system based on quality embedding in the case of label noise, including:

Network picture label collection module: obtain a large number of pictures and label information provided by users from the network picture sharing platform, and filter and organize according to the required types;

Label quality factor embedding module: used to introduce the label quality factor in the traditional supervised image classification model to control the generation of the predicted value of the noisy label and absorb the error return information from the wrong label, and calculate the overall network model (after embedding the quality factor The corresponding logarithmic likelihood function of the image classification model) is used as the optimization objective function of training;

Network model building block: it is used to use the deep neural network to model the optimization objective function, and obtain four sub-models of encoding model, sampling model, decoding model and classification model respectively;

Network parameter training module: input the training pictures and noisy labels into the overall network model, use the variant stochastic gradient descent method to train the four sub-models end-to-end linkage, and update the parameters of the four sub-models at the same time;

New image classification task processing module: For a new picture that requires classification, input it to the trained classification model to obtain the prediction of the real label of the picture.

The implementation technical features of the specific modules of the image classification system based on quality embedding in the case of noisy tags above correspond to the steps of the image classification method based on quality embedding in the case of noisy tags.

It should be noted that the steps in the method provided by the present invention can be realized by using the corresponding modules, devices, units, etc. in the system, and those skilled in the art can refer to the technical solution of the system to realize the step flow of the method , that is, the embodiments in the system can be understood as preferred examples for implementing the method, which will not be described in detail here.

Those skilled in the art know that, in addition to realizing the system provided by the present invention and its various devices in a purely computer-readable program code mode, the system provided by the present invention and its various devices can be completely programmed with logic gates, logic gates, The same functions can be realized in the form of switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system provided by the present invention and its various devices can be considered as a hardware component, and the devices included in it for realizing various functions can also be regarded as the structure in the hardware component; Means for implementing various functions can be regarded as either a software module implementing a method or a structure within a hardware component.

Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims (10)

1. A kind of 1. image classification method based on quality insertion in the case of label is noisy, it is characterised in that：Including：

   Network picture tag collection step：A large amount of pictures and the label letter of user's offer are provided from network picture sharing platform Breath, is filtered and is arranged according to required species, for use in the training of Image Classifier；

   Label quality factor Embedded step：The label quality factor is introduced in the image classification model for have supervision, for controlling band Make an uproar label predicted value generation and absorb the error back information from error label；Using maximizing log-likelihood function, if The optimization object function that meter is added after the label quality factor；

   Network model construction step：Optimization object function is modeled using deep neural network, obtains four models, respectively For encoding model, sampling model, decoded model and disaggregated model；

   Network parameter training step：The training picture that network picture tag collection step is obtained and the label with noise input net The above-mentioned network model that network model construction step obtains, being linked end to end using the stochastic gradient descent method of mutation, it is above-mentioned to train Four models, while model parameter is updated, the network model trained；

   Image classification step：New picture for requiring classification, inputs to the disaggregated model trained, obtains to the true mark of picture The prediction of label, while produce the classification results of image.
2. 2. the image classification method based on quality insertion in the case of label is noisy according to claim 1, it is characterised in that： In the label quality factor Embedded step, have the image classification model of supervision existing, add picture tag quality because The insertion of son, makes the new optimization object function be：

   <mrow> <mi>ln</mi> <mi> </mi> <mi>P</mi> <mrow> <mo>(</mo> <mi>Y</mi> <mo>|</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>ln</mi> <mi> </mi> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>ln</mi> <mi> </mi> <msub> <mi>E</mi> <mrow> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>,</mo> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> </mrow> </msub> <mo>&amp;lsqb;</mo> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>z</mi> <mi>m</mi> </msub> <mo>,</mo> <msub> <mi>s</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow>

   Wherein x_mAnd y_mBe respectively m pictures pixel set and relative users provide noise label, z_mAnd s_mIt is generation respectively The hidden variable of table picture true tag and label quality, M represent the picture sum for training；

   New optimization object function is bad caused by the label for concentrate mistake to training data due to adding the label quality factor Influence has absorption, meanwhile, the gradient function of the new optimization object function is difficult to calculate, therefore transfers to optimize its card first Simplify the required computing resource of training according to lower bound, while using skill is joined again, obtain final optimization object function formula combinations.
3. 3. the image classification method based on quality insertion in the case of label is noisy according to claim 2, it is characterised in that： The network model construction step, final optimization object function formula combinations are built respectively using deep neural network Mould, obtain four models：Encoding model, sampling model, decoded model and disaggregated model；Wherein：

   The encoding model, using convolutional neural networks, for generating noise label from image content XPriori prediction And combine noise label y q (S | X, Y) and true tag distribution q (Z | X, Y) are distributed to label quality and be predicted；

   The sampling model, label quality distribution q (S | X, Y) and true tag for encoding model to be generated be distributed q (Z | X, Y explicit value S and Z) are mapped as；

