CN110321952A

CN110321952A - A kind of training method and relevant device of image classification model

Info

Publication number: CN110321952A
Application number: CN201910592095.3A
Authority: CN
Inventors: 王晓宁; 孙钟前; 付星辉; 郑瀚; 杨巍
Original assignee: Tencent Healthcare Shenzhen Co Ltd
Current assignee: Tencent Healthcare Shenzhen Co Ltd
Priority date: 2019-07-02
Filing date: 2019-07-02
Publication date: 2019-10-11
Anticipated expiration: 2039-07-02
Also published as: CN110321952B

Abstract

The embodiment of the invention discloses the training methods and relevant device of a kind of image classification model, which comprises obtains training image collection；By related training image input picture classification learning model, obtain the prediction class probability that related training image is directed to each target category, and unrelated training image is inputted into described image classification learning model, obtain the prediction class probability that unrelated training image is directed to each target category；Determine probability distribution variances degree of the target category between the true probability distribution and prediction probability distribution under the related training image；It determines the prediction classification uncertainty of correlation training image, and determines the prediction classification uncertainty of unrelated training image；According to the prediction classification uncertainty of probability distribution variances degree, the prediction classification uncertainty of related training image and unrelated training image, the "current" model parameter of image classification learning model is adjusted.The generalization ability of image classification model can be improved through the embodiment of the present invention.

Description

A kind of training method and relevant device of image classification model

Technical field

This application involves machine learning field more particularly to a kind of training methods and relevant device of image classification model.

Background technique

With the development of artificial intelligence field, more and more work can be realized by computer, wherein image classification It is exactly feature of the computer by extraction image, is divided an image into according to the feature of image a certain in several graphic collections Class carries out the technology of interpretation classification instead of the mankind by vision.Presently, there are many algorithms for being used for image classification, such as KNN (k nearest neighbor classifier), SVM (support vector machines), CNN (convolutional neural networks) etc., CNN is manually pre-processed due to not needing And the operation such as additional feature extraction, become the main stream approach in image classification field.

During carrying out image classification by the image classification model based on CNN, image classification model extraction input figure The characteristics of image of picture, and then be the prediction probability of each pre-set categories according to box counting algorithm input picture, and then will most The corresponding classification of big prediction probability is determined as the classification of input object.For example, the default class that image classification model is classified There are two types of not, respectively cat and dog, i.e., after any image input picture disaggregated model, image classification model can calculate the image It is the prediction probability of cat and dog, and then output phase is to the corresponding classification of higher prediction probability (one of cat and dog).One In a little situations, the image input picture disaggregated model of cat or dog can be exported into accurate classification, in other cases, by cat Or after the image input picture disaggregated model other than dog, cat or the one such classification of dog are still exported, and corresponding inside model Prediction probability it is higher, such as 0.8 or 0.8 or more, that is to say, that image classification model is more for certain by the figure other than cat and dog As ground erroneous judgement is for cat or dog, illustrate that the training method of current image classification model prevents image classification model to not belonging to Reasonable prediction is carried out in the input picture of pre-set categories, so that the generalization ability of image classification model is weaker.

Summary of the invention

The application provides the training method and relevant device of a kind of image classification model, and image can be improved through the invention The generalization ability of disaggregated model.

On the one hand the embodiment of the present invention provides a kind of training method of image classification model, comprising:

Training image collection is obtained, the training image collection includes the related training image under different target classification, and not Belong to the unrelated training image of the target category, the correlation training image carries the corresponding classification mark of respective target category Label；

By the related training image input picture classification learning model, the related training image is obtained for each institute The prediction class probability of target category is stated, and the unrelated training image is inputted into described image classification learning model, obtains institute State the prediction class probability that unrelated training image is directed to each target category；

According to the corresponding class label of target category of the related training image, the related training image for each The prediction class probability of the target category determines true probability distribution of the target category under the related training image With the probability distribution variances degree between prediction probability distribution；

It is directed to the prediction class probability of each target category according to the related training image, determines the related instruction Practice the prediction classification uncertainty of image, and the prediction according to the unrelated training image for each target category is classified Probability determines the prediction classification uncertainty of the unrelated training image；

According to the probability distribution variances degree, the prediction classification uncertainty of the related training image and described unrelated The prediction classification uncertainty of training image, adjusts the "current" model parameter of described image classification learning model, so that passing through tune The prediction point of probability distribution variances degree, the related training image determined that image classification learning model after section is determined The prediction classification uncertainty of class uncertainty and the unrelated training image determined meets preset adjusted result item Part.

Wherein, the corresponding class label of target category according to the related training image, the related training figure Prediction class probability as being directed to each target category determines that the target category is true under the related training image Real probability distribution and prediction probability distribution between probability distribution variances degree include:

According to the corresponding class label of target category of the related training image, the related training image for each The prediction class probability of the target category determines the classification cross entropy of the related training image, and the correlation is trained The classification cross entropy of image is determined as the probability distribution variances degree.

Wherein, the prediction class probability that each target category is directed to according to the related training image, determines The prediction classification uncertainty of the correlation training image, and each target category is directed to according to the unrelated training image Prediction class probability, determine the unrelated training image prediction classification uncertainty include:

It is directed to the prediction class probability of each target category according to the related training image, determines the related instruction Practice the classification information entropy of image, and the classification information entropy of the related training image is determined as the pre- of the related training image Survey classification uncertainty；

It is directed to the prediction class probability of each target category according to the unrelated training image, determines the unrelated instruction Practice the classification information entropy of image, and the classification information entropy of the unrelated training image is determined as the pre- of the unrelated training image Survey classification uncertainty.

Wherein, the method also includes:

Classify respectively to the prediction of the prediction classification uncertainty and the unrelated training image of the related training image Uncertainty is normalized, obtain the related training image normalization uncertainty and the unrelated training image Normalization uncertainty；

It is described according to the probability distribution variances degree, the prediction classification uncertainty of the related training image and described The prediction classification uncertainty of unrelated training image, the "current" model parameter for adjusting described image classification learning model include:

According to the normalization uncertainty and the unrelated instruction of the probability distribution variances degree, the related training image Practice the normalization uncertainty of image, adjusts the "current" model parameter of described image classification learning model.

Wherein, described respectively to the prediction classification uncertainty of the related training image and the unrelated training image Prediction classification uncertainty is normalized, and obtains the normalization uncertainty of the related training image and described unrelated The normalization uncertainty of training image includes:

Determine maximum classification uncertainty of the target category in the case where probability is uniformly distributed；

By the ratio of the prediction classification uncertainty of the related training image and the maximum classification uncertainty, determine For the normalization uncertainty of the related training image, and the prediction of the unrelated training image is classified uncertainty and institute The ratio for stating maximum classification uncertainty, is determined as the normalization uncertainty of the unrelated trained object.

