CN110110734A - Opener recognition methods, information processing equipment and storage medium - Google Patents
Opener recognition methods, information processing equipment and storage medium Download PDFInfo
- Publication number
- CN110110734A CN110110734A CN201810103385.2A CN201810103385A CN110110734A CN 110110734 A CN110110734 A CN 110110734A CN 201810103385 A CN201810103385 A CN 201810103385A CN 110110734 A CN110110734 A CN 110110734A
- Authority
- CN
- China
- Prior art keywords
- data
- classification
- discrimination model
- model
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Abstract
Present disclose provides opener recognition methods, information processing equipment and storage mediums.The opener recognition methods includes: that the feature of data to be identified is extracted using trained discrimination model；And the classification of the data to be identified is identified based on extracted feature, wherein, the discrimination model is obtained by following manner: using original auxiliary data, the generation auxiliary data obtained by generation model and for the classification based training data with known class of opener identification, respectively to identify the classification of classification based training data and differentiate that original or generation auxiliary data as task, carries out alternately training to discrimination model.
Description
Technical field
The disclosure relates generally to field of information processing, can refuse in particular to one kind from unknown classification
The opener recognition methods of data, the information processing equipment and storage medium that can be realized the opener recognition methods.
Background technique
In recent years, speech recognition, image recognition for being realized based on machine learning etc. carry out various input datas
The technology of identification (or classification) obtains more and more extensive concern and application.For example, confirming the person of speaking by speech recognition
The voice authentication technology (also referred to as vocal print confirmation) of part can be applied in scenes such as information security, authentications.Known using image
OCR (optical character identification) technology that do not realize can be applied in various documents, to identify text therein.
In the various identification technologies of such as speech recognition and image recognition, it can know using closed set recognition methods and opener
Other method, the classification for the training data that the former uses in training pattern cover number to be processed when identifying using the model
According to all categories, and the classification of training data that the latter uses in training is not covered by all categories of pending data.?
In practical application, often occur needing using opener due to training data is limited, label training data workload is huge etc.
The case where recognition methods.In this case, for the pending data from unknown classification, opener recognition methods should be judged
It is not belonging to known class when training, and classifies to avoid by its mistake.
Accordingly, it is desired to provide a kind of opener recognition methods that can refuse the data from unknown classification.
Summary of the invention
It has been given below about brief overview of the invention, in order to provide about the basic of certain aspects of the invention
Understand.It should be appreciated that this summary is not an exhaustive overview of the invention.It is not intended to determine pass of the invention
Key or pith, nor is it intended to limit the scope of the present invention.Its purpose only provides certain concepts in simplified form,
Taking this as a prelude to a more detailed description discussed later.
In view of the demand improved to existing opener recognition methods, an object of the present invention is to provide a kind of opener
Recognition methods and the information processing equipment and storage medium that can be realized the opener recognition methods can be refused to come from unknown
The data of classification.
According to one aspect of the disclosure, a kind of opener recognition methods is provided comprising: utilize trained differentiation mould
Type extracts the features of data to be identified；And the classification of the data to be identified is identified based on extracted feature, wherein
The discrimination model is obtained by following manner: being assisted using original auxiliary data, by the generation that generation model obtains
Data and classification based training data with known class for opener identification, respectively to identify the classes of classification based training data
Not and differentiate original or generate auxiliary data as task, alternately training is carried out to discrimination model.
According to another aspect of the present disclosure, a kind of opener identification equipment is provided comprising: feature extraction unit is used for
The feature of data to be identified is extracted using trained discrimination model；And classification recognition unit, for based on extracted
Feature identifies the classifications of the data to be identified, wherein the discrimination model is obtained by following manner: using former
Beginning auxiliary data, the generation auxiliary data obtained by generation model and the classification with known class identified for opener
Training data, respectively to identify the classification of classification based training data and differentiate original or generation auxiliary data as task, to differentiation mould
Type carries out alternately training.
According to the disclosure in another aspect, additionally providing a kind of information processing equipment comprising processor, the processor
It is configured as: extracting the feature of data to be identified using trained discrimination model；And known based on extracted feature
The classification of the not described data to be identified, wherein the discrimination model is obtained by following manner: original supplementary number is utilized
According to, the generation auxiliary data that is obtained by generation model and for the classification based training data with known class of opener identification,
Respectively to identify the classification of classification based training data and differentiate that original or generation auxiliary data as task, replaces discrimination model
Training.
According to the other aspects of the disclosure, additionally provides one kind and computer is made to realize voice authentication method as described above
Program.
According to the another aspect of the disclosure, corresponding storage medium is additionally provided, the instruction of machine-readable is stored with
Code, described instruction code enable to machine to execute above-mentioned opener recognition methods when being read by machine and being executed.The finger
Enabling code includes instruction code portion, for carrying out operations described below: data to be identified are extracted using trained discrimination model
Feature；And the classification of the data to be identified is identified based on extracted feature, wherein the discrimination model is to pass through
Following manner obtains: knowing using original auxiliary data, the generation auxiliary data obtained by generation model and for opener
Other classification based training data with known class, respectively to identify the classification of classification based training data and differentiate that original or generation is auxiliary
Helping data is task, carries out alternately training to discrimination model.
The above-mentioned various aspects according to the embodiment of the present disclosure, can at least obtain following benefit: be provided using the disclosure
Opener recognition methods, opener identification equipment, information processing equipment and storage medium, can be using passing through auxiliary data enhancing
The discrimination model of character representation ability extracts and is conducive to opener and knows another characteristic, thus the data to known class into
Refuse the data from unknown classification while row is correctly classified.
By the detailed description below in conjunction with attached drawing to the most preferred embodiment of the disclosure, the these and other of the disclosure is excellent
Point will be apparent from.
Detailed description of the invention
The disclosure can be by reference to being better understood, wherein in institute below in association with description given by attached drawing
Have and has used the same or similar appended drawing reference in attached drawing to indicate same or similar component.The attached drawing is together with following
It is described in detail together comprising in the present specification and forming a part of this specification, and is used to that this is further illustrated
Disclosed preferred embodiment and the principle and advantage for explaining the disclosure.Wherein:
Fig. 1 is the flow chart for schematically showing the example flow of the opener recognition methods according to the embodiment of the present disclosure.
Fig. 2 is one of discrimination model used in the opener recognition methods schematically shown through training acquisition Fig. 1
The flow chart of example process.
Fig. 3 is to illustrate how to obtain showing for discrimination model used in the opener recognition methods of Fig. 1 by training
It is intended to.
Fig. 4 is the flow chart of the example process carried out at the step S203 being shown schematically in Fig. 2.
Fig. 5 is the schematic diagram of an exemplary construction for discrimination model shown in explanatory diagram 3.
Fig. 6 is to show the opener recognition methods of the embodiment of the present disclosure to identify with the opener of the prior art as reference examples
The schematic diagram of the obtained ROC curve of method.
Fig. 7 is the schematic frame for schematically showing the exemplary construction that equipment is identified according to the opener of the embodiment of the present disclosure
Figure.
Fig. 8 be show can be used to realize it is possible according to one kind of the information processing method of the embodiment of the present disclosure and equipment
The structure diagram of hardware configuration.