   The decoded model, using neutral net, it inputs the output label quality S and true tag Z for sampling model, is used for Generate and predict the posteriority of noise label q (Y | Z, S)；

   The disaggregated model, using convolutional neural networks, it generates the prediction to true tag Z using picture.
4. 4. the image classification method based on quality insertion in the case of label is noisy according to claim 3, it is characterised in that： The network parameter training step, the noise label posteriority recovered using decoded model predict that q (Y | Z, S) carries out the mould for having supervision Type training, calculation code model, sampling model, the passback gradient of decoded model, model parameter is updated, meanwhile, using encoding The true tag distribution q (Z | X, Y) obtained in model carries out the model training for having supervision to disaggregated model, calculates neutral net and returns Gradient is passed, updates model parameter.
5. 5. the image classification method according to claim any one of 1-4 based on quality insertion in the case of label is noisy, its It is characterised by：In the network picture tag collection step, web crawlers technology has been used, institute is collected on picture social network sites The label of a large amount of pictures and user annotation that need.
6. A kind of 6. image classification system based on quality insertion in the case of label is noisy, it is characterised in that：Including：

   Network picture tag collection module：A large amount of pictures and the label information of user's offer are provided from network picture sharing platform And filtered and arranged according to required species；

   The label quality factor is embedded in module：The label quality factor is introduced in the image classification model that tradition has supervision to control band Make an uproar label predicted value generation and absorb the error back information from error label, calculate image classification model corresponding to logarithm Optimization object function of the likelihood function as training；

   Network model builds module：For being modeled using deep neural network to the optimization object function, respectively obtain Encoding model, sampling model and decoded model and four models of disaggregated model；

   Network parameter training module：Training picture and the label with noise are inputted into network model, use the stochastic gradient of mutation Descent method links end to end trains four models, while updates model parameter；

   New images classification task processing module：New picture for requiring classification, input are obtained pair to the disaggregated model trained The prediction of picture true tag.
7. 7. the image classification system based on quality insertion in the case of label is noisy according to claim 6, it is characterised in that： The label quality factor is embedded in module, has the image classification model of supervision existing, adds picture tag quality factor Insertion, make the new optimization object function be：

   <mrow> <mi>ln</mi> <mi> </mi> <mi>P</mi> <mrow> <mo>(</mo> <mi>Y</mi> <mo>|</mo> <mi>X</mi> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>ln</mi> <mi> </mi> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>M</mi> </munderover> <mi>ln</mi> <mi> </mi> <msub> <mi>E</mi> <mrow> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>z</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>x</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>,</mo> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> </mrow> </msub> <mo>&amp;lsqb;</mo> <mi>P</mi> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mo>|</mo> <msub> <mi>z</mi> <mi>m</mi> </msub> <mo>,</mo> <msub> <mi>s</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> </mrow>

   Wherein x_mAnd y_mBe respectively m pictures pixel set and relative users provide noise label, z_mAnd s_mIt is generation respectively The hidden variable of table picture true tag and label quality, M represent the picture sum for training；

   New optimization object function is bad caused by the label for concentrate mistake to training data due to adding the label quality factor Influence has absorption, meanwhile, the gradient function of the new optimization object function is difficult to calculate, therefore transfers to optimize its card first Simplify the required computing resource of training according to lower bound, while using skill is joined again, obtain final optimization object function formula combinations.
8. 8. the image classification method based on quality insertion in the case of label is noisy according to claim 7, it is characterised in that： The network model builds module, and final optimization object function formula combinations are built respectively using deep neural network Mould, obtain four models：Encoding model, sampling model, decoded model and disaggregated model；Wherein：

   The encoding model, using convolutional neural networks, for generating noise label from image content XPriori prediction And combine noise label y q (S | X, Y) and true tag distribution q (Z | X, Y) are distributed to label quality and be predicted；

   The sampling model, label quality distribution q (S | X, Y) and true tag for encoding model to be generated be distributed q (Z | X, Y explicit value S and Z) are mapped as；

   The decoded model, used method are neutral net, and it is inputted as the output label quality S of sampling model and true Label Z, q (Y | Z, S) is predicted the posteriority of noise label for generating；

   The disaggregated model, used method are convolutional neural networks, and it generates the prediction to true tag Z using picture.
9. 9. the image classification method based on quality insertion in the case of label is noisy according to claim 8, it is characterised in that： The network parameter training module, the noise label posteriority recovered using decoded model predict that q (Y | Z, S) carries out the mould for having supervision Type training, calculation code model, sampling model, the passback gradient of decoded model, model parameter is updated, meanwhile, using encoding The true tag distribution q (Z | X, Y) obtained in model carries out the model training for having supervision to disaggregated model, calculates neutral net and returns Gradient is passed, updates model parameter.
10. 10. the image classification system according to claim any one of 6-7 based on quality insertion in the case of label is noisy, its It is characterised by：The network picture tag collection module, has used web crawlers technology, on picture social network sites needed for collection The label of a large amount of pictures and user annotation wanted.