Wherein, it is described according to the probability distribution variances degree, the prediction classification uncertainty of the related training image with And the prediction classification uncertainty of the unrelated training image, adjust the "current" model parameter packet of described image classification learning model It includes:

According to the probability distribution variances degree, the prediction classification uncertainty of the related training image and described unrelated The prediction classification uncertainty of training image constructs the corresponding classification damage of "current" model parameter of described image classification learning model Lose function；

Local derviation is asked to the Classification Loss function, determines that the "current" model parameter of described image classification learning model is corresponding Lose gradient；

According to the corresponding loss gradient of the "current" model parameter of described image classification learning model and preset parametrics Habit rate adjusts the "current" model parameter of described image classification learning model.

Wherein, it is described according to the probability distribution variances degree, the prediction classification uncertainty of the related training image with And the prediction classification uncertainty of the unrelated training image, construct the "current" model parameter pair of described image classification learning model The Classification Loss function answered includes:

The corresponding Classification Loss function L of the "current" model parameter of described image classification learning model are as follows:

L=L_c+αL_r-βL_ur,

Wherein, L_cFor the probability distribution variances degree, α is that the prediction classification of the preset related training image is uncertain Spend corresponding weight, L_rFor the prediction classification uncertainty of the related training image, β is the preset unrelated training image The corresponding weight of prediction classification uncertainty, L_urFor the prediction classification uncertainty of the unrelated training image.

Wherein, the corresponding class label of target category according to the related training image, the related training figure Prediction class probability as being directed to each target category determines the classification cross entropy packet of each related training image It includes:

The classification cross entropy L of the correlation training image_ce-urAre as follows:

Wherein, i is the index of the related training image, and N is the quantity of the related training image, and i and N are positive whole Number, and i≤N, j are the index of the target category, K are the number of species of the target category, and j and K are positive integer, and j≤ K, y_jFor the corresponding class label of target category j, p_ijThe prediction class probability of target category j is directed to for related training image i.

Wherein, the prediction class probability that each target category is directed to according to the related training image, determines It is described correlation training image classification information entropy include:

The classification information entropy L of the correlation training image_e-rAre as follows:

Wherein, i is the index of the related training image, and N is the quantity of the related training image, and i and N are positive whole Number, and i≤N, j are the index of the target category, K are the number of species of the target category, and j and K are positive integer, and j≤ K, p_ijThe prediction class probability of target category j is directed to for related training image i；

The prediction class probability that each target category is directed to according to the unrelated training image, determines the nothing Close training image classification information entropy include:

The classification information entropy L of the unrelated training image_e-urAre as follows:

Wherein, l is the index of the unrelated training image, and M is the quantity of the unrelated training image, and l and M are positive whole Number, and l≤M, j are the index of the target category, K are the number of species of the target category, and j and K are positive integer, and j≤ K, g_ljThe prediction class probability of target category j is directed to for unrelated training image l.

On the other hand the embodiment of the present invention provides a kind of training device of image classification model, comprising:

Image set obtains module, and for obtaining training image collection, the training image collection includes under different target classification Related training image, and it is not belonging to the unrelated training image of the target category, the correlation training image carries respective The corresponding class label of target category；

Confidence determination module, for obtaining the phase for the related training image input picture classification learning model The prediction class probability that training image is directed to each target category is closed, and the unrelated training image is inputted into the figure As classification learning model, the prediction class probability that the unrelated training image is directed to each target category is obtained；

Diversity factor determining module, for according to the corresponding class label of target category of the related training image, described Related training image is directed to the prediction class probability of each target category, determines the target category in the related training Probability distribution variances degree between true probability distribution under image and prediction probability distribution；

Uncertainty determining module, for the prediction point according to the related training image for each target category Class probability determines the prediction classification uncertainty of the related training image, and according to the unrelated training image for each The prediction class probability of a target category determines the prediction classification uncertainty of the unrelated training image；

Parameter adjustment module, for being classified not according to the prediction of the probability distribution variances degree, the related training image The prediction classification uncertainty of degree of certainty and the unrelated training image, adjusts the current mould of described image classification learning model Shape parameter, so that the probability distribution variances degree determined by the image classification learning model after adjusting, the phase determined The prediction classification uncertainty of the prediction classification uncertainty and the unrelated training image determined of closing training image is full The preset adjusted result condition of foot.

Wherein, the diversity factor determining module, it is corresponding specifically for the target category according to the related training image Class label, the related training image are directed to the prediction class probability of each target category, determine the related training The classification cross entropy of image, and the classification cross entropy of the related training image is determined as the probability distribution variances degree.

Wherein, the uncertainty determining module is specifically used for being directed to each mesh according to the related training image The prediction class probability for marking classification, determines the classification information entropy of the related training image, and by the related training image Classification information entropy is determined as the prediction classification uncertainty of the related training image；

Wherein, described device further includes normalization module, not for the prediction classification respectively to the related training image The prediction classification uncertainty of degree of certainty and the unrelated training image is normalized, and obtains the related training image Normalization uncertainty and the unrelated training image normalization uncertainty；

The parameter adjustment module, specifically for being returned according to the probability distribution variances degree, the related training image One changes the normalization uncertainty of uncertainty and the unrelated training image, adjusts working as described image classification learning model Preceding model parameter.

Wherein, the normalization module includes maximum uncertainty determination unit and uncertainty normalization unit:

The maximum uncertainty determination unit is used to determine the target category in the case where probability is uniformly distributed Maximum classification uncertainty；

The uncertainty normalization unit, for by the prediction classification uncertainty of the related training image with it is described The ratio of maximum classification uncertainty, is determined as the normalization uncertainty of the related training image, and by the unrelated instruction Practice the prediction classification uncertainty of image and the ratio of the maximum classification uncertainty, is determined as the unrelated trained object Normalize uncertainty.

Wherein, the parameter adjustment module includes loss function structural unit, gradient determination unit and parameter adjustment unit:

The loss function structural unit is used for the prediction according to the probability distribution variances degree, the related training image The prediction classification uncertainty of uncertainty of classifying and the unrelated training image, construction described image classification learning model The corresponding Classification Loss function of "current" model parameter；

The gradient determination unit determines described image classification learning model for seeking local derviation to the Classification Loss function The corresponding loss gradient of "current" model parameter；

The parameter adjustment unit is used for the corresponding loss of "current" model parameter according to described image classification learning model Gradient and preset parameter learning rate adjust the "current" model parameter of described image classification learning model.