Specific embodiment
Exemplary embodiment of the invention is described hereinafter in connection with attached drawing.For clarity and conciseness,
All features of actual implementation mode are not described in the description.It should be understood, however, that developing any this actual implementation
Much decisions specific to embodiment must be made during example, to realize the objectives of developer, for example, symbol
Restrictive condition those of related to system and business is closed, and these restrictive conditions may have with the difference of embodiment
Changed.In addition, it will also be appreciated that although development is likely to be extremely complex and time-consuming, to having benefited from the disclosure
For those skilled in the art of content, this development is only routine task.
Here, and also it should be noted is that, in order to avoid having obscured the present invention because of unnecessary details, in the accompanying drawings
Illustrate only with closely related apparatus structure and/or processing step according to the solution of the present invention, and be omitted and the present invention
The little other details of relationship.
Some opener recognition methods are had existed in the prior art.However, these opener recognition methods are assumed mostly from
Know the training data of classification concentrate the feature extracted can Efficient Characterization known class and unknown classification simultaneously data, but in this way
Hypothesis be not always set up.For example, for such as the system of neural network, above-mentioned validity feature may not deposited end to end
So as to cause cannot achieve effective opener recognition methods, and the error in data of unknown classification may be identified as known
Classification.
In this regard, inventor providing a kind of opener recognition methods, the information processing that can be realized the opener recognition methods is set
Standby and storage medium can extract and be conducive to using the discrimination model of character representation ability is enhanced by auxiliary data
Opener knows another characteristic, to refuse the data from unknown classification while correctly being classified to known class.
According to one aspect of the disclosure, a kind of opener recognition methods is provided.Fig. 1 is schematically shown according to this public affairs
Open the flow chart of the example flow of the opener recognition methods 100 of embodiment.As shown in Figure 1, opener recognition methods 100 can wrap
Include: characteristic extraction step S101 extracts the feature of data to be identified using trained discrimination model；And classification identification step
Rapid S103 identifies the classification of the data to be identified based on extracted feature, wherein the discrimination model is by following
Mode obtains: using original auxiliary data, the generation auxiliary data obtained by generation model and for opener identification
Classification based training data with known class, respectively to identify the classification of classification based training data and differentiate original or generation supplementary number
According to for task, alternately training is carried out to discrimination model.
According to the opener recognition methods of the present embodiment, sentencing for character representation ability can be enhanced using by auxiliary data
Other model, extract be conducive to opener know another characteristic, can not only Efficient Characterization known class data, also characterize unknown
The data of classification, so as to refuse the data from unknown classification while correctly being classified to known class.
The opener recognition methods of the present embodiment can be applied to various types of data.As an example, original auxiliary data, point
Class training data and data to be identified can respectively include image data or motor behavior data, also may include in addition to movement is gone
For other one-dimensional datas other than data, such as voice data.
Preferably, original auxiliary data can belong to same area with classification based training data.For example, in classification based training data
When for image data, original auxiliary data may be image data.If original auxiliary data belongs to classification based training data
Completely unrelated field, then extracted feature may be unfavorable for identifying the classification of classification based training data.
The generation model and discrimination model used in above-mentioned opener recognition methods can have various implementations, as long as its
Corresponding functions can be realized.As an example, discrimination model can be model neural network based.Similarly, it generates
Model is also possible to model neural network based.For example, discrimination model and/or generation model can be based on convolutional Neural net
The model of network, such model for example can be particularly advantageous when handling image data.However, those skilled in the art can manage
It solves, the discrimination model and/or generation model applied in the opener recognition methods of the embodiment of the present disclosure are without being limited thereto, but can root
Different neural networks, such as full Connection Neural Network, Recognition with Recurrent Neural Network or recurrent neural network are used according to practical application,
Not reinflated description herein.
It will mainly be described below by image data and discrimination model based on convolutional neural networks and for generating model
The opener recognition methods of the present embodiment.However, it will be understood by those skilled in the art that the opener recognition methods of the embodiment of the present disclosure
It can be applied similarly to other kinds of data and use the model based on other networks.
Now, consider following examples of ancient documents Chinese Character Recognition: with the image of the preparatory Dunhuang ancient documents Chinese character marked
As the classification based training data with known class identified for opener, and using CASIA hand-written data collection as original auxiliary
Data.For trained discrimination model, data to be identified may be the image of Dunhuang ancient documents Chinese character, it is also possible to other texts
Offering text even non-legible image, model will recognize that the former classification, and after refusing the data as unknown classification
Person.Note that as described above, original auxiliary data belongs to same area with classification based training data and is conducive to extract and can characterize
The class another characteristic of classification based training data, but can have certain difference therebetween.For example, although with CASIA in this example
Chinese handwritten data set, but can be with the text image of other language, even non-textual image data as original auxiliary data
As original auxiliary data.For example, the color image data collection of such as CIFAR 10 can be used alternatively.As long as original auxiliary
There is certain correlation between data and classification based training data, advantageous feature can be extracted.
The opener recognition methods and wherein of the present embodiment is further described referring to Fig. 2 to Fig. 5 below in conjunction with the example
The model used.
Describe how to obtain discrimination model used in the opener recognition methods of Fig. 1 with reference first to Fig. 2 to Fig. 4.Fig. 2
It is the stream for schematically showing the example process that discrimination model used in the opener recognition methods of Fig. 1 is obtained by training
Cheng Tu, Fig. 3 are the schematic diagrames illustrated how through the training discrimination model, and Fig. 4 is shown schematically in Fig. 2
Step S203 at carry out an example process flow chart.
As shown in Fig. 2, including alternately iteration progress to identify classification for training the example process 200 of discrimination model
The classification of training data is the step S201 of task training discrimination model, to differentiate that original or generate auxiliary data instructs as task
Practice the step S203 of discrimination model and judges whether the step S205 of convergence.
In the training process of the example process, firstly, in step s 201, as shown in the solid arrow of upside in Fig. 3,
The classification based training data 301 of the Dunhuang ancient documents text for example, marked in advance are input in discrimination model D, differentiation is utilized
Model D identifies it, output classification C1, C2 ..., Cn as recognition result (n is for example, more than or equal to 2 natural number,
The number of whole classifications in its presentation class training data, each classification can correspond to a Chinese character), and based on knowledge
The accuracy of other result updates the parameter of discrimination model D.
Herein, various prior art manners be can use to realize that above-mentioned training and parameter therein update.As an example,
The first-loss function of the accuracy of the classification based on discrimination model identification classification based training data can be set, and be based on the loss
Function updates the parameter of discrimination model, so that the recognition result of discrimination model is as accurate as possible.
For example, the Softmax loss function based on cross entropy can be usedAs above-mentioned
One loss function, wherein SiIt is for example to be connected to after the full articulamentum of the neural network model as discrimination model
I-th of value of Softmax layers of output vector S indicates that training data (or being sample) belongs to the probability of i-th of classification,
yiFor sample label, i value is 1 to N, and N indicates the classification number of all training samples.It can be by making Softmax loss function
Minimize the not reinflated description herein to update the parameter of discrimination model and realize the training in step S201 to discrimination model.
In a preferred embodiment, in addition to the first damage related with classification accuracy of such as Softmax loss function
It loses except function, it can also be in the training of step S201 additionally using unknown losses function to advanced optimize discrimination model
Characteristic present ability.For example, optionally, in the training of step S201, when the classification to identify classification based training data is to appoint
Be engaged in training discrimination model when, the parameter of discrimination model can be updated based on first-loss function and the second loss function, this
The accuracy of classification of one loss function based on discrimination model identification classification based training data, which, which is based on utilizing, sentences
Other model is from the similarity between the feature that each classification based training data are extracted respectively.
As an example, the second loss function based on above-mentioned similarity can be based on each training data (or sample)
The loss function L of the distance between featuredis:
Wherein, f is sample characteristics, and i indicates sample number, and M indicates total number of samples mesh.It can be based on above-mentioned first-loss letter
Number LSoftmaxWith the second loss function LdisTotal loss function L is set:
L=α1Lsoft max+α2LdisFormula (2)
Wherein, α1And α2It indicates weight coefficient, can adjust and be arranged according to experiment.It can be by making total loss function
It minimizes to update the parameter of discrimination model and realize the training in step S201 to discrimination model.Note that although provide herein
First-loss function L based on cross entropySoftmaxWith the second loss function L based on distancedis, but in present disclosure
Under teaching, those skilled in the art can be similarly using other classifications based on discrimination model identification classification based training data
The first-loss function of accuracy and based on using discrimination model between the feature that each classification based training data are extracted respectively
Similarity the second loss function, not reinflated description herein.
In the preferred embodiment, similar between the feature due to additionally having used each classification based training data of expression
Second loss function of degree, be conducive to make it is increasingly similar between the feature of extracted each classification based training data (or these
" distance " is smaller between feature), the data of known class and the data field of unknown classification are separated to be more advantageous to.
It in step s 201 after training,, will for example as shown in the dotted arrow of downside in Fig. 3 in step S203
It is assisted for the original auxiliary data 302 of the image from CASIA hand-written data collection and by generating the generation that model G is generated
Data 303 (such as generating the generation auxiliary data based on stochastic variable) are input to has updated sentencing for parameter in step s 201
In other model D, the two is differentiated using discrimination model D, output differentiates result R ("true") or F ("false") to indicate to sentence
Not Wei original auxiliary data still generate auxiliary data, and continue to update discrimination model D's based on the accuracy of result is differentiated
Parameter.
In a preferred embodiment, can be optimized in step S203 by generation-confrontation type training and more newborn
At both model G and discrimination model D.That is, at the step S203 in alternately training process, it can be by discrimination model original
Differentiated between auxiliary data and the generation auxiliary data obtained by generation model, and generation model and discrimination model are carried out
Dual training.After dual training reception, the data for being difficult to differentiate input are original auxiliary datas or by giving birth to by discrimination model
The generation auxiliary data obtained at model.
It herein, can be using various known for training by the way of generation-confrontation network (GAN) come in step S203
Training generates both model and discrimination model.For example, can construct respectively indicates discrimination model D to being obtained by generation model G
Generate the third loss function L of the discriminant accuracy of auxiliary dataG, and indicate discrimination model D to generation auxiliary data and original
4th loss function L of the discriminant accuracy of both beginning auxiliary datasD, and alternately so that LGThe discriminant accuracy of expression is minimum
Turn to purpose mode of the loss function of more newly-generated model in the case where keeping discrimination model parameter constant be trained, with
So that LDThe discriminant accuracy of expression updates discrimination model in the case where keeping generation model parameter constant for the purpose of maximizing
The mode of loss function be trained, until discrimination model is difficult to differentiate between original assistant images and the life that is obtained by generation model
Until assistant images.
As an example, discrimination model D is to original assistant images x and the generation assistant images G generated using model G is generated
(z) differentiation can be considered as two classification problems, and export the probability value between 0 to 1 as classification results D (x) or D (G
(z)), wherein the probability value presentation class result less than 0.5 is the probability value presentation class for generating assistant images, and being greater than 0.5
It as a result is original assistant images.Correspondingly, loss function l can be constructed1And l2To indicate that discrimination model D carries out two kinds of images
The logistic regression loss (alternatively referred to as LR loss, i.e. Logistic Regression Loss) of classification, that is, l1=log (D
(x)), l2=log (1-D (G (z))).Based on l1And l2To construct the loss function L of discrimination model DD=l1+l2=log (D (x))
+ log (1-D (G (z))), it is more big, indicate that discrimination model D is quasi- to the differentiation for generating both auxiliary data and original auxiliary data
True property is higher, and constructs the loss function L for generating model GG=l2=log (1-D (G (z))), it is smaller, indicate discrimination model D
It is lower to the discriminant accuracy of the generation auxiliary data obtained by generation model G (that is, discrimination model D is obtained by generation model G
Generation auxiliary data confused).It can be based on above-mentioned loss function LDAnd LGOptimize the parameter of two models by training.
Fig. 4 schematically shows the above-mentioned generation carried out at the step S203 for example in Fig. 2-confrontation type training one
The flow chart of a example process.
Specifically, firstly, as shown in the step S401 in Fig. 4, discrimination model D is individually trained.As an example, ought for the first time into
When entering step S401, discrimination model D and the initial parameter for generating model G can be and be randomly provided；In iteration hereafter, with
Initial parameter of the Optimal Parameters that last iteration obtains as model.Training in step S401 is regarded as to two classification
The Training of classifier, without regard to any change for the parameter for generating model G.That is, in the processing of step S401, it will
Original assistant images and marked respectively using the obtained generation assistant images of model G are generated with the label of " R ", " F ", then by these
Image with label, which is input in discrimination model D, classifies.At this point, so that the loss function L of discrimination model DD(its energy
Enough indicating discrimination model D to the discriminant accuracy for generating both auxiliary data and original auxiliary data) acquisitions maximum value is target,
Discrimination model D is trained, and in loss function LDThe Optimal Parameters of discrimination model D are obtained when obtaining maximum value.Here, may be used
With the above-mentioned training for using the existing method for arbitrarily training classifier neural network based to realize to discrimination model D,
This is repeated no more.
After the discrimination model D for obtaining optimization, processing 400 enters next step S403, carries out generating model G herein
With the training of the concatenation network of discrimination model D, is remained unchanged with the parameter in discrimination model D and (remain in step S401 and obtain
The Optimal Parameters obtained) in the case where, the parameter of more newly-generated model G.Due to the ginseng of the only more newly-generated model G of training at this time
Number, thus can also as shown in the step S403 of Fig. 5, referred to as to generate model G training.
In the processing of step S403, the generation assistant images generated by generation model G (are generated into auxiliary figure in this step
As being marked as " R ") it is input in discrimination model D.Model G is generated so as to confuse discrimination model D in order to optimize, it is desirable to
Generate the loss function L of model GG(it can indicate discrimination model D sentencing to the generation auxiliary data obtained by generation model G
Other accuracy) it is the smaller the better.Correspondingly, in step S403, so that loss function LGAcquisition minimum value is target, to generation
The concatenation network of model G and discrimination model D is trained, and in LGThe Optimal Parameters for generating model G are obtained when obtaining minimum value,
Obtain the generation model G of optimization.