Wherein, the loss function structural unit, specifically for constructing the "current" model of described image classification learning model The corresponding Classification Loss function L of parameter are as follows:

L=L_c+αL_r-βL_ur,

Wherein, the diversity factor determining module is specifically used for determining the classification cross entropy L of the related training image_ce-ur Are as follows:

Wherein, the uncertainty determining module is specifically used for determining the classification information entropy L of the related training image_e-r Are as follows:

And determine the classification information entropy L of the unrelated training image_e-urAre as follows:

On the other hand the embodiment of the present invention provides a kind of training device of image classification model, comprising: processor and deposit Reservoir；

The processor is connected with memory, wherein for storing program code, the processor is used for the memory Said program code is called, to execute such as the method in the embodiment of the present invention.

The embodiment of the invention also provides a kind of computer readable storage medium, the computer-readable recording medium storage There is computer program, the computer program includes program instruction, and described program is instructed when being executed by a processor, executed such as this Method in inventive embodiments.

In the embodiment of the present invention, the training image collection comprising related training image and unrelated training image is obtained, it will be related Training image input picture classification learning model obtains the prediction class probability that related training image is directed to each target category, And by unrelated training image input picture classification learning model, unrelated training image is obtained for the prediction point of each target category Then class probability is directed to each target according to the corresponding class label of target category of related training image, related training image Prediction class probability respectively determines that true probability distribution and prediction probability of the target category under related training image are distributed it Between probability distribution variances degree, and be directed to according to related training image the prediction class probability of each target category, determine related The prediction classification of training image is uncertain, and the prediction class probability of each target category is directed to according to unrelated training image, is determined The prediction classification uncertainty of unrelated training image, and then classified according to the prediction of probability distribution variances degree, related training image The prediction classification uncertainty of uncertainty and unrelated training image, adjusts the "current" model ginseng of image classification learning model Number.In the training process of image classification learning model, optimization image classification learning model is instructed by probability distribution variances degree Nicety of grading, pass through related training image prediction classification uncertainty and unrelated training image prediction classify uncertainty The confidence level of the prediction probability of guidance optimization image classification learning model, to improve the image classification model that training obtains Generalization ability.

Detailed description of the invention

It to describe the technical solutions in the embodiments of the present invention more clearly, below will be to needed in the embodiment Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for ability For the those of ordinary skill of domain, without creative efforts, it can also be obtained according to these attached drawings other attached Figure.

Fig. 1 a is a kind of schematic network structure of DenseNet provided in an embodiment of the present invention；

Fig. 1 b is a kind of classification mechanism schematic diagram of the image classification model based on CNN provided in an embodiment of the present invention；

Fig. 1 c is a kind of schematic diagram of training image collection provided in an embodiment of the present invention；

Fig. 2 is a kind of flow diagram of the training method of image classification model provided in an embodiment of the present invention；

Fig. 3 is the schematic diagram of a kind of true probability distribution and prediction probability distribution provided in an embodiment of the present invention；

Fig. 4 is the flow diagram of the training method of another image classification model provided in an embodiment of the present invention；

Fig. 5 is a kind of structural schematic diagram of the training device of image classification model provided in an embodiment of the present invention；

Fig. 6 is the structural schematic diagram of the training device of another image classification model provided in an embodiment of the present invention.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that the described embodiment is only a part of the embodiment of the present invention, instead of all the embodiments.Based on this Embodiment in invention, every other reality obtained by those of ordinary skill in the art without making creative efforts Example is applied, shall fall within the protection scope of the present invention.

For image classification field CNN network there are many kinds of, such as DenseNET (intensive convolutional network), ResNet (residual error network), VGG (super-resolution cycle tests) network etc., in the instruction for introducing the image classification model in the embodiment of the present invention Before practicing method, brief introduction is done to the image classification model based on CNN by taking DenseNet (intensive convolutional network) as an example first, A refering to fig. 1, Fig. 1 a are a kind of schematic network structure of DenseNet provided in an embodiment of the present invention, and image to be classified is defeated After entering DenseNET, after DenseNET is handled, the classification results of image to be classified are obtained.As shown, DenseNet packet Containing convolutional layer, pond layer, intensive block (Dense Block) and Softmax layers, wherein convolutional layer is used to carry out characteristics of image It extracts or Feature Compression is carried out to the characteristics of image that intensive block extracts；Pond layer will be for that will extract the characteristic pattern that characteristics of image obtains It is converted into fixed dimension；Each intensive block may include 5 convolutional layers, and the image for extracting input picture different dimensions is special Sign, every layer of convolutional layer pass through this using the output of convolutional layer all before the convolutional layer in the same intensive block as input Kind network connection can alleviate the problem of gradient network disappears, and strengthen the multiplexing of intensive block internal image feature, reduce meter Calculation amount；Softmax layers for calculating input image for the prediction probability of each classification.

Referring next to Fig. 1 b, Fig. 1 b is a kind of point of image classification model based on CNN provided in an embodiment of the present invention Class schematic diagram of mechanism, image classification model shown in Fig. 1 b are DenseNET, which is two points for cat and dog Class model, that is, any image inputs DenseNET, and there are two types of possible: cat or dog for the classification results meeting of output.Such as Fig. 1 b institute Show, there are four images to be sorted, respectively the first image, the second image, third image and the 4th image, by aforementioned four figure As that can be respectively obtained for the respective classification results of aforementioned four image, successively after input DenseNET are as follows: cat, dog, dog and cat.

The application scenarios of the training method of image classification model provided in an embodiment of the present invention can be in medical imaging knowledge Other field is trained when carrying out disorder in screening (i.e. target classification is illness and non-illness) according to medical imaging for picture Table images, the image classification mould of the low disorder in screening result of output confidence level are not conformed to caused by the reasons such as reflective, shake, foreign matter Type.

Next the specific implementation for introducing the training method of image classification model provided in an embodiment of the present invention, refering to Fig. 2, Fig. 2 are a kind of flow diagram of the training method of image classification model provided in an embodiment of the present invention, as shown, institute The method of stating may comprise steps of:

S201 obtains training image collection.

Here, the training image that training image is concentrated includes the related training image under different target classification, and is not belonged to In the unrelated training image of target category, related training image carries the corresponding class label of respective target category.Wherein mesh Marking classification, the image classification model to obtain after preset training is directed to the classification that input picture can be sorted out.For example, in Fig. 1 b DenseNET includes dog and cat for its preset target category before training, then the image classification model obtained after training is Two disaggregated models, that is, input picture is classified as cat or dog, such as the image that the first image in Fig. 1 b is a width cat, is obtained Classification results are cats；Second image is the image of a width dog, and obtained classification results are dogs；Third image is the figure of a width aircraft Picture, obtained classification results are dogs；4th image is the image of a width lion, and obtained classification results are cats.For another example, if it is default Target category include bicycle, aircraft and flower, then train after image classification model be three disaggregated models, be similar to Fig. 1 b It is middle that any input picture is classified as one of cat and dog, input picture can be classified as to bicycle, aircraft or flower in this example One of piece, specific example is not listed.