It is connected in series since the above-mentioned training to the concatenation network for generating model G and discrimination model D may be considered one
" long network " classifier training, be only based only upon loss function L at this timeGModel is generated to update to constitute in " the long network "
The parameter (parameter for being somebody's turn to do the part those of composition discrimination model D in " long network " remains unchanged) of the part those of G, thus can also
To be considered as the training to two classifier neural network based.It therefore, can be using arbitrary instruction in the processing of step S403
Practice the concatenation network realized based on the existing method of two classifiers of neural network network to model G and discrimination model D is generated
Above-mentioned training, details are not described herein.
After the generation model G optimized, processing 400 can enter next step S405, utilize step herein
The generation model G of the optimization obtained in S403 obtains new generation assistant images.
Then, in step S 407, the discrimination model D based on the optimization obtained in step S401, to the life using optimization
It is distinguished at the obtained generation assistant images of model G and original assistant images, such as both images is exported make respectively
For the probability value of classification results.
Then, in step S409, judge whether the discrimination model D of optimization has been difficult to differentiate between and generate assistant images and original
Assistant images.For example, when discrimination model D both images are exported close to 0.5 probability value as classification results when, then
Think that discrimination model D has been difficult to differentiate between two kinds of images, at this point, the processing 400 of dual training terminates.On the other hand, when optimization
When discrimination model D can still distinguish two kinds of images, processing 400 returns to step S401, and optimization ginseng is obtained in the above an iteration
Number carries out the processing of step S401 to S409 as initial parameter, until judging that discrimination model G is difficult to area in step S409
It is mitogenetic at assistant images and original assistant images until, processing 400 terminate.
Via above-mentioned dual training, can use and generate model and obtain generating auxiliary data, and generate auxiliary data with
Original auxiliary data is used for the training of discrimination model to enhance its characteristic present ability, without additionally to auxiliary data together
It is labeled with the training for discrimination model.In other words, this preferred embodiment can be to avoid additionally increasing cumbersome time-consuming mark
Work is infused, while enhancing the feature obtained using discrimination model to the characterization ability of unknown data.
In step S203 after training, in step S205, judge whether discrimination model D has restrained, and in model
Terminate to train when having restrained；Otherwise, step S201 is returned to, and repeats the processing of step S201 to step S205, Zhi Daomo
Until type restrains.As an example, can be by judging the parameter of updated discrimination model D and step S201 in step S203
In the parameter of updated discrimination model D whether essentially identical judge whether discrimination model D restrains: if above-mentioned two step
In updated discrimination model D parameter it is essentially identical, then may determine that discrimination model D has restrained.It as an alternative, can be by sentencing
The loss function related with classification accuracy (such as above-mentioned first of discrimination model D after disconnected undated parameter in step s 201
Loss function LSoftmaxOr total loss function L) whether essentially identical with the corresponding loss function in last iteration and judge
The loss function related with original or generation auxiliary data is differentiated of discrimination model D in step S203 after undated parameter is (all
Such as above-mentioned loss function LD) whether essentially identical with the corresponding loss function in last iteration, to judge whether discrimination model D receives
It holds back.If above-mentioned two loss function is essentially identical respectively in iteration twice, it may determine that discrimination model D has restrained.
The discrimination model obtained using above-mentioned training, enhances the character representation ability of discrimination model by auxiliary data,
Can extract be conducive to opener know another characteristic, can simultaneously Efficient Characterization known class and unknown class another characteristic, from
And refuse the data from unknown classification while correctly being classified to known class.
Fig. 5 shows the discrimination model used in the training described above by reference to Fig. 2 to Fig. 4 based on convolutional neural networks
An exemplary construction.As shown in figure 5, the convolutional neural networks discrimination model 500 include input layer 501, be arranged alternately two
A convolutional layer 502 and 504 and two pond layers 503 and 505 and the full articulamentum of full articulamentum i.e. first for being divided into two parts
506 and the second full articulamentum 507.When training, such as at step S201 shown in Fig. 2, classification based training data are input to
Input layer 501, with the processing via each convolutional layer and pond layer the first full articulamentum 506 obtain it is all as shown in Figure 3
Classification C1, C2 ..., the recognition result of Cn, and the accuracy based on recognition result update in addition to the second full articulamentum 507 with
The parameter of outer each layer.In addition, original or auxiliary generation data are input to defeated for example at step S203 shown in Fig. 2
Enter layer 501, R or F in such as Fig. 3 are obtained in the second full articulamentum 507 with the processing via each convolutional layer and pond layer
Differentiation as a result, and based on differentiating that the accuracy of result updates the parameter of each layer other than the first full articulamentum 506.Alternately
Above-mentioned training is carried out, until model convergence, the convolutional neural networks discrimination model that may finally be optimized.
Note that although Fig. 5 is shown including an input layer, two full articulamentums, two pond layers, two full articulamentums
Convolutional neural networks structure, it will be appreciated by a person skilled in the art that being answered in the opener recognition methods of the embodiment of the present disclosure
The structure of discrimination model based on convolutional neural networks is without being limited thereto, but can be set according to practical application to carry out difference
Meter, such as may include different number of convolutional layer, pond layer or full articulamentum, and such as regularization layer can be also comprised
Deng other layers.
As an example, (being not limited to convolution shown in fig. 5 using discrimination model neural network based such as shown in fig. 5
Neural network), extract shown in the step S101 of example as shown in figure 1 the feature of data to be identified processing may include: by
Identification data are input to the neural network as trained discrimination model, extract the output number of a layer of the neural network
According to feature as data to be identified.
Preferably, said one layer can be the pond layer of neural network.For example, discrimination model shown in Fig. 5 shows
In example, be extracted feature can be the tightly pond layer before full articulamentum.The pond layer is for classification based training for identification
The full articulamentum of the first of data and for differentiate it is original or generate auxiliary data the second full articulamentum be public, that is, from
This feature that this layer extracts can simultaneously effective characterize the feature of the classification based training data of known class and as unknown classification
The feature of the auxiliary data of data.Therefore, the feature being achieved in that is advantageously implemented effective opener identification.Certainly, the disclosure
Embodiment is without being limited thereto, and can be from other layers of neural network (such as convolutional layer or other ponds of convolutional neural networks
Layer etc.) extract feature.
In one example, spy is being extracted from data to be identified using trained discrimination model for example shown in fig. 5
After sign, the processing of the classification of identification data to be identified may include: that will be extracted shown in the step S103 of example as shown in figure 1
Feature respectively with using trained discrimination model from the feature extracted in the classification based training data of each classification compared with, with
Identify the classification of data to be identified.
For example, if the feature of data to be identified and the feature extracted from the classification based training data of one of classification it
Between similarity be greater than predetermined threshold, then be identified as the category；Otherwise, that is, if this feature with from any one classification
Similarity is respectively less than predetermined threshold between the feature extracted in classification based training data, then is identified as unknown classification.As an alternative,
Can calculate separately the features of data to be identified to from similar between the feature extracted in the classification based training data of each classification
Degree, and by maximum similarity obtained compared with predetermined threshold.If maximum similarity is greater than predetermined threshold, will be to be identified
Data are identified as classification corresponding with maximum similarity；Otherwise, data to be identified are identified as unknown classification.
As an example, the average value for the feature extracted from all classification based training data of a classification can be calculated, make
For the average characteristics (MF) of the training data of the category, with the calculating for above-mentioned similarity.Differentiation is utilized at this point it is possible to calculate
Model from the feature that data to be identified the are extracted similarity between the average characteristics of the training data of each classification respectively, as
Its similarity between the feature of the training data of the category.As an example, euclidean distance metric can be used to characterize
State similarity.It will be understood by those skilled in the art that other measurements such as COS distance measurement, joint shellfish can also be applied similarly
Ye Si measurement or a combination thereof comes the similarity between characteristic feature, herein not reinflated description.
Using the opener recognition methods of the present embodiment described above by reference to Fig. 1 to Fig. 5, it can utilize and pass through auxiliary data
The discrimination model of character representation ability is enhanced, extracts and is conducive to opener knowledge another characteristic, to be carried out to known class
Refuse the data from unknown classification while correct classification.
It describes to carry out the beneficial of opener identification acquisition using the opener recognition methods of the embodiment of the present disclosure referring to Fig. 6
Effect.Fig. 6 is to show opener recognition methods and the opener identification side of the prior art as reference examples of the embodiment of the present disclosure
The schematic diagram of the obtained ROC curve of method, wherein curve 601 shows the opener recognition methods using the embodiment of the present disclosure
ROC curve, curve 602 show the ROC curve of the art methods as comparative example.
In two methods, 300 classes (that is, indicating about 300 Chinese characters) of Dunhuang ancient documents text have equally been used
About 40,000 training samples in addition used as classification based training data, and in the opener recognition methods of the present embodiment
About 100,000 original auxiliary datas of about 4000 classes (that is, indicating about 4000 Chinese characters) from CASIA hand-written data collection.More
Further, in the method for the present embodiment, use about 40,000 training samples of 300 classes in the ancient documents of Dunhuang as
Classification based training data and about 100,000 samples from hand-written data collection are as original auxiliary data, such as extremely above by reference to Fig. 1
Training discrimination model common like that described in Fig. 5；In the art methods of reference examples, it is used alone in the ancient documents of Dunhuang
About 40,000 training samples of above-mentioned 300 classes are individually trained in reference examples in the way of the general classifier of the prior art
Discrimination model (that is, not using any auxiliary data).After training, disclosure reality is similarly used in a manner of previously described
The art methods of the method and reference examples of applying example identify test sample.Namely based on the test extracted using discrimination model
In similarity between the feature of the Dunhuang ancient documents text of the feature of sample and each classification for utilizing discrimination model to extract
Maximum similarity identify the classification of test sample, i.e., whether be greater than by the maximum similarity judged in above-mentioned similarity pre-
Threshold value is determined to identify whether pending data belongs to respective classes.If the feature of test sample and the Dunhuang Ancient from each classification
The maximum similarity between feature extracted in document text is greater than predetermined threshold, then is identified as test sample similar to maximum
Spend the classification of that corresponding Dunhuang ancient documents text；Otherwise, test sample is identified as unknown classification.By in similarity
Change above-mentioned predetermined threshold in entire threshold range, then can obtain entire ROC curve.
ROC curve is also referred to as recipient's operating characteristics (receiver operating characteristic) curve,
On each point reflect the sensitivity to same signal stimulus, in the opener identification problem discussed here, it is each point reflection
For the accuracy of classification or the identification of a given similarity threshold.More specifically, on ROC curve each point abscissa table
Show false positive rate corresponding with a given similarity threshold, ordinate indicates real ratio corresponding with the given threshold value
Rate, entire ROC curve indicate false positive rate corresponding with entire similarity threshold range and real ratio.
Herein, defining true positives indicates that the sample of unknown classification is correctly identified as unknown classification, known to false positive expression
The sample of classification is mistakenly identified as unknown classification, and false negative indicates that the sample of unknown classification is mistakenly identified as known class
Not, true negative indicates that the sample of known class is correctly identified as known class.Abscissa is the false positive rate of opener identification,
Indicate the ratio for being erroneously identified as unknown classification but the practical sample that all known class are accounted for for known sample, i.e. FP/N=
FP/ (FP+TN), wherein FP is false positive number, and TN is true negative number, and N=FP+TN is negative sum.Opener identification is asked
Topic, false positive rate FPR are better closer to 0.Ordinate shows the real ratio (TPR) of opener identification, indicates to be correctly identified as
The ratio of unknown classification and the practical sample that all unknown classifications are accounted for for unknown sample, i.e. TP/P=TP/ (TP+FN), wherein
TP is true positives number, and FN is false negative number, and P=TP+FN is positive sum.Problem, real ratio (TPR) are identified for opener
It is better closer to 1.
As can be seen that the ROC curve 601 of the opener recognition methods of the present embodiment is compared to right from two curves of Fig. 6
The real ratio TPR of ROC curve 602 as usual is closer to 1, and vacation positive rate FPR can be identified more closer to 0
Unknown classification.In other words, the opener recognition methods of the present embodiment can preferably refuse unknown classification compared to the prior art
Data avoid being mistaken for known class.
Opener recognition methods according to an embodiment of the present disclosure is described above by reference to Fig. 1 to Fig. 6.It is identified using the opener
Method can extract using the discrimination model of character representation ability is enhanced by auxiliary data and be conducive to opener identification
Feature, can not only Efficient Characterization known class data, characterize the data of unknown classification, also so as to known class
Refuse the data from unknown classification while classification correctly.
It will be understood by those skilled in the art that in addition to the image for text in document described herein as example is known
Except not, the opener recognition methods that can refuse the present embodiment of the data from unknown classification handles (example for general pattern
Such as recognition of face), speech recognition (such as speaker's identity certification), motor behavior identify (such as recognizing model of movement) field
All have wide application prospect.
In addition, this method can also be applied to a lot of other fields.For example, for there may be the electric system numbers of failure
According to, motor data, chemical process data etc., training data when training can not cover all fault categories or only cover part therefore
Hinder classification, but need to identify the failure of these unknown classifications in application, in this case the opener identification of the embodiment of the present disclosure
Method will be particularly useful.It can be based on the opener recognition methods of the embodiment of the present disclosure, using using original supplementary number appropriate
According to the differentiation mould of (such as the partial fault data that have obtained or other data for belonging to normal data same area) training
Type, to differentiate the fault data of unknown classification.
Similarly, in electric business field, system needs to detect abnormal transaction row generally to normal e-commerce data modeling
For and/or fraud, that is, need various abnormal datas unknown when recognition training (abnormal behaviour for identifying unknown classification),
In this case the opener recognition methods of the embodiment of the present disclosure can also be applied.At this point it is possible to opening based on the embodiment of the present disclosure
Set identification method, using using original auxiliary data appropriate (such as the part abnormal data obtained or and normal data
Belong to other data of same area) training discrimination model, to differentiate the abnormal data of unknown classification.
It is meant only to more fully understand the embodiment of the present disclosure convenient for those skilled in the art note that providing above-mentioned application example,
And it is not intended to constitute any restrictions to the embodiment of the present disclosure or its application field or scene.Those skilled in the art can be according to reality
Border needs to carry out the embodiment of the present disclosure various modifications appropriate and/or change, and is applied to any field appropriate and field
Scape.
According to root another aspect of the present disclosure, a kind of opener identification equipment is provided.Fig. 7 is schematically shown according to this
The schematic block diagram of the exemplary construction of the opener identification equipment of open embodiment.