Above-mentioned target category includes at least two, and related training image is the training figure under above-mentioned at least two target category Picture, unrelated training image are the unrelated training image for being not belonging to above-mentioned at least two target category.For example, if target category includes Cat and dog two categories, referring to Fig. 1 c, Fig. 1 c is a kind of schematic diagram of training image collection provided in an embodiment of the present invention, such as Fig. 1 c Shown, training image collection includes related training image and unrelated training image, wherein related training image can be as shown in the figure Multiple include cat images and multiple include dog images, unrelated training image can be multiple as shown in the figure include aircraft But does not include the image of cat and dog, comprising bicycle but do not include the image of cat and dog, comprising lion but do not include cat and dog Image comprising mouse but does not include the image of cat and dog, comprising keyboard but does not include the image of cat and dog and comprising house but not Image comprising cat and dog.

Here, the related training image and unrelated training image that training image is concentrated can be what user's identification filtered out, It is also possible to what the identification of other machines model filtered out.For ease of calculation machine to training image concentrate training image classification into Row is distinguished and the calculating of subsequent related training parameter, can distribute different classifications for the training image in training set of images Label, and classification mark is carried out to training image, so that related training image carries the corresponding classification mark of respective target category Label, unrelated training image carry the corresponding unified class label of classification other than target category.For example, if target category Only include cat and dog, class label 1 can be marked to the related training image in training set of images including cat, to training set of images Related training image in conjunction comprising dog marks class label 2, the unrelated training to dog and cat is not included in training set of images conjunction Image labeling class label 3.

Related training image input picture classification learning model is obtained related training image for each target by S202 The prediction class probability of classification, and by unrelated training image input picture classification learning model, it obtains unrelated training image and is directed to The prediction class probability of each target category.

Here image classification learning model can be the archetype after training initial stage model parameter initialization, to model Parameter initialization includes being pre-configured with to weight matrix in image classification learning model, the image classification after initializing at this time Practising model has certain image classification ability (classification capacity under normal conditions at this time is lower), or to image classification Learning model carries out the mid-module after model parameter adjusting several times.What prediction class probability here can indicate to be classified Related training image or unrelated training image, are classified as the confidence level of each classification in target category.For in Fig. 1 a Shown in DenseNet network, Softmax layers of input picture that can be exported are directed to the prediction class probability of each target category. For example, target category only includes cat and dog, after related training image or unrelated training image input picture classification learning model, obtain To the probability of related training image or unrelated training image comprising cat for input and comprising the probability of dog, and it is directed to same Input picture, the probability of the two and be 1.If image A is 0.9 for the prediction class probability of cat, the prediction for dog is classified generally Rate is 0.1, and image B is 0.6 for the prediction class probability of cat, and the prediction class probability for dog is 0.4, although the two can It is classified as cat, but image A is higher than image B for the prediction class probability of cat, that is, image A is classified as the confidence journey of cat Degree is higher than image B.

S203 is directed to each mesh according to the corresponding class label of the target category of related training image, related training image The prediction class probability for marking classification determines that true probability distribution of the target category under related training image is distributed with prediction probability Between probability distribution variances degree.

Probability distribution, for expressing the probabilistic law of stochastic variable value.Here it is possible to become using target category as random The value range of amount, the stochastic variable is the corresponding class label of each target category, according to the respective mark of related training image The true probability distribution of the available training image of classification is signed, the corresponding probability of label classification of related training image is 1, The corresponding probability of label classification of his target category is 0；The prediction for each target category point obtained according to step S202 The prediction probability distribution of the available related training image of class probability.Refering to Fig. 3, Fig. 3 is one kind provided in an embodiment of the present invention The schematic diagram of true probability distribution and prediction probability distribution, it is assumed that in two classification of cat and dog, the class label of cat is 0, dog Class label be 1, the corresponding class label of other classifications is 2, one include cat related training image, obtain it for cat Prediction class probability be 0.8, for dog prediction class probability be 0.2, then its true probability distribution and prediction probability distribution As shown in Figure 3.It can be appreciated that the prediction probability of the related training image obtained by image classification learning model be distributed it is closer The true probability of related training image is distributed, and illustrates that the classification capacity of image classification learning model is better, therefore mesh can be used Probability distribution variances degree of the classification between the true probability distribution under related training image and prediction probability distribution is marked, is reversely referred to The adjusting of the model parameter in image classification learning model is led, to optimize the classification capacity of image classification learning model.

Wherein, probability point of the target category between the true probability distribution under related training image and prediction probability distribution Cloth diversity factor can there are many forms to indicate, such as Euclidean distance between the two, manhatton distance, Chebyshev's distance, Fujian Can Paderewski distance, mahalanobis distance, cosine angle, cross entropy, relative entropy etc., for example, the example based on Fig. 3, corresponding Europe Formula distance isCorresponding cross entropy is H=- (0 × log0.8+1 × log0.2) ≈ 0.70, the concrete form of probability distribution variances degree is without limitation here.

S204 is directed to the prediction class probability of each target category according to related training image, determines correlation training image Prediction classify uncertainty, and according to unrelated training image be directed to each target category prediction class probability, determine unrelated The prediction classification uncertainty of training image.

Here, each related training image or unrelated training image are respectively directed to the prediction class probability of each target classification Characterize its confidence level for being classified as each classification in target category, then it is true according to the prediction class probability of related training image The prediction classification uncertainty of fixed related training image, can characterize the classification confidence level of related training image, according to nothing The prediction classification uncertainty for closing the unrelated training image for predicting that class probability is determining of training image, can characterize unrelated training The classification confidence level of image.The high image classification learning model of generalization ability to the classification confidence level of related training image compared with Height, and it is lower to the classification confidence level of unrelated training image, therefore can be uncertain with the prediction classification of related training image The prediction of degree and unrelated training image is classified uncertainty, the tune of the model parameter in reversed guide image classification learning model Section, to optimize the generalization ability of image classification learning model.

In a kind of implementation, the prediction classification uncertainty of related training image can be by the classification of related training image Comentropy indicates that the prediction classification uncertainty of unrelated training image can be indicated by the classification information entropy of unrelated training image.

S205 schemes according to probability distribution variances degree, the prediction classification uncertainty of related training image and unrelated training The prediction classification uncertainty of picture, adjusts the "current" model parameter of image classification learning model.

Specifically, can be classified uncertainty according to the prediction of probability distribution variances degree, related training image first, and The prediction classification uncertainty of unrelated training image constructs the corresponding classification damage of "current" model parameter of image classification learning model Function is lost, for example, using L_cIndicate probability distribution variances degree, L_rIndicate the prediction classification uncertainty of related training image, L_urTable Show the prediction classification uncertainty of unrelated training image, above-mentioned Classification Loss function can be constructed as: L=L_c+L_r-L_ur, by phase Close the prediction classification uncertainty L of training image_rWith the prediction classification uncertainty L of unrelated training image_ur, as Classification Loss The regularization term of the confidence level of image classification learning model prediction probability is constrained in function；Then the Classification Loss function is asked Local derviation determines the corresponding loss gradient of the "current" model parameter of described image classification learning model, and then according to described image point The corresponding loss gradient of the "current" model parameter of class learning model and preset parameter learning rate adjust described image taxology Practise model "current" model parameter so that by adjusting after image classification learning model determine probability distribution variances degree, The prediction classification uncertainty for the related training image determined and the prediction classification for the unrelated training image determined be not true Fixed degree meets preset adjusted result condition.

In a kind of optional mode, preset adjusted result condition be can be for probability distribution variances degree, related training The prediction classification uncertainty of image and prediction classification this three of uncertainty of unrelated training image, preset respective tune respectively Nodule really bar part.For example, above-mentioned adjusted result condition may include: to be redefined out by the image classification model after adjusting Probability distribution variances degree, compared to the obtained probability distribution variances degree rate of descent before adjusting within the scope of the first rate threshold； The prediction classification uncertainty of the related training image redefined out, it is pre- compared to the related training image obtained before adjusting Classification uncertainty rate of descent is surveyed within the scope of the second rate threshold；The prediction classification of the unrelated training image redefined out is not Degree of certainty, the prediction compared to the unrelated training image obtained before adjusting classify uncertainty ascending amount in third proportion threshold value model In enclosing.

In another optional mode, preset adjusted result condition be can be by for probability distribution variances degree, phase The unified preset tune of the prediction classification uncertainty of the prediction classification uncertainty and unrelated training image of closing training image Nodule really bar part.For example, can be according to the prediction classification uncertainty and unrelated instruction of probability distribution variances degree, related training image The prediction classification uncertainty for practicing image constructs Classification Loss function, and then above-mentioned adjusted result condition can be to damage for classification Lose the constraint condition of function: by the value of the corresponding Classification Loss function of image classification learning model after adjusting, compared to tune The value rate of descent of the corresponding Classification Loss function of image classification learning model obtained before section is in preset loss suppression ratio In example threshold range.

In the training process of actual image classification learning model, image classification model provided in an embodiment of the present invention Training method iteration can execute in a computer, until the probability distribution of the related training image of image classification learning model is poor Different degree is less than preset diversity factor threshold value, and the prediction classification uncertainty of related training image is less than preset first and does not know Degree, and stop iteration in the case that the prediction classification uncertainty of unrelated training image is greater than preset second uncertainty, or Person stops iteration in the case that the value of Classification Loss function is less than preset Classification Loss threshold value.

In neural network model, data are normalized not only can be by numerical definiteness in certain range It is interior, solve the problems, such as in computer calculating process that numerical value is excessive or too small bring numerical value overflows, it can also be ensured that under gradient During drop method training neural network model, the instruction of neural network model is can be completed in one unified parameter learning rate of setting Practice, reduces calculating data volume.

Therefore, optionally, after step s 204, the training method of image classification model provided in this embodiment can be with The following steps are included: not true to the prediction classification of the prediction classification uncertainty of related training image and unrelated training image respectively Fixed degree is normalized, and the normalization of the normalization uncertainty and unrelated training image that obtain related training image is not true Fixed degree.In step S205 specifically can according to probability distribution variances degree, related training image normalization uncertainty and The normalization uncertainty of unrelated training image adjusts the "current" model parameter of image classification learning model.

It is further alternative, classify to the prediction of the prediction classification uncertainty and unrelated training image of related training image The concrete mode that uncertainty is normalized can be with are as follows: determines target category in the case where probability is uniformly distributed most Macrotaxonomy uncertainty；By the ratio of the prediction classification uncertainty of related training image and maximum classification uncertainty, determine For the normalization uncertainty of related training image, and not by the prediction of unrelated training image classification uncertainty and maximum classification The ratio of degree of certainty is determined as the normalization uncertainty of unrelated trained object.

Referring to fig. 4, Fig. 4 is the process signal of the training method of another image classification model provided in an embodiment of the present invention Figure, as shown, the method may include following steps:

S401 obtains training image collection.

Wherein, training image collection includes the related training image under different target classification, and is not belonging to target category Unrelated training image, related training image carry the corresponding class label of respective target category.

Related training image input picture classification learning model is obtained related training image for each target by S402 The prediction class probability of classification, and unrelated training image is inputted into described image classification learning model, obtain unrelated training image For the prediction class probability of each target category.

The specific implementation of step S401 and step S402 can respectively refering to step S201 in the corresponding embodiment of Fig. 2 and The specific implementation of step S202, details are not described herein again.

S403 is directed to each mesh according to the corresponding class label of the target category of related training image, related training image The prediction class probability of classification is marked, determines the classification cross entropy of correlation training image.

In a kind of optional specific implementation, the classification cross entropy of related training image can be determined according to the following formula L_ce-ur:

In another optional concrete mode, the classification cross entropy of related training image can be determined according to the following formula L_ce-ur:

Wherein, consistent in the meaning of parameters and formula (1) in formula (2), determinations is related trained in formula (1) The total classification cross entropy of image, determination is the average classification cross entropy of related training image in formula (2), and the two can characterize Difference degree of the target category between the true probability distribution under related training image and prediction probability distribution, does not limit herein It is fixed.

S404 is directed to the prediction class probability of each target category according to related training image, determines correlation training image Classification information entropy, and according to unrelated training image be directed to each target category prediction class probability, determine unrelated instruction Practice the classification information entropy of image.

In the specific implementation, can determine the classification information entropy L of related training image by following formula_e-r:

The classification information entropy L of unrelated training image can be determined by following formula_e-ur:

S405 determines maximum classification information entropy of the target category in the case where probability is uniformly distributed.

Here, according to principle of maximum entropy, in the case where not more information about stochastic variable, it is assumed that stochastic variable Be it is equally distributed, that is, be directed to a training image, the probability for being classified as various target categories is equal, at this time probability point The comentropy of cloth is maximum, for maximum classification information entropy.Specifically, determining maximum classification information entropy by following formula H_max-entropy:

Wherein, K is the number of species of target category.

The ratio of the classification information entropy of related training image and maximum classification information entropy is determined as related training by S406 The normalization classification information entropy of image, and by the ratio of the classification information entropy of unrelated training image and maximum classification information entropy, really It is set to the normalization classification information entropy of unrelated training image.

Specifically, obtaining the normalization classification information entropy L of related training image to formula (5) according to formula (3)_N-e-rAre as follows:

Wherein, consistent in the meaning of parameters and formula (3) in formula (6).It is obtained according to formula (4) and formula (5) Unrelated training image normalization classification information entropy L_N-e-urAre as follows:

Wherein, the meaning of the parameters in formula (7) is consistent in formula (4).

S407, according to classification cross entropy, the related normalization classification information entropy of training image and returning for unrelated training image One changes classification information entropy, adjusts the "current" model parameter of image classification learning model.

Here, according to classification cross entropy, the related normalization classification information entropy of training image and returning for unrelated training image One, which changes classification information entropy, constructs Classification Loss function, then adjusts image classification learning model according to direction of error propagation algorithm "current" model parameter.

It according to Classification Loss function can be L=L specifically, described_ce-ur+L_N-e-r-L_N-e-ur, or L=L_ce-ur+ αL_N-e-r-βL_N-e-ur, wherein α is the corresponding weight of normalization classification information entropy of preset related training image, and β is the nothing Close the corresponding weight of normalization classification information entropy of training image.And then each related training image and unrelated training image are respectively The parameters such as corresponding class label, prediction class probability substitute into Classification Loss function, obtain to weigh in image classification learning model The weight model parameters such as W and offset b are the loss function form of independent variable, the loss function can be sought local derviation to weight W, are determined Weight W in "current" model parameter out₀Gradient dL/dW, and then pass through formula W_t=W₀- a × dL/dW determines that "current" model is joined Weight W in number₀More fresh target W_t, wherein a is preset parameter learning rate.Similarly, local derviation is asked to offset b, and then passes through Formula b_t=b₀- a × dL/db determines the offset b in "current" model parameter₀More fresh target b_t。

In the training process of actual image classification learning model, image classification model provided in an embodiment of the present invention Training method iteration can execute in a computer, until the classification cross entropy of the related training image of image classification learning model Normalization classification information entropy less than preset intersection entropy threshold, and related training image is less than preset first information entropy threshold Value, and stop iteration in the case that the normalization classification information entropy of unrelated training image is greater than preset second information entropy threshold, Or stop in the case that the value of the Classification Loss function of image classification learning model is less than preset Classification Loss threshold value Only iteration.

In the embodiment of the present invention, the training image collection comprising related training image and unrelated training image is obtained, it will be related Training image input picture classification learning model obtains the prediction class probability that related training image is directed to each target category, And by unrelated training image input picture classification learning model, unrelated training image is obtained for the prediction point of each target category Then class probability is directed to each target according to the corresponding class label of target category of related training image, related training image Prediction class probability respectively determines the classification cross entropy of correlation training image, and is directed to each mesh according to related training image The prediction class probability of classification is marked, determines the classification information entropy of correlation training image, each mesh is directed to according to unrelated training image The prediction class probability for marking classification, determines the classification information entropy of unrelated training image, then to related training image and unrelated instruction The classification information entropy for practicing image carries out the normalization of maximum informational entropy, and then according to the classification cross entropy of related training image, phase The normalization classification information after normalization classification information entropy and the normalization of unrelated training image after closing training image normalization Entropy adjusts the "current" model parameter of image classification learning model.In the training process of image classification learning model, pass through correlation The nicety of grading of the classification cross entropy guidance optimization image classification learning model of training image, is normalized by related training image Normalization classification information entropy after rear normalization classification information entropy and the normalization of unrelated training image, guidance optimization image point The confidence level of the prediction probability of class learning model, to improve the generalization ability for the image classification model that training obtains.

Referring to Fig. 5, Fig. 5 is a kind of structural representation of the training device of image classification model provided in an embodiment of the present invention Figure, as shown, the training device 50 of shown image classification model, which can include at least image set, obtains module 501, confidence level Determining module 502, diversity factor determining module 503, uncertainty determining module 504 and parameter adjustment module 505, in which:

Image set obtains module 501, and for obtaining training image collection, the training image collection includes under different target classification Related training image, and be not belonging to the unrelated training image of the target category, the correlation training image carries respective The corresponding class label of target category；

Confidence determination module 502, for obtaining the related training image input picture classification learning model described Related training image is directed to the prediction class probability of each target category, and will be described in the unrelated training image input Image classification learning model obtains the prediction class probability that the unrelated training image is directed to each target category；

Diversity factor determining module 503, for the corresponding class label of target category according to the related training image, institute The prediction class probability that related training image is directed to each target category is stated, determines the target category in the related instruction Practice the probability distribution variances degree between the true probability distribution under image and prediction probability distribution；

Uncertainty determining module 504, for being directed to the pre- of each target category according to the related training image Class probability is surveyed, determines the prediction classification uncertainty of the related training image, and according to the unrelated training image needle To the prediction class probability of each target category, the prediction classification uncertainty of the unrelated training image is determined；

Parameter adjustment module 505, for being classified according to the prediction of the probability distribution variances degree, the related training image The prediction classification uncertainty of uncertainty and the unrelated training image, adjusts the current of described image classification learning model Model parameter so that by adjust after image classification learning model determine probability distribution variances degree, determine it is described The prediction classification uncertainty of related training image and the prediction classification uncertainty for the unrelated training image determined Meet preset adjusted result condition.

Optionally, the diversity factor determining module 503, specifically for the target category pair according to the related training image Class label, the related training image answered are directed to the prediction class probability of each target category, determine the correlation The classification cross entropy of training image, and the classification cross entropy of the related training image is determined as the probability distribution variances Degree.

Optionally, the uncertainty determining module 504, is specifically used for: according to the related training image for each The prediction class probability of the target category determines the classification information entropy of the related training image, and the correlation is trained The classification information entropy of image is determined as the prediction classification uncertainty of the related training image；

And the prediction class probability of each target category is directed to according to the unrelated training image, determine the nothing The classification information entropy of training image is closed, and the classification information entropy of the unrelated training image is determined as the unrelated training image Prediction classify uncertainty.

Optionally, described device further includes normalization module 506, for the prediction point respectively to the related training image The prediction classification uncertainty of class uncertainty and the unrelated training image is normalized, and obtains the related training The normalization uncertainty of the normalization uncertainty of image and the unrelated training image；

The parameter adjustment module 505, specifically for according to the probability distribution variances degree, the related training image The normalization uncertainty of uncertainty and the unrelated training image is normalized, described image classification learning model is adjusted "current" model parameter.

Optionally, the normalization module 506 includes maximum uncertainty determination unit 5061 and uncertainty normalization Unit 5062:

The maximum uncertainty determination unit 5061, for determining the target category the case where probability is uniformly distributed Under maximum classification uncertainty；

The uncertainty normalization unit 5062, for by the prediction classification uncertainty of the related training image with The ratio of the maximum classification uncertainty, is determined as the normalization uncertainty of the related training image, and by the nothing The prediction classification uncertainty of training image and the ratio of the maximum classification uncertainty are closed, the unrelated training pair is determined as The normalization uncertainty of elephant.

Optionally, the parameter adjustment module 505 includes loss function structural unit 5051,5052 and of gradient determination unit Parameter adjustment unit 5053:

The loss function structural unit 5051, for according to the probability distribution variances degree, the related training image Prediction classification uncertainty and the unrelated training image prediction classify uncertainty, construct described image classification learning The corresponding Classification Loss function of the "current" model parameter of model；

The gradient determination unit 5052 determines described image taxology for seeking local derviation to the Classification Loss function Practise the corresponding loss gradient of "current" model parameter of model；

The parameter adjustment unit 5053, it is corresponding for the "current" model parameter according to described image classification learning model Gradient and preset parameter learning rate are lost, the "current" model parameter of described image classification learning model is adjusted.

Optionally, the loss function structural unit 5051, specifically for working as construction described image classification learning model The corresponding Classification Loss function L of preceding model parameter are as follows:

L=L_c+αL_r-βL_ur,

Optionally, the diversity factor determining module 503 is specifically used for determining the classification cross entropy of the related training image L_ce-urAre as follows:

Optionally, the uncertainty determining module 504 is specifically used for determining the classification information of the related training image Entropy L_e-rAre as follows:

In the specific implementation, the training device of described image disaggregated model can be executed such as by each functional module built in it Each step in the training method of the image classification model of Fig. 2 or Fig. 4, specific implementation details sees Fig. 2 and Fig. 4 is corresponding The realization details of each step in embodiment, details are not described herein again.

In the embodiment of the present invention, image set obtains module and obtains the training comprising related training image and unrelated training image Related training image input picture classification learning model is obtained related training image and is directed to by image set, confidence determination module The prediction class probability of each target category, and by unrelated training image input picture classification learning model, obtain unrelated training Image is directed to the prediction class probability of each target category, and then diversity factor determining module is according to the target class of related training image Not corresponding class label, related training image are directed to the prediction class probability of each target respectively, determine target category in phase Close the probability distribution variances degree between the true probability distribution under training image and prediction probability distribution, uncertainty determining module It is directed to the prediction class probability of each target category according to related training image, determines that the prediction classification of correlation training image is not true It is fixed, and the prediction point of unrelated training image is determined for the prediction class probability of each target category according to unrelated training image Class uncertainty, and then parameter adjustment module is according to the prediction classification uncertainty of probability distribution variances degree, related training image And the prediction classification uncertainty of unrelated training image, adjust the "current" model parameter of image classification learning model.In image In the training process of classification learning model, parameter adjustment module learns mould by probability distribution variances degree guidance optimization image classification The nicety of grading of type is classified uncertain by the prediction classification uncertainty of related training image and the prediction of unrelated training image The confidence level of the prediction probability of degree guidance optimization image classification learning model, to improve the image classification model that training obtains Generalization ability.

Referring to Fig. 6, Fig. 6 is the structural representation of the training device of another image classification model provided in an embodiment of the present invention Figure, as shown, the training device 60 of described image disaggregated model includes: at least one processor 601, such as CPU, at least one A network interface 604, user interface 603, memory 605, at least one communication bus 602.Wherein, communication bus 602 is used for Realize the connection communication between these components.Wherein, user interface 603 may include display screen (Display), camera (Camera), optional user interface 603 can also include standard wireline interface and wireless interface.Network interface 604 optionally may be used To include standard wireline interface and wireless interface (such as WI-FI interface).Memory 605 can be high speed RAM memory, can also To be non-labile memory (non-volatile memory), for example, at least a magnetic disk storage.Memory 605 can Choosing can also be that at least one is located remotely from the storage device of aforementioned processor 601.As shown in fig. 6, as a kind of computer It may include operating system, network communication module, Subscriber Interface Module SIM and image classification mould in the memory 605 of storage medium The training application program of type.

In terminal 6 shown in Fig. 6, user interface 603 is mainly used for receiving the training that user is directed to training image concentration The interface of the class label of image；And processor 601 can be used for calling the instruction of the image classification model stored in memory 605 Practice application program, and specifically execute following operation:

According to the probability distribution variances degree, the prediction classification uncertainty of the related training image and described unrelated The prediction classification uncertainty of training image, adjusts the "current" model parameter of described image classification learning model, so that passing through tune The prediction of the probability distribution variances degree that image classification learning model after section is determined and the related training image determined The prediction classification uncertainty of classification uncertainty and the unrelated training image determined meets preset adjusted result Condition.

It should be appreciated that the executable Fig. 2 above of the training device 60 of image classification model described in the embodiment of the present invention Or the description in embodiment corresponding to Fig. 4 to the training method of described image disaggregated model, it also can be performed real corresponding to Fig. 5 above The description in example to the training device 50 of described image disaggregated model is applied, details are not described herein.In addition, to using same procedure Beneficial effect description, is also no longer repeated.

The embodiment of the present invention also provides a kind of computer storage medium, and the computer storage medium is stored with computer journey Sequence, the computer program include program instruction, and described program instruction executes the computer such as Method described in previous embodiment, the computer can be one of the training device of image classification model mentioned above Point.

Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access Memory, RAM) etc..

The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.

Claims

1. a kind of training method of image classification model characterized by comprising

Training image collection is obtained, the training image collection includes the related training image under different target classification, and is not belonging to The unrelated training image of the target category, the correlation training image carry the corresponding class label of respective target category；

By the related training image input picture classification learning model, the related training image is obtained for each mesh The prediction class probability of classification is marked, and the unrelated training image is inputted into described image classification learning model, obtains the nothing Close the prediction class probability that training image is directed to each target category；

According to the corresponding class label of target category of the related training image, the related training image for each described The prediction class probability of target category, determine true probability distribution of the target category under the related training image with it is pre- Survey the probability distribution variances degree between probability distribution；

It is directed to the prediction class probability of each target category according to the related training image, determines the related training figure The prediction classification uncertainty of picture, and the prediction according to the unrelated training image for each target category is classified generally Rate determines the prediction classification uncertainty of the unrelated training image；

According to the prediction classification uncertainty and the unrelated training of the probability distribution variances degree, the related training image The prediction classification uncertainty of image, adjusts the "current" model parameter of described image classification learning model, so that after by adjusting The prediction classification of probability distribution variances degree, the related training image determined determined of image classification learning model not The prediction classification uncertainty of degree of certainty and the unrelated training image determined meets preset adjusted result condition.

2. the method according to claim 1, wherein the target category pair according to the related training image Class label, the related training image answered are directed to the prediction class probability of each target category, determine the target True probability of the classification under the related training image is distributed the probability distribution variances degree between prediction probability distribution and includes:

According to the corresponding class label of target category of the related training image, the related training image for each described The prediction class probability of target category, determines the classification cross entropy of the related training image, and by the related training image Classification cross entropy be determined as the probability distribution variances degree.

3. the method according to claim 1, wherein it is described according to the related training image for each described The prediction class probability of target category determines the prediction classification uncertainty of the related training image, and according to described unrelated Training image is directed to the prediction class probability of each target category, determines that the prediction classification of the unrelated training image is not true Degree includes: calmly

It is directed to the prediction class probability of each target category according to the related training image, determines the related training figure The classification information entropy of picture, and the prediction that the classification information entropy of the related training image is determined as the related training image is divided Class uncertainty；

It is directed to the prediction class probability of each target category according to the unrelated training image, determines the unrelated training figure The classification information entropy of picture, and the prediction that the classification information entropy of the unrelated training image is determined as the unrelated training image is divided Class uncertainty.

4. the method according to claim 1, wherein the method also includes:

The prediction classification of the prediction classification uncertainty to the related training image and the unrelated training image is not true respectively Fixed degree is normalized, and obtains the related normalization uncertainty of training image and returning for the unrelated training image One changes uncertainty；

It is described according to the probability distribution variances degree, the prediction classification uncertainty of the related training image and described unrelated The prediction classification uncertainty of training image, the "current" model parameter for adjusting described image classification learning model include:

According to the probability distribution variances degree, the normalization uncertainty of the related training image and the unrelated training figure The normalization uncertainty of picture adjusts the "current" model parameter of described image classification learning model.

5. according to the method described in claim 4, it is characterized in that, described respectively classify to the prediction of the related training image The prediction classification uncertainty of uncertainty and the unrelated training image is normalized, and obtains the related training figure The normalization uncertainty of picture and the normalization uncertainty of the unrelated training image include:

By the ratio of the prediction classification uncertainty of the related training image and the maximum classification uncertainty, it is determined as institute State the normalization uncertainty of related training image, and by the prediction of unrelated training image classification uncertainty and it is described most The ratio of macrotaxonomy uncertainty is determined as the normalization uncertainty of the unrelated trained object.

6. method according to any one of claims 1 to 3, which is characterized in that it is described according to the probability distribution variances degree, The prediction classification uncertainty of the correlation training image and the prediction classification uncertainty of the unrelated training image, are adjusted The "current" model parameter of described image classification learning model includes:

According to the prediction classification uncertainty and the unrelated training of the probability distribution variances degree, the related training image The prediction classification uncertainty of image, constructs the corresponding Classification Loss letter of "current" model parameter of described image classification learning model Number；

Local derviation is asked to the Classification Loss function, determines the corresponding loss of "current" model parameter of described image classification learning model Gradient；

According to the corresponding loss gradient of the "current" model parameter of described image classification learning model and preset parameter learning rate, Adjust the "current" model parameter of described image classification learning model.

7. according to the method described in claim 6, it is characterized in that, described according to the probability distribution variances degree, the correlation The prediction classification uncertainty of training image and the prediction classification uncertainty of the unrelated training image, construct described image The corresponding Classification Loss function of the "current" model parameter of classification learning model includes:

L=L_c+αL_r-βL_ur,

Wherein, L_cFor the probability distribution variances degree, α is the prediction classification uncertainty pair of the preset related training image The weight answered, L_rFor the prediction classification uncertainty of the related training image, β is the pre- of the preset unrelated training image Survey the corresponding weight of classification uncertainty, L_urFor the prediction classification uncertainty of the unrelated training image.

8. according to the method described in claim 2, it is characterized in that, the target category pair according to the related training image Class label, the related training image answered are directed to the prediction class probability of each target category, determine each described The classification cross entropy of related training image includes:

Wherein, i is the index of the related training image, and N is the quantity of the related training image, and i and N are positive integer, and i ≤ N, j are the index of the target category, and K is the number of species of the target category, and j and K are positive integer, and j≤K, y_jFor The corresponding class label of target category j, p_ijThe prediction class probability of target category j is directed to for related training image i.

9. according to the method described in claim 3, it is characterized in that, it is described according to the related training image for each described The prediction class probability of target category determines that the classification information entropy of the related training image includes:

Wherein, i is the index of the related training image, and N is the quantity of the related training image, and i and N are positive integer, and i ≤ N, j are the index of the target category, and K is the number of species of the target category, and j and K are positive integer, and j≤K, p_ijFor Related training image i is directed to the prediction class probability of target category j；

The prediction class probability that each target category is directed to according to the unrelated training image, determines the unrelated instruction Practice image classification information entropy include:

Wherein, l is the index of the unrelated training image, and M is the quantity of the unrelated training image, and l and M are positive integer, and l ≤ M, j are the index of the target category, and K is the number of species of the target category, and j and K are positive integer, and j≤K, g_ljFor Unrelated training image l is directed to the prediction class probability of target category j.

10. a kind of training device of image classification model characterized by comprising

Image set obtains module, and for obtaining training image collection, the training image collection includes the correlation under different target classification Training image, and it is not belonging to the unrelated training image of the target category, the correlation training image carries respective target The corresponding class label of classification；

Confidence determination module, for obtaining the related instruction for the related training image input picture classification learning model Practice the prediction class probability that image is directed to each target category, and the unrelated training image input described image is divided Class learning model obtains the prediction class probability that the unrelated training image is directed to each target category；

Diversity factor determining module, for the corresponding class label of target category according to the related training image, the correlation Training image is directed to the prediction class probability of each target category, determines the target category in the related training image Under true probability distribution prediction probability distribution between probability distribution variances degree；

Uncertainty determining module, it is general for being classified according to the related training image for the prediction of each target category Rate, determines the prediction classification uncertainty of the related training image, and is directed to each institute according to the unrelated training image The prediction class probability for stating target category determines the prediction classification uncertainty of the unrelated training image；

Parameter adjustment module, it is uncertain for being classified according to the prediction of the probability distribution variances degree, the related training image The prediction classification uncertainty of degree and the unrelated training image, adjusts the "current" model ginseng of described image classification learning model Number, so that the probability distribution variances degree determined by the image classification learning model after adjusting, the related instruction determined The prediction classification uncertainty of the prediction classification uncertainty and the unrelated training image determined of practicing image meets pre- If adjusted result condition.

11. a kind of training device of image classification model characterized by comprising processor and memory；The processor and Memory is connected, wherein and the memory is used to call said program code for storing program code, the processor, with Execute method as claimed in any one of claims 1 to 9.

12. a kind of computer readable storage medium, which is characterized in that the computer storage medium is stored with computer program, The computer program includes program instruction, and described program instructs when being executed by a processor, executes such as claim 1 to 9 times Method described in meaning one.