As shown in fig. 7, opener identification equipment 700 may include: feature extraction unit 701, for being sentenced using trained
Other model extracts the features of data to be identified；And classification recognition unit 702, for based on extracted feature to identify
State the classification of data to be identified, wherein the discrimination model is obtained by following manner: using original auxiliary data,
The generation auxiliary data obtained by generation model and the classification based training data with known class for opener identification, point
Using identify classification based training data classification and differentiate it is not original or generate auxiliary data as task, to discrimination model carry out alternately instruction
Practice.
Above-mentioned opener identification equipment and its each unit can for example carry out the opener described above by reference to Fig. 1 to Fig. 6 knowledge
The operation and/or processing of other method and its each step simultaneously realize similar effect, no longer carry out repeated explanation herein.
According to basic disclosed another aspect, a kind of information processing equipment is provided.The information processing equipment may be implemented
It may include processor according to the opener recognition methods of the embodiment of the present disclosure, which is configured as: utilizing trained
Discrimination model extracts the features of data to be identified；And the class of the data to be identified is identified based on extracted feature
Not, wherein the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
Auxiliary data and the classification based training data with known class for opener identification are generated, respectively to identify classification based training
The classification of data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
The processor of information processing equipment, which for example can be configured as the opener describe above by reference to Fig. 1 to Fig. 6, to be known
The operation and/or processing of other method and its each step simultaneously realize similar effect, no longer carry out repeated explanation herein.
As an example, original auxiliary data, classification based training data and data to be identified can respectively include image data or
Motor behavior data.
As an example, discrimination model can be model neural network based.Similarly, model is generated to be also possible to be based on
The model of neural network.For example, discrimination model and/or generation model can be the model based on convolutional neural networks.
It in a preferred embodiment, can be by discrimination model in original auxiliary data during alternately training
Differentiated between the generation auxiliary data obtained by generation model, and confrontation instruction is carried out to generation model and discrimination model
Practice.
Optionally, in addition, during alternately training, when the classification to identify classification based training data is task to differentiation
When model is trained, the parameter of discrimination model can be updated based on first-loss function and the second loss function, this first
The accuracy of classification of the loss function based on discrimination model identification classification based training data, second loss function are based on utilizing differentiation
Model is from the similarity between the feature that each classification based training data are extracted respectively.
As an example, processor can be configured as through following manner the feature for extracting data to be identified: will be wait know
Other data are input to the neural network as trained discrimination model, extract the output data of a layer of the neural network,
Feature as data to be identified.
Preferably, said one layer can be the pond layer of neural network.
In addition, as an example, processor can be configured as by come the classification that identifies data to be identified: will be extracted
Feature respectively with using trained discrimination model from the feature extracted in the classification based training data of each classification compared with, with know
The classification of data not to be identified.
Fig. 8 be show can be used to realize it is possible according to one kind of the information processing equipment of the embodiment of the present disclosure and method
The structure diagram of hardware configuration 800.
In fig. 8, central processing unit (CPU) 801 is according to the program stored in read-only memory (ROM) 802 or from depositing
The program that storage part 808 is loaded into random access memory (RAM) 803 executes various processing.In RAM 803, also according to need
Store the data required when CPU 801 executes various processing etc..CPU 801, ROM 802 and RAM 803 are via bus
804 are connected to each other.Input/output interface 805 is also connected to bus 804.
Components described below is also connected to input/output interface 805: importation 806 (including keyboard, mouse etc.), output
Part 807 (including display, such as cathode-ray tube (CRT), liquid crystal display (LCD) etc. and loudspeaker etc.), storage section
808 (including hard disks etc.), communications portion 809 (including network interface card such as LAN card, modem etc.).Communications portion 809
Communication process is executed via network such as internet.As needed, driver 810 can be connected to input/output interface 805.
Detachable media 811 such as disk, CD, magneto-optic disk, semiconductor memory etc., which can according to need, is installed in driver
On 810, so that the computer program read out can be mounted to as needed in storage section 808.
In addition, the disclosure also proposed a kind of program product of instruction code for being stored with machine-readable.Above-metioned instruction
When code is read and executed by machine, the above-mentioned voice authentication method according to the embodiment of the present disclosure can be performed.Correspondingly, for holding
The various storage mediums such as disk, CD, magneto-optic disk, semiconductor memory for carrying this program product are also included within this public affairs
In the disclosure opened.
That is, the disclosure also proposed a kind of storage medium, it is stored with the instruction code of machine-readable, described instruction generation
Code enables to machine to execute the above-mentioned voice authentication method according to the embodiment of the present disclosure when being read by machine and being executed.Institute
Stating instruction code includes instruction code portion, for carrying out operations described below: being extracted using trained discrimination model to be identified
The feature of data；And the classification of the data to be identified is identified based on extracted feature, wherein the discrimination model is
It is obtained by following manner: using original auxiliary data, the generation auxiliary data obtained by generation model and for opening
Collect the classification based training data with known class of identification, respectively to identify the classification of classification based training data and differentiate original or raw
It is task at auxiliary data, alternately training is carried out to discrimination model.
Above-mentioned storage medium for example can include but is not limited to disk, CD, magneto-optic disk, semiconductor memory etc..
In the description above to disclosure specific embodiment, for the feature a kind of embodiment description and/or shown
It can be used in one or more other embodiments in a manner of same or similar, with the feature in other embodiment
It is combined, or the feature in substitution other embodiment.
In addition, the method for the presently disclosed embodiments be not limited to specifications described in or it is shown in the accompanying drawings when
Between sequentially execute, can also be according to other time sequencings, concurrently or independently execute.Therefore, it is described in this specification
Method execution sequence scope of the presently disclosed technology is not construed as limiting.
It should be further understood that can also can be stored in various machines according to each operating process of the above method of the disclosure
The mode of computer executable program in the storage medium of reading is realized.
Moreover, the purpose of the disclosure can also be accomplished in the following manner: above-mentioned executable program code will be stored with
Storage medium is directly or indirectly supplied to system or equipment, and computer or central processing in the system or equipment
Unit (CPU) reads and executes above procedure code.
As long as embodiment of the present disclosure is not limited at this point, the system or equipment have the function of executing program
Program, and the program is also possible to arbitrary form, for example, program that target program, interpreter execute or being supplied to behaviour
Make the shell script etc. of system.
These above-mentioned machine readable storage mediums include but is not limited to: various memories and storage unit, semiconductor equipment,
Disk cell such as light, magnetic and magneto-optic disk and other media etc. for being suitable for storing information.
In addition, customer information processing terminal is by the corresponding website that is connected on internet, and will be according to the disclosure
Computer program code is downloaded and is installed in the information processing terminal and then executes the program, and each reality of the disclosure also may be implemented
Apply example.
To sum up, according to the embodiment of the present disclosure, present disclose provides following scheme, but not limited to this:
1. a kind of information processing equipment, the equipment include:
Processor is configured as:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
The generation auxiliary data arrived and the classification based training data with known class for opener identification, respectively to identify classification
The classification of training data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
2. information processing equipment according to scheme 1, wherein discrimination model is model neural network based.
3. the information processing equipment according to scheme 1 or 2, wherein during alternately training, pass through discrimination model
Differentiated between original auxiliary data and the generation auxiliary data obtained by generation model, and to generation model and differentiates mould
Type carries out dual training.
4. the information processing equipment according to scheme 1 or 2, wherein processor is configured as knowing by following manner
The classification of data not to be identified: extracted feature is instructed with using trained discrimination model from the classification of each classification respectively
Practice the feature extracted in data to compare, to identify the classification of data to be identified.
5. the information processing equipment according to scheme 1 or 2, wherein original auxiliary data, classification based training data and wait know
Other data respectively include image data or motor behavior data.
6. according to information processing equipment described in scheme 2, wherein during alternately training, when with identification classification instruction
When the classification for practicing data is that task is trained discrimination model, sentenced based on first-loss function and the second loss function to update
The parameter of other model, the accuracy of classification of the first-loss function based on discrimination model identification classification based training data, this second
Loss function based on using discrimination model from the similarity between the feature that each classification based training data are extracted respectively.
7, the information processing equipment according to scheme 2, wherein processor be configured as extracting by following manner to
It identifies the feature of data: data to be identified being input to the neural network as trained discrimination model, extract the nerve net
The output data of one layer of network, the feature as data to be identified.
8, information processing equipment according to scheme 7, wherein one layer is the pond layer of neural network.
9. a kind of opener recognition methods, comprising:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
The generation auxiliary data arrived and the classification based training data with known class for opener identification, respectively to identify classification
The classification of training data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
10. opener recognition methods according to scheme 9, wherein discrimination model is model neural network based.
11. the opener recognition methods according to scheme 9 or 10, wherein during alternately training, by differentiating mould
Type differentiated between original auxiliary data and the generation auxiliary data obtained by generation model, and to generating model and differentiation
Model carries out dual training.
12. the opener recognition methods according to scheme 9 or 10, wherein the classification for identifying data to be identified includes: by institute
The feature of extraction is respectively compared with the feature extracted from the classification based training data of each classification using trained discrimination model
Compared with to identify the classification of data to be identified.
13. the opener recognition methods according to scheme 9 or 10, wherein original auxiliary data, classification based training data and to
Identification data respectively include image data or motor behavior data.
14. opener recognition methods according to scheme 10, wherein during alternately training, when to identify classification
When the classification of training data is that task is trained discrimination model, updated based on first-loss function and the second loss function
The parameter of discrimination model, the accuracy of classification of the first-loss function based on discrimination model identification classification based training data, this
Two loss functions based on using discrimination model from the similarity between the feature that each classification based training data are extracted respectively.
15, opener recognition methods according to scheme 10, wherein extract data to be identified feature include: will be wait know
Other data are input to the neural network as trained discrimination model, extract the output data of a layer of the neural network,
Feature as data to be identified.
16, opener recognition methods according to scheme 15, wherein one layer is the pond layer of neural network.
17, a kind of storage medium, is stored with the instruction code of machine-readable, and described instruction code is read by machine
And when executing, machine is enabled to execute a kind of opener recognition methods, described instruction code includes:
Instruction code portion, for carrying out operations described below:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
The generation auxiliary data arrived and the classification based training data with known class for opener identification, respectively to identify classification
The classification of training data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
Finally, it is to be noted that, in the disclosure, relational terms such as first and second and the like are used merely to
It distinguishes one entity or operation from another entity or operation, without necessarily requiring or implying these entities or behaviour
There are any actual relationship or orders between work.Moreover, the terms "include", "comprise" or its any other variant
It is intended to non-exclusive inclusion, so that the process, method, article or equipment including a series of elements may be not only
It including those elements, but also including other elements that are not explicitly listed, or further include for this process, method, object
Product or the intrinsic element of equipment.In the absence of more restrictions, the element limited by sentence "including a ...",
It is not precluded in the process, method, article or apparatus that includes the element that there is also other identical elements.
Although being had been disclosed above by the description of the specific embodiment of the disclosure to the disclosure, however, it should
Understand, those skilled in the art can design the various modifications to the disclosure in the spirit and scope of the appended claims, improve
Or equivalent.These modifications, improvement or equivalent should also be as being to be considered as included in disclosure range claimed.
Claims (10)
1. a kind of information processing equipment, the equipment include:
Processor is configured as:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
Auxiliary data and the classification based training data with known class for opener identification are generated, respectively to identify classification based training
The classification of data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
2. information processing equipment according to claim 1, wherein discrimination model is model neural network based.
3. information processing equipment according to claim 1 or 2, wherein during alternately training, pass through discrimination model
Differentiated between original auxiliary data and the generation auxiliary data obtained by generation model, and to generation model and differentiates mould
Type carries out dual training.
4. information processing equipment according to claim 1 or 2, wherein processor is configured as knowing by following manner
The classification of data not to be identified: extracted feature is instructed with using trained discrimination model from the classification of each classification respectively
Practice the feature extracted in data to compare, to identify the classification of data to be identified.
5. information processing equipment according to claim 1 or 2, wherein original auxiliary data, classification based training data and wait know
Other data respectively include image data or motor behavior data.
6. information processing equipment according to claim 2, wherein during alternately training, when with identification classification instruction
When the classification for practicing data is that task is trained discrimination model, sentenced based on first-loss function and the second loss function to update
The parameter of other model, the accuracy of classification of the first-loss function based on discrimination model identification classification based training data, this second
Loss function based on using discrimination model from the similarity between the feature that each classification based training data are extracted respectively.
7. information processing equipment according to claim 2, wherein processor be configured as extracting by following manner to
It identifies the feature of data: data to be identified being input to the neural network as trained discrimination model, extract the nerve net
The output data of one layer of network, the feature as data to be identified.
8. information processing equipment according to claim 7, wherein one layer is the pond layer of neural network.
9. a kind of opener recognition methods, comprising:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
Auxiliary data and the classification based training data with known class for opener identification are generated, respectively to identify classification based training
The classification of data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
10. a kind of storage medium, is stored with the instruction code of machine-readable, described instruction code is being read by machine and is being held
When row, machine is enabled to execute a kind of opener recognition methods, described instruction code includes:
Instruction code portion, for carrying out operations described below:
The feature of data to be identified is extracted using trained discrimination model；And
The classification of the data to be identified is identified based on extracted feature,
Wherein, the discrimination model is obtained by following manner: being obtained using original auxiliary data, by generation model
Auxiliary data and the classification based training data with known class for opener identification are generated, respectively to identify classification based training
The classification of data and differentiation are original or generate auxiliary data as task, carry out alternately training to discrimination model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103385.2A CN110110734B (en) | 2018-02-01 | 2018-02-01 | Open set identification method, information processing apparatus, and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810103385.2A CN110110734B (en) | 2018-02-01 | 2018-02-01 | Open set identification method, information processing apparatus, and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110110734A true CN110110734A (en) | 2019-08-09 |
| CN110110734B CN110110734B (en) | 2023-04-07 |
Family
ID=67483583
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810103385.2A Active CN110110734B (en) | 2018-02-01 | 2018-02-01 | Open set identification method, information processing apparatus, and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110734B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110704618A (en) * | 2019-09-20 | 2020-01-17 | 阿里巴巴集团控股有限公司 | Method and device for determining standard problem corresponding to dialogue data |
| CN112017670A (en) * | 2020-08-13 | 2020-12-01 | 北京达佳互联信息技术有限公司 | Target account audio identification method, device, equipment and medium |
| CN112508062A (en) * | 2020-11-20 | 2021-03-16 | 普联国际有限公司 | Open set data classification method, device, equipment and storage medium |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003168113A (en) * | 2001-12-04 | 2003-06-13 | Nec Corp | System, method and program of image recognition |
| US20040170318A1 (en) * | 2003-02-28 | 2004-09-02 | Eastman Kodak Company | Method for detecting color objects in digital images |
| US20040208390A1 (en) * | 2003-04-18 | 2004-10-21 | Medispectra, Inc. | Methods and apparatus for processing image data for use in tissue characterization |
| CN103646244A (en) * | 2013-12-16 | 2014-03-19 | 北京天诚盛业科技有限公司 | Methods and devices for face characteristic extraction and authentication |
| US20140079297A1 (en) * | 2012-09-17 | 2014-03-20 | Saied Tadayon | Application of Z-Webs and Z-factors to Analytics, Search Engine, Learning, Recognition, Natural Language, and Other Utilities |
| US20150254555A1 (en) * | 2014-03-04 | 2015-09-10 | SignalSense, Inc. | Classifying data with deep learning neural records incrementally refined through expert input |
| CN105913081A (en) * | 2016-04-08 | 2016-08-31 | 西安电子科技大学 | Improved PCAnet-based SAR image classification method |
| CN106815566A (en) * | 2016-12-29 | 2017-06-09 | 天津中科智能识别产业技术研究院有限公司 | A kind of face retrieval method based on multitask convolutional neural networks |
| US20170262737A1 (en) * | 2016-03-11 | 2017-09-14 | Magic Leap, Inc. | Structure learning in convolutional neural networks |
| WO2017175282A1 (en) * | 2016-04-04 | 2017-10-12 | オリンパス株式会社 | Learning method, image recognition device, and program |
-
2018
- 2018-02-01 CN CN201810103385.2A patent/CN110110734B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003168113A (en) * | 2001-12-04 | 2003-06-13 | Nec Corp | System, method and program of image recognition |
| US20040170318A1 (en) * | 2003-02-28 | 2004-09-02 | Eastman Kodak Company | Method for detecting color objects in digital images |
| US20040208390A1 (en) * | 2003-04-18 | 2004-10-21 | Medispectra, Inc. | Methods and apparatus for processing image data for use in tissue characterization |
| US20140079297A1 (en) * | 2012-09-17 | 2014-03-20 | Saied Tadayon | Application of Z-Webs and Z-factors to Analytics, Search Engine, Learning, Recognition, Natural Language, and Other Utilities |
| CN103646244A (en) * | 2013-12-16 | 2014-03-19 | 北京天诚盛业科技有限公司 | Methods and devices for face characteristic extraction and authentication |
| WO2015090126A1 (en) * | 2013-12-16 | 2015-06-25 | 北京天诚盛业科技有限公司 | Facial characteristic extraction and authentication method and device |
| US20150254555A1 (en) * | 2014-03-04 | 2015-09-10 | SignalSense, Inc. | Classifying data with deep learning neural records incrementally refined through expert input |
| US20170262737A1 (en) * | 2016-03-11 | 2017-09-14 | Magic Leap, Inc. | Structure learning in convolutional neural networks |
| WO2017175282A1 (en) * | 2016-04-04 | 2017-10-12 | オリンパス株式会社 | Learning method, image recognition device, and program |
| CN105913081A (en) * | 2016-04-08 | 2016-08-31 | 西安电子科技大学 | Improved PCAnet-based SAR image classification method |
| CN106815566A (en) * | 2016-12-29 | 2017-06-09 | 天津中科智能识别产业技术研究院有限公司 | A kind of face retrieval method based on multitask convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| YUJIE LIU ET AL: "Car Plate Character Recognition Using a Convolutional Neural Network with Shared Hidden Layers", 《2015 CHINESE AUTOMATION CONGRESS (CAC)》 * |
| ZILI YI ET AL: "DualGAN: Unsupervised Dual Learning for Image-to-Image Translation", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION》 * |
| ZONGYUAN GERMANIUM ET AL: "Generative OpenMax for Multi-Class Open Set Classification", 《HTTPS://ARXIV.ORG/PDF/1707.07418.PDF》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110704618A (en) * | 2019-09-20 | 2020-01-17 | 阿里巴巴集团控股有限公司 | Method and device for determining standard problem corresponding to dialogue data |
| CN110704618B (en) * | 2019-09-20 | 2023-06-27 | 创新先进技术有限公司 | Method and device for determining standard problem corresponding to dialogue data |
| CN112017670A (en) * | 2020-08-13 | 2020-12-01 | 北京达佳互联信息技术有限公司 | Target account audio identification method, device, equipment and medium |
| CN112017670B (en) * | 2020-08-13 | 2021-11-02 | 北京达佳互联信息技术有限公司 | Target account audio identification method, device, equipment and medium |
| CN112508062A (en) * | 2020-11-20 | 2021-03-16 | 普联国际有限公司 | Open set data classification method, device, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110110734B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109993236B (en) | One-shot Simese convolutional neural network-based small-sample Manchu matching method | |
| KR101446376B1 (en) | Identification and verification of an unknown document according to an eigen image process | |
| US7360686B2 (en) | Method and system for discovering significant subsets in collection of documents | |
| CN109784325A (en) | Opener recognition methods and equipment and computer readable storage medium | |
| CN109960727B (en) | Personal privacy information automatic detection method and system for unstructured text | |
| CN112257441B (en) | Named entity recognition enhancement method based on counterfactual generation | |
| CN110110734A (en) | Opener recognition methods, information processing equipment and storage medium | |
| CN110084609B (en) | Transaction fraud behavior deep detection method based on characterization learning | |
| CN108681735A (en) | Optical character recognition method based on convolutional neural networks deep learning model | |
| CN110164455A (en) | Device, method and the storage medium of user identity identification | |
| CN109446345A (en) | Nuclear power file verification processing method and system | |
| CN114329034A (en) | Image text matching discrimination method and system based on fine-grained semantic feature difference | |
| CN108681532B (en) | Sentiment analysis method for Chinese microblog | |
| CN114707571A (en) | Credit data anomaly detection method based on enhanced isolation forest | |
| CN109145704A (en) | A kind of human face portrait recognition methods based on face character | |
| CN110349585A (en) | Voice authentication method and information processing equipment | |
| WO2023070843A1 (en) | Multi-modal currency recognition system, method, storage medium and terminal | |
| CN112732908B (en) | Test question novelty evaluation method and device, electronic equipment and storage medium | |
| Muhtar et al. | A Survey of Offline Handwritten Signature Verification Based on Deep Learning | |
| CN109767430A (en) | The quality determining method and quality detecting system of valuable bills | |
| Mehta et al. | Optical music notes recognition for printed piano music score sheet | |
| CN117648409B (en) | OCR-based format file anti-counterfeiting recognition method | |
| CN117253097B (en) | Semi-supervision domain adaptive image classification method, system, equipment and storage medium | |
| CN117764160A (en) | Training method and device for relation extraction model, electronic equipment and storage medium | |
| CN113988226B (en) | Data desensitization validity verification method and device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |