CN108229692A - A kind of machine learning recognition methods based on double contrast's study - Google Patents
A kind of machine learning recognition methods based on double contrast's study Download PDFInfo
- Publication number
- CN108229692A CN108229692A CN201810128018.8A CN201810128018A CN108229692A CN 108229692 A CN108229692 A CN 108229692A CN 201810128018 A CN201810128018 A CN 201810128018A CN 108229692 A CN108229692 A CN 108229692A
- Authority
- CN
- China
- Prior art keywords
- sample
- contrast
- machine learning
- learning model
- input
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of machine learning recognition methods based on double contrast's study, and a certain amount of multi-medium data sample of known class can be utilized to put in order using different contrast sample's inputs to machine learning model f1The learning training of multiple differentiated is carried out, to carry out multi-medium data classification identifying processing, machine learning model f1It is designed as the built-up pattern framework of convolutional neural networks model or/and full Connection Neural Network model, significantly reduce the dependence to magnanimity training sample, and it can easily extend to carrying out classification identification without the multi-medium data classification of learning training, existing multimedia data classification machine learning recognition methods is solved well because the dependence to a large amount of training samples and because leading to practical application when directly can not carry out Classification and Identification to the classification without learning training, the problem of versatility is limited, it more can effectively be applied to extensively in more specific multimedia data classification use occasions.
Description
Technical field
The present invention relates to multimedia data processing and machine learning techniques field more particularly to one kind based on dual right
Than the machine learning recognition methods of study.
Background technology
Multimedia (Multimedia) is the synthesis of media, in computer systems, multimedia refer to combination two kinds or
A kind of man-machine interactive information interchange of two or more media and communications media, the media used include word, picture, photo,
Interaction function that sound, animation and film and formula are provided etc..
With the arrival in big data epoch, the classification of mass multimedia data and digging technology are particularly important.In sea
It measures in data mining, how to utilize the information classified and excavated from data with existing to instruct the classification of new data and excavation
A new research hotspot is become.Particularly when the sample size of certain tasks is less, can have using multi-task learning
The reduction mass data classification of effect and the time cost excavated simultaneously improve acquisition of information accuracy.For example, in face of being known based on face
Other access control system of residential community development task, if the facial image of each owner is respectively divided into an independent image data class
Not, it is necessary to which the Classification and Identification to facial image is realized in system processing, judges that collected facial image is to belong at current gate inhibition
Which who in the face (judging image data classification belonged to) of owner, and then judge whether to release gate inhibition.
A kind of effective, robust information classification approach is proved to be based on deep learning method in practice.Depth nerve
Network (such as depth convolutional neural networks) is most representative machine learning method.Deep learning model usually has tens of layers
The data analysis layer that can learn, have it is hundreds thousand of, even millions of can be with learning parameter.Since quantity of parameters composition is extremely huge
Studying space, optimal model parameter in order to obtain, it usually needs a large amount of training data.But in order to train depth
Practise model, it is necessary to which structure possesses the training dataset of great amount of samples, and usual training samples number is more than tens thousand of.However, structure
Such training set is extremely difficult in practical applications, and costs dearly.For example, in face of based on the small of recognition of face
Area's access control system development task, if the facial image of each owner is respectively divided into an independent image data classification,
When carrying out Classification and Identification training to machine learning model, if necessary to acquire ten hundreds of facial images for each owner
Training sample is very unpractical.Deep learning method is caused many to the greediness of big data which results in depth model
Field is all difficult to obtain concrete application, is difficult with reliable technology realizability in other words.
When deep learning method is used for classification task, traditional deep learning method requirement disaggregated model contrast sample's
Class must be identical with producing the class of sample, i.e. model can only classify the class learnt, if there is the sample of new class needs to classify,
Must re -training machine learning model or machine learning model is done some adaptability training study.For example, in face of base
In the access control system of residential community development task of recognition of face, if the facial image of each owner is respectively divided into an independent figure
As data category, using current deep learning method, it is required for carrying out study instruction to the facial image of each current owner
Practice；When there is a new owner to occur, even if the facial image of the new owner is added directly into identification contrast sample's data
In library, due to not carrying out learning training to the facial image of the new owner before machine learning model, at gate inhibition again
It is secondary collect the new owner facial image when, machine learning model still can not be based on contrast sample's database in the new owner
Face image data and Direct Classification identifies the new owner.This also results in the machine learning mould based on deep learning method
The training of type needs to consume a large amount of training computing resource and longer training learning time, limits it in practical application field
Ease of use and versatility in conjunction.
Invention content
For above-mentioned deficiency in the prior art, present invention solves the technical problem that being that how to provide one kind is based on
The machine learning recognition methods of double contrast's study needs to solve existing multimedia data classification machine learning recognition methods
The problem of relying on a large amount of training sample and leading to practical application, further solves existing multimedia data classification machine
The problem of device study recognition methods directly can not carry out Classification and Identification to the classification without learning training and versatility is caused to be limited.
In order to solve the above technical problems, present invention employs following technological means：
Based on the machine learning recognition methods of double contrast's study, among the multi-medium data of multiple and different known class
Target identification sample and contrast sample are chosen, as a machine learning model f1Input, to machine learning model f1It is learned
Training is practised, and then utilizes the machine learning model f after learning training1Classification identification is carried out to multi-medium data to be identified；It is described
Machine learning model f1Including the first sub- learning model fDPWith the second sub- learning model fDE, the first sub- learning model fDPFor
Convolutional neural networks model or full Connection Neural Network model, the second sub- learning model fDEFor convolutional neural networks model or
Full Connection Neural Network model；Selected contrast sample includes more than two different classes of multiple multi-medium datas, and
Setting contrast sample is input to machine learning model f1Input put in order, and according to contrast sample input put in order, will
Target identification sample is combined with contrast sample with presetting rule of combination, is consequently formed and is remained with contrast sample input row
Each multi-medium data sample group cooperation is learnt mould by multiple data samples combination of row Cahn-Ingold-Prelog sequence rule for the described second son respectively
Type fDEInput, and by each corresponding second sub- learning model fDEOutput put in order rule according to contrast sample input
It then sorts and forms a data vector, as the described first sub- learning model fDPInput vector, and it is described first son study mould
Type fDPResult vector of the output vector as the machine learning model；Learn from there through training so that learning training institute
The machine learning model f obtained1Each result vector element in the result vector of output to characterize target identification sample with
The correlation between contrast sample's generic on corresponding arrangement ordinal position, so as to utilize the more of known class
Media data sample is put in order using different contrast sample's inputs to the machine learning model f1Carry out multiple study
Training.
In the above-mentioned machine learning recognition methods based on double contrast's study, preferably, as machine learning mould
Type f1The target identification sample of input is one or more, and belongs to same category；
If as machine learning model f1The target identification sample of input is one, by target identification sample and comparative sample
When this is combined with presetting rule of combination, the presetting rule of combination is one kind among following manner：
Rule of combination mode is 1.：The target identification sample is established into combinations of pairs between each contrast sample respectively
Relationship carries out combinations of pairs respectively；
Rule of combination mode is 2.：Each contrast sample is first subjected to category division, then by the target identification sample
Syntagmatic is established between the contrast sample of each classification respectively, is respectively combined；
If as machine learning model f1The target identification sample of input is multiple, by target identification sample and comparative sample
When this is combined with presetting rule of combination, the presetting rule of combination is one kind among following manner：
Rule of combination mode a：Each target identification sample is established into matched group between each contrast sample respectively
Conjunction relationship carries out combinations of pairs respectively；
Rule of combination mode b：Each contrast sample is first subjected to category division, then by each target identification sample
Syntagmatic is established between the contrast sample of each classification respectively, is respectively combined；
Rule of combination mode c：Target complete identification sample is established between each contrast sample respectively as a whole
Combinations of pairs relationship carries out combinations of pairs respectively；
Rule of combination mode d：Each contrast sample is first subjected to category division, then makees target complete identification sample
Syntagmatic is established between the contrast sample of each classification respectively to be whole, is respectively combined.
In the above-mentioned machine learning recognition methods based on double contrast's study, preferably, to machine learning mould
Type f1During carrying out learning training, the target identification sample and contrast sample are from preset multi-medium data sample database
It is chosen, chooses the multi-medium data conduct of a part of known class included in the multi-medium data sample database every time
Target identification sample and contrast sample are to machine learning model f1Learning training is carried out, and several times from the multi-medium data sample
Target identification sample and contrast sample are chosen in this library to machine learning model f1Learning training is carried out, to ensure target identification sample
The selection of this and contrast sample traverse each multi-medium data classification included in the multi-medium data sample database, and are directed to
Each multi-medium data classification in multi-medium data sample database has been performed both by contrast sample's selection operation of at least H times, H be with
Frequency threshold value is chosen in the training of setting.
It is preferably, described to utilize study instruction in the above-mentioned machine learning recognition methods based on double contrast's study
Machine learning model f after white silk1To multi-medium data to be identified carry out classification identification concrete mode be：
The multi-medium data as object to be identified is obtained as sample to be identified and from multiple and different known class
The contrast sample chosen among multi-medium data, as the machine learning model f after learning training1Input, selected pair
More than two different classes of multiple multi-medium datas are included, and contrast sample is set to be input to machine learning model than sample
f1Input put in order, and put in order according to contrast sample's input, by sample to be identified and contrast sample with presetting
Rule of combination is combined, and is consequently formed and is remained with multiple data samples combination that contrast sample's input puts in order regular, point
It is the described second sub- learning model f not by each multi-medium data sample group cooperationDEInput, and by it is each it is corresponding second son learn
Practise model fDEOutput form a data vector according to the contrast sample input rule compositor that puts in order, as described the
One sub- learning model fDPInput vector, and the first sub- learning model fDPOutput vector as the machine learning mould
The result vector of type；In classification identification process, the machine learning model f1Each result in the result vector of output to
Secondary element is related between sample to be identified and contrast sample's generic on corresponding arrangement ordinal position to characterize
Property, so as to determine the generic of sample to be identified according to the correlation.
In the above-mentioned machine learning recognition methods based on double contrast's study, preferably, acquired is to be identified
Sample is one or more, and belongs to same category；
If it is input to machine learning model f1Sample to be identified for one, by sample to be identified with contrast sample with pre-
When the rule of combination of setting is combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode is 1.：The sample to be identified is established combinations of pairs respectively to close between each contrast sample
System, carries out combinations of pairs respectively；
Rule of combination mode is 2.：Each contrast sample is first subjected to category division, then by the sample to be identified point
Syntagmatic is not established between the contrast sample of each classification, is respectively combined；
If it is input to machine learning model f1Sample to be identified to be multiple, by sample to be identified with contrast sample with pre-
When the rule of combination of setting is combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode a：Each sample to be identified is established into combinations of pairs between each contrast sample respectively
Relationship carries out combinations of pairs respectively；
Rule of combination mode b：Each contrast sample is first subjected to category division, then by each sample to be identified point
Syntagmatic is not established between the contrast sample of each classification, is respectively combined；
Rule of combination mode c：All samples to be identified between each contrast sample are established respectively as a whole and are matched
To syntagmatic, combinations of pairs is carried out respectively；
Rule of combination mode d：Each contrast sample is first subjected to category division, it then will all sample conducts to be identified
It is whole to establish syntagmatic between the contrast sample of each classification respectively, it is respectively combined.
In the above-mentioned machine learning recognition methods based on double contrast's study, preferably, acquired is to be identified
Sample is one or more, and belongs to same category；
If the sample to be identified obtained is multiple, it may be used and be input to the machine learning model f in batches1It carries out
Identifying processing is input to machine learning model f in batches1Concrete mode be one kind among following manner：
Input mode is 1. in batches：Whole contrast samples are formed a sample with each sample to be identified respectively to input
Set；Multiple sample input set are consequently formed, by several times as the machine learning model f1Input；
Input mode is 2. in batches：Each contrast sample is first subjected to category division；Then it is selected from each classification
A contrast sample is taken, then chooses a sample to be identified, forms a sample input set；Multiple sample inputs are consequently formed
Set, by several times as the machine learning model f1Input；
Input mode is 3. in batches：Each contrast sample is first subjected to category division；Then it is selected from each classification
A contrast sample is taken, together with all samples to be identified, forms a sample input set；Multiple sample inputs are consequently formed
Set, by several times as the machine learning model f1Input；
Input mode is 4. in batches：Whole contrast samples are formed a sample input with all samples to be identified to gather,
As the machine learning model f1Input.
In the above-mentioned machine learning recognition methods based on double contrast's study, preferably, according to machine learning mould
Type f1The concrete mode that the result vector repeatedly exported carries out classification identifying processing is one kind among following manner：
Repeatedly export classification identification method 1.：Each result vector element in the result vector of each output of Statistical Comparison,
The highest result vector element of degree of correlation that correlation is embodied is found out, by the arrangement corresponding to the result vector element
Contrast sample's generic on ordinal position is determined as sample generic to be identified；
Repeatedly export classification identification method 2.：By machine learning model f1The result vector of each output adds up, and obtains
To accumulation result vector, so as to the correlation that each result vector element is characterized in accumulation result vector described in Statistical Comparison, look for
Go out the highest result vector element of degree of correlation that correlation is embodied, the arrangement corresponding to the result vector element is suitable
Contrast sample's generic that tagmeme is put is determined as sample generic to be identified.
It is above-mentioned based on double contrast study machine learning recognition methods in, preferably, the contrast sample from
Chosen in preset multi-medium data sample database, choose every time to as machine learning model f1The comparative sample of input
This categorical measure L is less than the categorical measure S of the multi-medium data of known class included in the multi-medium data sample database,
L and S is the integer more than 1, and needs repeatedly to choose contrast sample from the multi-medium data sample database respectively as machine
Device learning model f1Input, carry out multiple classification identifying processing to the sample identified, with ensure the selection of contrast sample traverse
Each multi-medium data classification included in the multi-medium data sample database, and for every in multi-medium data sample database
A multi-medium data classification has been performed both by contrast sample's selection operation of at least K times, and K is that number threshold is chosen in the identification with setting
Value；Then, Statistical Comparison machine learning model f1Each result vector in the result vector that each secondary classification identifying processing is exported
Element finds out the highest result vector element of degree of correlation that correlation is embodied, will be corresponding to the result vector element
Put in order on position contrast sample's generic be determined as sample generic to be identified.
Compared with the prior art, the invention has the advantages that：
1st, the present invention is based on the machine learning recognition methods of double contrast's study, can utilize a certain amount of of known class
Multi-medium data sample is put in order using different contrast sample's inputs to machine learning model f1Carry out multiple differentiated
Learning training can carry out a large amount of learning training to reach expected using a small amount of training sample to machine learning model
Classification recognition effect so as to significantly reduce the dependence to magnanimity training sample, solves existing multimedia data classification machine
The problem of device study recognition methods needs to rely on a large amount of training sample and lead to practical application.
2nd, the present invention is based on the machine learning recognition methods of double contrast's study, even for some multi-medium data class
Not without learning training, as long as the multi-medium data sample of the multi-medium data classification is added to identification contrast sample's data
In library, when sample to be identified is the multi-medium data of the category, machine learning model f1The result vector of output still can
Embody the otherness between the sample to be identified and other different classes of contrast samples and the comparative sample with the same category
Correlation between this, so as to can still determine the generic of sample to be identified according to the correlation, so as to convenient
Extension to carrying out classification identification without the multi-medium data classification of learning training, can solve because can not be directly to the non-study of Confucian classics
The problem of practising the classification progress Classification and Identification of training and versatility caused to be limited.
3rd, the present invention is based on the machine learning recognition methods of double contrast's study, in the process for carrying out classification identifying processing
In, may be used the classification identifying processing mode that data are locally chosen allow choose every time to as machine learning model f1It is defeated
The contrast sample's categorical measure L entered is less than the multi-medium data of known class included in the multi-medium data sample database
Categorical measure S, then by repeatedly choosing contrast sample respectively as machine learning model f1Input to the sample identified into
The mode of the multiple classification identifying processing of row, to reduce machine learning model f1Each classification identification processing procedure performs data operation
The data volume of processing avoids machine learning model f1The problem for the treatment of effeciency is too low or can not perform effectively processing.
4th, the present invention is based on the machine learning recognition methods of double contrast's study, existing multi-medium data is solved well
Sorting machine learns recognition methods because of the dependence to a large amount of training samples and because can not be directly to the classification without learning training
The problem of carrying out Classification and Identification and causing practical application, versatility limited more can effectively be applied to more extensively
In specific multimedia data classification use occasion, there is wide technology application and promotion prospect.
Description of the drawings
Fig. 1 is the flow diagram of machine learning training process in machine learning recognition methods of the present invention.
Fig. 2 is the flow diagram of another machine learning training process in machine learning recognition methods of the present invention.
Fig. 3 is the flow diagram of multi-medium data classification identification process in machine learning recognition methods of the present invention.
Specific embodiment
It needs to rely on a large amount of training sample for existing multimedia data classification machine learning recognition methods and cause
The problem of practical application, needs the recognition principle from existing machine learning recognition methods to be analyzed, and discovery causes to ask
The reason of topic occurs.Existing sorting machine learns recognition methods, typically by sample to be identified and the comparative sample of known class
This progress individually compares, and calculates the similarity between sample to be identified and contrast sample or calculates sample to be identified and comparison
Difference distance value between sample, to judge whether sample to be identified belongs to same category with contrast sample, so as to fulfill treating
Identify the classification identification of sample.Such machine learning recognition methods, applies the application scenarios identified in multimedia data classification
In, easily limited by technology application：
On the one hand, because in multi-medium data, all there may be larger data differences for same category of data sample；Example
Such as, for the access control system of residential community development task based on recognition of face, if the facial image of each owner is respectively divided
The image data classification independent for one needs system processing to realize the Classification and Identification to facial image, but even same position
The facial image of owner, also easily because ambient light, shooting angle, owner itself dressing dressing situations such as difference and exist
Image difference, and the training sample under each ambient light, shooting angle, dressing dressing cond, only for identical
Sample Similarity to be identified or difference distance value under cond calculate and identification judges there is direct help, which results in
Need owner's facial image largely under the conds such as varying environment light, different shooting angles, different dressing dressing
As training sample and identification correction data, learning training is carried out to machine learning model, just can guarantee preferable recognition of face
Effect so as to increase the operation difficulty for training model learning in practical application, causes to limit to technology application.
On the other hand, in existing sorting machine study recognition methods, it is different from sample generic to be identified it is each its
His class training sample is difficult to symbolize otherness for the learning training influence of the specimen discerning result to be identified；For example, it faces
For access control system of residential community development task based on recognition of face, if the facial image of each owner is respectively divided into one solely
Vertical image data classification is identified or during learning training, contrast sample's database for the facial image of an owner
In other arbitrary owners facial image, for the owner to be identified facial image identification or training result be similarity not
Foot or difference distance are larger；Therefore a large amount of non-similar contrast sample is for the identification of sample to be identified or learning training result
Significant differentiated can not be brought to influence, this also results in and is only capable of for the identification of sample to be identified or learning training indirectly
Enough dependent on generic contrast sample, dependence of the machine learning recognition methods to magnanimity training sample is increased.
Correspondingly, based on above-mentioned both sides limitation reason, this also results in another as a result, being directed to multimedia number
According in Classification and Identification application scenarios, for not passing through the data category of learning training, existing machine learning recognition methods pair
The sample to be identified of the category carries out effective classification identification.
Above-mentioned analysis result is directed to, based on the technological thought to solve the above problems, the present invention proposes a kind of based on double
The machine learning recognition methods of study is compared again, which employs the learning training mode different from the prior art to machine
Device learning model is trained, as shown in Figure 1, choosing target identification sample among the multi-medium data of multiple and different known class
This R and contrast sample a, as a machine learning model f1Input, to machine learning model f1Learning training is carried out, and then
Utilize the machine learning model f after learning training1Classification identification is carried out to multi-medium data to be identified；The machine learning model
f1Including the first sub- learning model fDPWith the second sub- learning model fDE, the first sub- learning model fDPFor convolutional neural networks
Model or full Connection Neural Network model, the second sub- learning model fDEConnect for convolutional neural networks model or full nerve net
Network model；Selected contrast sample includes more than two different classes of multiple multi-medium datas, and sets contrast sample
It is input to machine learning model f1Input put in order, such as in Fig. 1, by contrast sample input put in order list it is multiple
Contrast sample is respectively labeled as a1、a2、…、an, n is denoted as machine learning model f1Contrast sample's quantity of input, and according to
Contrast sample's input puts in order, by target identification sample R and contrast sample a1、a2、…、anWith presetting rule of combination into
Row combination, such as in Fig. 1 simple example be by target identification sample R by contrast sample input put in order respectively with comparative sample
This1、a2、…、anIt is combined, is consequently formed and remains with multiple data sample groups that contrast sample inputs the rule that puts in order
It closes, is respectively the described second sub- learning model f by each multi-medium data sample group cooperationDEInput, and by each corresponding
Two sub- learning model fDEOutput DE1、DE2、…、DEnThe rule compositor that puts in order is inputted according to the contrast sample and forms one
Data vector, as the described first sub- learning model fDPInput vector A, and the first sub- learning model fDPOutput to
Measure the result vector C as the machine learning model；Learn from there through training so that the machine learning obtained by learning training
Model f1Each result vector element c in the result vector C of outputi∈ C (i ∈ 1,2 ..., n }) know to characterize target
Very originally and accordingly arrange a contrast sample a on ordinal positioniCorrelation between (i ∈ { 1,2 ..., n }) generic
Property, so as to be put in order using the multi-medium data sample of known class using different contrast sample's inputs to the machine
Device learning model f1Carry out multiple learning training.For example, as shown in Fig. 2, using n contrast sample identical in Fig. 1, but it is logical
It crosses setting change contrast sample's input to put in order so that a originally1The contrast sample of ordinal position is adjusted to a4Ordinal position,
So as to machine learning model f1Carry out different learning trainings.
The present invention is based on double contrast study machine learning recognition methods, compared with the prior art for, employ not
With the technical implementation way of thinking, target identification sample and contrast sample are used as by the multi-medium data for choosing known class,
It is input to machine learning model f together1Learning training is carried out, selected contrast sample is needed comprising more than two inhomogeneities
Other multiple multi-medium datas, the difference between different classes of multiple contrast samples on input puts in order can be embodied
The opposite sex.Simultaneously as machine learning model f1Learning training target be so that learning training obtained by machine learning model f1It is defeated
Each result vector element in the result vector gone out is arranged with corresponding on ordinal position to characterize target identification sample
Correlation between one contrast sample's generic, also, due to machine learning model f1It is designed as including the first son study
Model fDPWith the second sub- learning model fDEBuilt-up pattern framework, the first sub- learning model f thereinDPWith the second son study mould
Type fDECan residual error can be selected with selected as convolutional neural networks model or full Connection Neural Network model, convolutional neural networks
Neural network model (Residual Neural Network, be abbreviated as ResNet), intensive convolutional network model
(DenseConvolutional Network, be abbreviated as DenseNet) etc., full Connection Neural Network can then select this field
The more common neural network model with full articulamentum of technical staff；Using including the first sub- learning model fDPWith the second son
Learning model fDEBuilt-up pattern framework form machine learning model f1, during learning training, by target identification sample
After (classification identifying processing phase should be sample to be identified) is combined with contrast sample, formation remains with contrast sample and inputs row
Multiple data samples combination of row Cahn-Ingold-Prelog sequence rule is input to the second sub- learning model fDE, obtained output is according to the comparative sample
This input put in order rule compositor form a data vector, as the described first sub- learning model fDPInput vector, from
And by the first sub- learning model fDPOutput is as machine learning model f1Result vector, thereby guarantee that machine learning model f1It is defeated
Pair for remaining and being inputted with contrast sample between putting in order that put in order of each result vector element in the result vector gone out
It should be related to, and due to the first sub- learning model fDPWith the second sub- learning model fDEIt can be with selected as convolutional neural networks model
Or full Connection Neural Network model, therefore each result vector element is suitable by contrast sample's input arrangement in result vector
The influence of sequence, so that the machine learning model f of training gained1Each result vector element in the result vector of output
The correlation characterized, all putting in order with contrast sample's input, there are relevance influences.Since in this way, in contrast sample with
Target identification sample belongs to difference of the multi-medium data of the same category on contrast sample's input puts in order, can be to engineering
Practise model f1Learning training result generate different influence, therefore, belong to each of the same category with target identification sample
Multi-medium data is in sample as a comparison in use, can be by adjusting its sequence in contrast sample's input puts in order
Position, to machine learning model f1Carry out the learning training of multiple differentiated.Meanwhile in contrast sample with target identification sample
Difference of the multi-medium data to belong to a different category on contrast sample's input puts in order, also can be to machine learning model f1's
Learning training result generates different influences, therefore, each multi-medium data to belong to a different category with target identification sample
In sample as a comparison in use, can also by adjusting its contrast sample input put in order in ordinal position, repeatedly
It participates in machine learning model f1Differentiated learning training.A certain amount of multimedia of known class just can be utilized as a result,
Data sample is put in order using different contrast sample's inputs to machine learning model f1Carry out the study instruction of multiple differentiated
Practice, a large amount of learning training can be carried out to machine learning model using a small amount of training sample and known to reach expected classification
Other effect so as to significantly reduce the dependence to magnanimity training sample, solves existing multimedia data classification machine learning
The problem of recognition methods needs to rely on a large amount of training sample and lead to practical application.
In specific application, in the machine learning recognition methods using the present invention to machine learning model f1Carry out study instruction
In experienced process, target identification sample and contrast sample are chosen from preset multi-medium data sample database, are chosen every time
The multi-medium data of a part of known class included in the multi-medium data sample database is as target identification sample and right
Than sample to machine learning model f1Learning training is carried out, and chooses target from the multi-medium data sample database several times and knows
Very this and contrast sample are to machine learning model f1Learning training is carried out, to ensure the choosing of target identification sample and contrast sample
Each multi-medium data classification included in the traversal multi-medium data sample database is taken, and for multi-medium data sample database
In each multi-medium data classification be performed both by contrast sample's selection operation of at least H times, H is that the training with setting is chosen time
Number threshold value.The multi-medium data for choosing a part of known class included in multi-medium data sample database every time is known as target
Very this and contrast sample are to machine learning model f1Learning training is carried out, is a kind of learning training processing side locally chosen
Formula.It is performed because if the multi-medium data overall situation of classifications whole included in multi-medium data sample database is chosen to machine
Learning model f1Learning training processing, be easy to cause comparison operational data amount it is huge, operation efficiency is too low, also, if machine
Device learning model f1Neural network level it is excessive, be also easy to lead to machine learning model f1There can not be so a large amount of data to hold
The effective calculation process of row.Therefore, more matchmakers of a part of known class included in multi-medium data sample database are chosen every time
Volume data is as target identification sample and contrast sample to machine learning model f1Learning training is carried out, then by repeatedly choosing
Target identification sample and contrast sample are taken to machine learning model f1The mode of learning training is carried out, to reduce machine learning model
f1Each learning training processing procedure performs the data volume of data operation processing, avoids machine learning model f1Treatment effeciency is too low
Or the problem of processing can not be performed effectively；But this classification identifying processing mode locally chosen, it is possible to multimedia number occur
According to the multi-medium data training included in sample database be underutilized during learning training the problem of, in view of this,
And due to machine learning model f1Each result vector element in the result vector of output is inputted by contrast sample
The influence to put in order, thus the other contrast sample of same class on different contrast sample's inputs put in order for target
The interdependence effects of identification sample can have differences, and may influence to machine learning model f1Learning training as a result, being therefore
Ensure as far as possible to machine learning model f1Learning training effect, learning training process carry out ensure target identification sample and
The selection of contrast sample traverses each multi-medium data classification included in the multi-medium data sample database, and for more matchmakers
Each multi-medium data classification in volume data sample database has been performed both by contrast sample's selection operation of at least H times, and H is and setting
Training choose frequency threshold value, the specific value of H can determine according to practical application experience.
Utilize the machine learning model f obtained by the machine learning training program learning training1, just can be used in multimedia
Data carry out classification identification.Specifically, utilize the machine learning model f after learning training1To multi-medium data to be identified into
Row classification identification concrete mode be：As shown in figure 3, the multi-medium data as object to be identified is obtained as sample to be identified
RxAnd the contrast sample a chosen among the multi-medium data of multiple and different known class, as the machine after learning training
Learning model f1Input, selected contrast sample includes more than two different classes of multiple multi-medium datas, and sets
It puts contrast sample and is input to machine learning model f1Input put in order, such as in Fig. 3, by contrast sample input put in order
The multiple contrast samples listed are respectively labeled as a1、a2、…、an, n is denoted as machine learning model f1The contrast sample of input
Quantity, and put in order according to contrast sample's input, by sample R to be identifiedxWith contrast sample a1、a2、…、anWith presetting
Rule of combination is combined, such as simple example is by sample R to be identified in Fig. 3xIt puts in order point by contrast sample's input
Not with contrast sample a1、a2、…、anIt is combined, is consequently formed and remains with multiple numbers that contrast sample inputs the rule that puts in order
It is combined according to sample, is respectively the described second sub- learning model f by each multi-medium data sample group cooperationDEInput, and will be each
Corresponding second sub- learning model fDEOutput DE1、DE2、…、DEnThe rule compositor that puts in order is inputted according to the contrast sample
A data vector is formed, as the described first sub- learning model fDPInput vector A, and the first sub- learning model fDP
Result vector C of the output vector as the machine learning model；In classification identification process, the machine learning model f1It is defeated
Each result vector element c in the result vector C gone outi∈ C (i ∈ { 1,2 ..., n }) are characterizing sample R to be identifiedxWith
A contrast sample a on corresponding arrangement ordinal positioniCorrelation between (i ∈ { 1,2 ..., n }) generic, so as to
To determine sample R to be identified according to the correlationxGeneric；If for example, machine learning model f obtained by learning training1Its is defeated
Result vector element c in the result vector C gone outiValue it is smaller show with accordingly arrange ordinal position on contrast sample belonging to
Degree of correlation between classification is higher, then as shown in figure 3, determining sample R to be identified in identificationxGeneric yxWhen, result vector
Contrast sample a on the corresponding arrangement ordinal position of one result vector element of C intermediate values minimumiAffiliated classification yi, it is possible to
It is determined as being sample R to be identifiedxGeneric, i.e.,i∈{1,2,…,n}。
The above-mentioned machine in application, in the machine learning recognition methods that the present embodiment is learnt based on double contrast is embodied
Learning training process and multi-medium data classification identification processing procedure can be loaded into machine by being configured after computer programming
In the processor for learning identification device so that its processor is configured as performing the machine learning of above-mentioned machine learning training flow
Training program or the multi-medium data classification recognizer for performing above-mentioned multi-medium data classification identifying processing flow.Based on this
Machine learning classification identification device designed by the machine learning recognition methods of invention has common technical characterstic and skill naturally
Art advantage.
In machine learning recognition methods of the present invention and its specific implementation of device, to cause the machine obtained by learning training
Learning model f1Each result vector element in the result vector of output arranges to characterize target identification sample with corresponding
The correlation between contrast sample's generic on ordinal position when specific training operates, is easily accomplished；Example
Such as, in the training process, whether be directed to target identification sample with inputting a contrast sample on putting in order is same class
Not, in machine learning model f1The result vector element to put in order accordingly in the result vector of output on position assigns
Preset correlation desired value, such as the same category imparting positive correlation desired value (such as being assigned a value of " 0 "), and different classes of tax
Give negative correlation desired value (such as being assigned a value of " 1 "), then trained by machine learning, machine learning model f1It just being capable of acquistion
For whether the correlation of the same category is distinguished between target identification sample and contrast sample, in the result vector by its output
Each result vector element characterize belonging to target identification sample and a contrast sample on corresponding arrangement ordinal position
Correlation between classification.Thus after training, when performing the classification identifying processing to multi-medium data, machine learning mould
Type f1Each result vector element in the result vector of output just can be good at differentiated symbolize sample to be identified with
The correlation between contrast sample's generic on corresponding arrangement ordinal position, wherein closer to positive correlation desired value
Result vector element, corresponding to the element sorting position in result vector contrast sample input puts in order on position
Contrast sample's generic, it is possible to be determined to be the generic of sample to be identified.For example, shown in FIG. 1
It practises in training flow, target identification sample R and contrast sample input put in order the contrast sample a of the 1st1Belong to mutually similar
Not, therefore the result vector element c of the 1st of putting in order in result vector is assigned1Value for " 0 " represent positive correlation it is expected
Value, the corresponding result vector element of remaining different classes of contrast sample's ordinal position are assigned a value of " 1 " and represent negative correlation desired value；
And in learning training flow shown in Fig. 2, target identification sample R and contrast sample input put in order the comparative sample of the 4th
This4Belong to the same category, therefore assign the result vector element c of the 1st of putting in order in result vector4Value be " 0 " represent
Positive correlation desired value, the corresponding result vector element of remaining different classes of contrast sample's ordinal position are assigned a value of " 1 " and represent negative
Correlation desired value.
It should be noted that in the machine learning recognition methods of the present invention, the target identification sample that is obtained in learning training
Originally it can be one, may be multiple, and need to belong to same category；Equally, it is targeted in classification identifying processing
Sample to be identified may be one, may be multiple, but be equally also required to belong to same category.And it is directed to target knowledge
Very this (or sample to be identified) is one or more different situations, and machine learning recognition methods of the invention is in concrete application
In implementation, it is related to the factor of several respects, it is also desirable to which a point different situation is illustrated.
Wherein, the factor of first aspect, in learning training process or classification identification processing procedure, by target identification sample
(mutually should be sample to be identified in classification identification processing procedure) is input to machine learning model f with contrast sample1When, it needs with pre-
The rule of combination of setting is combined, combined treatment as progress, is on the one hand to enable to the multiple data to be formed
Sample combination remains with contrast sample and inputs the rule that puts in order, and is on the other hand to just establish target in data input level
Identification sample (or sample to be identified) inputs being associated with and contact between putting in order contrast sample between and with contrast sample,
This is also the important technology difference of machine learning recognition methods of the present invention compared with the prior art.In data input level just
Establish target identification sample (or sample to be identified) between contrast sample and with contrast sample input put in order between
Association contact after, pass through machine learning model f1Processing output, output result vector in each result to
Secondary element just no longer only has with target identification sample (or sample to be identified) and the similitude of a contrast sample therebetween
It closes, pass of the target identification sample (or sample to be identified) among also being combined with each data sample of composition between contrast sample
Connection property is related, and passes through machine learning model f1Full connection calculation process effect so that each result vector element also with
It is related that each data sample as input combines retained contrast sample's input rule that puts in order, so as to preferably ensure that
Machine learning model f1Each result vector element in the result vector of output to characterize target identification sample with it is corresponding
Correlation between the contrast sample's generic to put in order on position.
Exactly by means of by target identification sample (mutually should be sample to be identified in classification identification processing procedure) with comparison
Machine learning model f is input to after sample combination1Carry out the specially treated mode of learning training or classification identifying processing so that this
The machine learning recognition methods that invention is learnt based on double contrast can also have the data category for not passing through learning training
Standby classification recognition capability.Because the machine learning model f obtained after being trained using the method for the present invention1To multi-medium data to be identified
When carrying out classification identification, machine learning model f1A result vector element in the result vector of output not only with sample to be identified
This is related with the similitude of a contrast sample therebetween, the target among also more being combined with each data sample of composition
Identification sample (or sample to be identified) relevance between contrast sample and combine institute with each data sample as input
Reservation contrast sample input put in order rule it is related, therefore, even for some multi-medium data classification without
Training is practised, as long as the multi-medium data sample of the multi-medium data classification is added in identification contrast sample's database, when treating
When identifying the multi-medium data that sample is the category, machine learning model f1The result vector of output can still embody this and treat
Identify the otherness between the sample contrast sample different classes of with other and the phase between the contrast sample of the same category
Guan Xing, so as to can still determine the generic of sample to be identified according to the correlation.Therefore, the present invention is based on double contrasts
The machine learning recognition methods of study can be easily extended to carrying out classification without the multi-medium data classification of learning training
Identification can solve to cause what versatility was limited to ask because directly can not carry out Classification and Identification to the classification without learning training due to
Topic.
Meanwhile just because of need to establish target identification sample (or sample to be identified) between contrast sample and with it is right
Association contact between putting in order than sample input, thus the quantity of target identification sample (or sample to be identified) for one or
For under multiple different situations, specific combination also can different from.
If as machine learning model f1The target identification sample (or sample to be identified) of input is one, is known by target
When very this (or sample to be identified) is combined with contrast sample with presetting rule of combination, the presetting combination rule
It is then one kind among following manner：
Rule of combination mode is 1.：By the target identification sample (or sample to be identified) respectively with each contrast sample it
Between establish combinations of pairs relationship, carry out combinations of pairs respectively；
Rule of combination mode is 2.：Each contrast sample is first subjected to category division, then by the target identification sample
(or sample to be identified) establishes syntagmatic between the contrast sample of each classification respectively, is respectively combined.
If as machine learning model f1The target identification sample (or sample to be identified) of input is multiple, is known by target
When very this (or sample to be identified) is combined with contrast sample with presetting rule of combination, the presetting combination rule
It is then one kind among following manner：
Rule of combination mode a：By each target identification sample (or sample to be identified) respectively with each contrast sample
Between establish combinations of pairs relationship, carry out combinations of pairs respectively；
Rule of combination mode b：Each contrast sample is first subjected to category division, then by each target identification sample
(or sample to be identified) establishes syntagmatic between the contrast sample of each classification respectively, is respectively combined；
Rule of combination mode c：Target complete identification sample (or sample to be identified) is right with each respectively as a whole
Than establishing combinations of pairs relationship between sample, combinations of pairs is carried out respectively；
Rule of combination mode d：Each contrast sample is first subjected to category division, target complete is then identified into sample
(or sample to be identified) establishes syntagmatic between the contrast sample of each classification respectively as a whole, carries out group respectively
It closes.
Combinations of the above regular fashion is 1. and rule of combination mode a is by each target identification sample (or sample to be identified
This) combinations of pairs relationship is established between each contrast sample respectively, combinations of pairs is carried out respectively, and such rule of combination can
To form the data sample combination as much as possible for remaining with contrast sample's input and putting in order regular, for learning training processing
For flow, quantity data sample combination as much as possible is conducive to put in order by changing different contrast sample's inputs
And be used for execution differentiated different more times and think learning training, for hoisting machine learning model f1Training effect have centainly
It helps.And rule of combination mode 2. and processing rule b, c, d be then by target complete identification sample (or sample to be identified) work
, for an entirety (belonging to a classification) or using each contrast sample of each classification as an entirety then carry out respectively again
Combination, the multiple data samples combination being consequently formed can not only retain contrast sample and input the rule that puts in order, and will be complete
Each comparison of the portion's target identification sample (or sample to be identified) as an entirety (belonging to a classification) or by each classification
Sample closed as one data sample group of an overall structure in component part, a data sample group being consequently formed close into
Enter machine learning model f1During middle progress calculation process, calculation process process is equivalent to the data for incorporating respective classes sample
Common feature, therefore have certain help for common feature Division identification rate of the raising between different classes of.
The factor of second aspect in classification identification processing procedure, is getting multiple same category of samples to be identified
It needs to carry out classification identifying processing, while in the case that contrast sample is there is also multiple classifications and multiple quantity, may be used point
Batch is input to machine learning model f1Mode processing is identified；During concrete operations, it is input to machine learning mould in batches
Type f1Concrete mode one kind among following manner may be used：
Input mode is 1. in batches：Whole contrast samples are formed a sample with each sample to be identified respectively to input
Set；Multiple sample input set are consequently formed, by several times as the machine learning model f1Input；
Input mode is 2. in batches：Each contrast sample is first subjected to category division；Then it is selected from each classification
A contrast sample is taken, then chooses a sample to be identified, forms a sample input set；Multiple sample inputs are consequently formed
Set, by several times as the machine learning model f1Input；
Input mode is 3. in batches：Each contrast sample is first subjected to category division；Then it is selected from each classification
A contrast sample is taken, together with all samples to be identified, forms a sample input set；Multiple sample inputs are consequently formed
Set, by several times as the machine learning model f1Input；
Input mode is 4. in batches：Whole contrast samples are formed a sample input with all samples to be identified to gather,
As the machine learning model f1Input.
Correspondingly, using machine learning model f is input in batches1Carry out the processing mode of classification identification, each batch
Input will all obtain a result vector, therefore according to machine learning model f1The result vector repeatedly exported carries out classification knowledge
Other places reason concrete mode can also in the following way among one kind：
Repeatedly export classification identification method 1.：Each result vector element in the result vector of each output of Statistical Comparison,
The highest result vector element of degree of correlation that correlation is embodied is found out, by the arrangement corresponding to the result vector element
Contrast sample's generic on ordinal position is determined as sample generic to be identified；
Repeatedly export classification identification method 2.：By machine learning model f1The result vector of each output adds up, and obtains
To accumulation result vector, so as to the correlation that each result vector element is characterized in accumulation result vector described in Statistical Comparison, look for
Go out the highest result vector element of degree of correlation that correlation is embodied, the arrangement corresponding to the result vector element is suitable
Contrast sample's generic that tagmeme is put is determined as sample generic to be identified.
Wherein, repeatedly output classification identification method be 1. directly according to each time export result vector in whole results to
Secondary element carries out correlation statistics comparison, finds out highest one of degree of correlation to determine sample generic to be identified.It is and more
2. the secondary classification identification method that exports is then just to compare the result vector of each output with the correlation statistics that add up later again, find out
Highest one of degree of correlation determines sample generic to be identified.In comparison, classification identification method 2. phase is repeatedly exported
When then to being input to machine learning model f in batches1Each secondary result vector output after classification identifying processing is carried out to carry out
The comprehensive consideration of cumulative mean for classification identification method is repeatedly exported 1., is more conducive to avoid because of accidental error
The wrong situation of sample class identification to be identified, helps to ensure that better recognition accuracy caused by and.
The factor of the third aspect, during classification identifying processing is carried out using machine learning recognition methods of the present invention,
Contrast sample can be chosen, and in specific application from preset multi-medium data sample database, can be operated every time
Choose to as machine learning model f1Contrast sample's categorical measure L of input is less than in the multi-medium data sample database
Comprising the categorical measure S, L and S of multi-medium data of known class be integer more than 1, and need from more matchmakers
Contrast sample is repeatedly chosen in volume data sample database respectively as machine learning model f1Input, to the sample identified carry out it is more
Secondary classification identifying processing, to ensure that the selection of contrast sample traverses each more matchmakers included in the multi-medium data sample database
Volume data classification, and it has been performed both by at least comparison of K times for each multi-medium data classification in multi-medium data sample database
Sample selection operation, K are that frequency threshold value is chosen in the identification with setting.Allow choose every time to as machine learning model f1It is defeated
The contrast sample's categorical measure L entered is less than the multi-medium data of known class included in the multi-medium data sample database
Categorical measure S is a kind of classification identifying processing mode locally chosen.Because if included in multi-medium data sample database
The multi-medium data overall situations of whole classifications choose and perform classification identifying processing to the sample identified, be easy to cause comparison operation
Data volume is huge, and operation efficiency is too low, also, if machine learning model f1Neural network level it is excessive, be also easy to cause
Machine learning model f1There can not be so a large amount of data to perform effective calculation process.Therefore, allow choose every time making
For machine learning model f1Contrast sample's categorical measure L of input is less than known included in the multi-medium data sample database
The categorical measure S of the multi-medium data of classification, then by repeatedly choosing contrast sample respectively as machine learning model f1's
Input carries out the mode of multiple classification identifying processing to the sample identified, to reduce machine learning model f1At each classification identification
Reason process performs the data volume of data operation processing, avoids machine learning model f1Treatment effeciency is too low or can not perform effectively place
The problem of reason；But this classification identifying processing mode locally chosen, may cause to be not present in the contrast sample of single pick
Classification belonging to sample to be identified and cause to cannot get effective classification recognition result, and due to machine learning model f1Output
Result vector in each result vector element put in order and influenced, therefore same class by contrast sample's input
Other contrast sample influences to deposit on different contrast sample's inputs puts in order for the correlation identification of sample to be identified
In difference, classification recognition result to the sample identified may be influenced, therefore in order to ensure classification to the sample identified as far as possible
The accuracy of identification needs to ensure that the selection of contrast sample traverses each more matchmakers included in the multi-medium data sample database
Volume data classification, and it has been performed both by at least comparison of K times for each multi-medium data classification in multi-medium data sample database
Sample selection operation, K are the selection frequency threshold value with setting, and the specific value of K can be determined according to practical application experience.Then,
Statistical Comparison machine learning model f1Each result vector element in the result vector that each secondary classification identifying processing is exported, finds out
The highest result vector element of degree of correlation that correlation is embodied, by putting in order corresponding to the result vector element
Contrast sample's generic on position is determined as sample generic to be identified.
Recognition effect comparative example：
The machine that the present embodiment will use in the machine learning identification device provided by the present invention for multimedia data classification
Device learns recognition methods, and the recognition methods of prior art machine learning model is used with some, is carried out using identical data set
Recognition effect contrast experiment, verifying the machine learning identification side used in machine learning identification device provided by the present invention
The feasibility and validity of method.
In the present embodiment, the method for the present invention is labeled as " LCNN ", and prior art machine learning mould as a comparison
Type includes document " Lake, B.M., Salakhutdinov, R.＆Tenenbaum, J.B.Human-level concept
learning through probabilistic program induction Supplementary
Be mentioned in Material.Science 350,1332-1338 (2015) " BPL (BayesianProgram Learning,
Bayes plans learning algorithm) model (being labeled as " BPL [Lake2015] "), document " Vinyals, O., Blundell, C.,
Lillicrap,T.,Wierstra,D.&others.Matching networks for one shot learning.in
It is mentioned in Advances in Neural Information Processing Systems 3630-3638 (2016) "
Matching Nets (matching network) model (being labeled as " Matching Nets [Vinyals2016] ") and document
“Koch,G.,Richard Zemel&Ruslan Salakhutdinov.Siamese neural networks for one-
The Convolutional being mentioned in shot image recognition.in (University of Toronto, 2015) "
Siamese Net (convolution connection network) model (is labeled as " Convolutional Siamese Net [Kock2015] ").
The present embodiment is based on Omniglot data sets, choose 30 respectively in the training set provided in Omniglot data sets,
60th, for the sample of 136,156,964 classifications as training set, each classification has 20 samples, and the model for participating in comparison is distinguished
It is trained；Then, using document " Koch, G., Richard Zemel＆Ruslan Salakhutdinov.Siamese
neural networks for one-shot image recognition.in(University of Toronto,
2015) 400 test samples (each classification has 20 samples) of 20 classifications provided in " carry out the 20 single samples for selecting 1
The identification test of (20-way) classification, the classification identification test for each model carry out 100 times, count respective identification essence respectively
Degree.
In the present embodiment, using the first sub- learning model f in the method for the present invention schemeDPThe full connection god of selected as individual layer
Through network, the second sub- learning model fDEThe residual error neural network (ResNet) of 121 layers of selected as.The present embodiment is to as a comparison
The accuracy of identification statistical data that prior art machine learning model carries out classification identification test is as shown in table 1, to the method for the present invention
The accuracy of identification statistical data that each embodiment scheme carries out classification identification test is as shown in table 2.
Table 1
Table 2
It can see by above-mentioned Tables 1 and 2, machine learning recognition methods of the present invention is due to can be based on identical training
Sample set is to machine learning model f1Carry out more differentiated learning training, thus in identical training sample classification book and
Under conditions of training samples number, the accuracy of identification of machine learning recognition methods of the present invention leads over the existing skill for participating in comparison
The accuracy of identification of art machine learning model, it is sufficient to illustrate that the machine learning recognition methods of the present invention is directed to multi-medium data classification
Identification has good feasibility and validity.
In conclusion the present invention is based on the machine learning recognition methods of double contrast's study, known class can be utilized
A certain amount of multi-medium data sample is put in order using different contrast sample's inputs to machine learning model f1It carries out multiple
The learning training of differentiated can carry out a large amount of learning training to reach using a small amount of training sample to machine learning model
To expected classification recognition effect, so as to significantly reduce the dependence to magnanimity training sample, solves existing multimedia number
The problem of needing to rely on a large amount of training sample according to sorting machine study recognition methods and leading to practical application；Meanwhile that is,
Just for some multi-medium data classification without learning training, as long as the multi-medium data sample by the multi-medium data classification
Originally it is added in identification contrast sample's database, when sample to be identified is the multi-medium data of the category, machine learning model
f1The result vector of output can still embody the difference between the sample to be identified and other different classes of contrast samples
Property and the correlation between the contrast sample of the same category, so as to can still determine sample to be identified according to the correlation
Generic, so as to easily extend to without learning training multi-medium data classification carry out classification identification, energy
It is enough to solve the problems, such as versatility to be caused to be limited because directly carry out Classification and Identification to the classification without learning training due to；In addition,
During classification identifying processing is carried out, the classification identifying processing mode that data are locally chosen may be used and allow what is chosen every time
To as machine learning model f1Contrast sample's categorical measure L of input is less than included in the multi-medium data sample database
Known class multi-medium data categorical measure S, then by repeatedly choosing contrast sample respectively as machine learning mould
Type f1Input carry out the mode of multiple classification identifying processing to the sample identified, to reduce machine learning model f1Each classification
Identification processing procedure performs the data volume of data operation processing, avoids machine learning model f1Treatment effeciency is too low or can not be effective
The problem of performing processing.It can be seen that the present invention is based on the machine learning recognition methods of double contrast's study, solve well
Existing multimedia data classification machine learning recognition methods because of the dependence to a large amount of training samples and because can not directly to without
The problem of classification of learning training carries out Classification and Identification and causes practical application, versatility limited, can be more effective extensively
Be applied in more specific multimedia data classification use occasions, there is wide technology application and promotion prospect.
Finally illustrate, the above embodiments are merely illustrative of the technical solutions of the present invention and it is unrestricted, although with reference to reality
Example is applied the present invention is described in detail, it will be understood by those of ordinary skill in the art that, it can be to the technical side of the present invention
Case is modified or replaced equivalently, and without departing from the objective and range of technical solution of the present invention, should all be covered in the present invention
Right in.
Claims (8)
1. a kind of machine learning recognition methods based on double contrast's study, which is characterized in that in multiple and different known class
Target identification sample and contrast sample are chosen among multi-medium data, as a machine learning model f1Input, to machine
Learning model f1Learning training is carried out, and then utilizes the machine learning model f after learning training1To multi-medium data to be identified into
Row classification identifies；The machine learning model f1Including the first sub- learning model fDPWith the second sub- learning model fDE, described first
Sub- learning model fDPFor convolutional neural networks model or full Connection Neural Network model, the second sub- learning model fDEFor volume
Product neural network model or full Connection Neural Network model；Selected contrast sample includes more than two different classes of more
A multi-medium data, and contrast sample is set to be input to machine learning model f1Input put in order, and according to contrast sample
Input puts in order, and target identification sample with contrast sample with presetting rule of combination is combined, reservation is consequently formed
There is multiple data samples combination that contrast sample's input puts in order regular, be by each multi-medium data sample group cooperation respectively
The second sub- learning model fDEInput, and by each corresponding second sub- learning model fDEOutput according to the comparative sample
This input put in order rule compositor form a data vector, as the described first sub- learning model fDPInput vector, and
The first sub- learning model fDPResult vector of the output vector as the machine learning model；It is learned from there through training
It practises so that the machine learning model f obtained by learning training1Each result vector element in the result vector of output is to table
The correlation between target identification sample and contrast sample's generic on corresponding arrangement ordinal position is levied, so as to
It is put in order using the multi-medium data sample of known class using different contrast sample's inputs to the machine learning model
f1Carry out multiple learning training.
2. the machine learning recognition methods according to claim 1 based on double contrast's study, which is characterized in that as machine
Learning model f1The target identification sample of input is one or more, and belongs to same category；
If as machine learning model f1The target identification sample of input is one, by target identification sample and contrast sample with
When presetting rule of combination is combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode is 1.：The target identification sample is established combinations of pairs respectively to close between each contrast sample
System, carries out combinations of pairs respectively；
Rule of combination mode is 2.：Each contrast sample is first subjected to category division, then distinguishes the target identification sample
Syntagmatic is established between the contrast sample of each classification, is respectively combined；
If as machine learning model f1The target identification sample of input to be multiple, by target identification sample and contrast sample with
When presetting rule of combination is combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode a：Each target identification sample is established combinations of pairs respectively to close between each contrast sample
System, carries out combinations of pairs respectively；
Rule of combination mode b：Each contrast sample is first subjected to category division, then distinguishes each target identification sample
Syntagmatic is established between the contrast sample of each classification, is respectively combined；
Rule of combination mode c：Target complete identification sample is established into pairing between each contrast sample respectively as a whole
Syntagmatic carries out combinations of pairs respectively；
Rule of combination mode d：Each contrast sample is first subjected to category division, then identifies sample as whole target complete
Body establishes syntagmatic between the contrast sample of each classification respectively, is respectively combined.
3. the machine learning recognition methods according to claim 1 based on double contrast's study, which is characterized in that machine
Learning model f1During carrying out learning training, the target identification sample and contrast sample are from preset multi-medium data sample
It is chosen in this library, chooses the multimedia number of a part of known class included in the multi-medium data sample database every time
According to as target identification sample and contrast sample to machine learning model f1Learning training is carried out, and several times from the multimedia
Target identification sample and contrast sample are chosen in data sample library to machine learning model f1Learning training is carried out, to ensure target
Identify that the selection of sample and contrast sample traverse each multi-medium data classification included in the multi-medium data sample database,
And the contrast sample that at least H times has been performed both by for each multi-medium data classification in multi-medium data sample database chooses behaviour
Make, H is that frequency threshold value is chosen in the training with setting.
4. the machine learning recognition methods according to claim 1 based on double contrast's study, which is characterized in that the utilization
Machine learning model f after learning training1To multi-medium data to be identified carry out classification identification concrete mode be：
Obtain more matchmakers of the multi-medium data as sample to be identified and from multiple and different known class as object to be identified
The contrast sample chosen among volume data, as the machine learning model f after learning training1Input, selected comparative sample
This includes more than two different classes of multiple multi-medium datas, and contrast sample is set to be input to machine learning model f1's
Input puts in order, and is put in order according to contrast sample's input, by sample to be identified and contrast sample with presetting combination
Rule is combined, and is consequently formed and is remained with multiple data samples combination that contrast sample's input puts in order regular, respectively will
Each multi-medium data sample group cooperation is the described second sub- learning model fDEInput, and by it is each it is corresponding second son study mould
Type fDEOutput according to the contrast sample input put in order rule compositor form a data vector, as described first son
Learning model fDPInput vector, and the first sub- learning model fDPOutput vector as the machine learning model
Result vector；In classification identification process, the machine learning model f1Each result vector member in the result vector of output
Element to characterize the correlation between contrast sample's generic on sample to be identified and corresponding arrangement ordinal position, from
And the generic of sample to be identified is determined according to the correlation.
5. the machine learning recognition methods according to claim 4 based on double contrast's study, which is characterized in that acquired
Sample to be identified is one or more, and belongs to same category；
If it is input to machine learning model f1Sample to be identified for one, by sample to be identified with contrast sample with presetting
Rule of combination when being combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode is 1.：The sample to be identified is established into combinations of pairs relationship between each contrast sample respectively,
Combinations of pairs is carried out respectively；
Rule of combination mode is 2.：First by each contrast sample carry out category division, then by the sample to be identified respectively with
Syntagmatic is established between the contrast sample of each classification, is respectively combined；
If it is input to machine learning model f1Sample to be identified to be multiple, by sample to be identified with contrast sample with presetting
Rule of combination when being combined, the presetting rule of combination is one kind among following manner：
Rule of combination mode a：Each sample to be identified is established into combinations of pairs relationship between each contrast sample respectively,
Combinations of pairs is carried out respectively；
Rule of combination mode b：First by each contrast sample carry out category division, then by each sample to be identified respectively with
Syntagmatic is established between the contrast sample of each classification, is respectively combined；
Rule of combination mode c：All samples to be identified are established into matched group between each contrast sample respectively as a whole
Conjunction relationship carries out combinations of pairs respectively；
Rule of combination mode d：Each contrast sample is first subjected to category division, then by all samples to be identified as a whole
Syntagmatic is established between the contrast sample of each classification respectively, is respectively combined.
6. the machine learning recognition methods learnt according to right wants 4 based on double contrast, which is characterized in that acquired treats
Sample is identified as one or more, and belongs to same category；
If the sample to be identified obtained is multiple, it may be used and be input to the machine learning model f in batches1It is identified
Processing, is input to machine learning model f in batches1Concrete mode be one kind among following manner：
Input mode is 1. in batches：Whole contrast samples are formed into a sample input set with each sample to be identified respectively
It closes；Multiple sample input set are consequently formed, by several times as the machine learning model f1Input；
Input mode is 2. in batches：Each contrast sample is first subjected to category division；Then one is chosen from each classification
A contrast sample, then a sample to be identified is chosen, form a sample input set；Multiple sample input sets are consequently formed
It closes, by several times as the machine learning model f1Input；
Input mode is 3. in batches：Each contrast sample is first subjected to category division；Then one is chosen from each classification
A contrast sample together with all samples to be identified, forms a sample input set；Multiple sample input sets are consequently formed
It closes, by several times as the machine learning model f1Input；
Input mode is 4. in batches：Whole contrast samples are formed a sample input with all samples to be identified to gather, as
The machine learning model f1Input.
7. the machine learning recognition methods according to claim 6 based on double contrast's study, which is characterized in that according to machine
Learning model f1The concrete mode that the result vector repeatedly exported carries out classification identifying processing is one kind among following manner：
Repeatedly export classification identification method 1.：Each result vector element in the result vector of each output of Statistical Comparison, finds out
The highest result vector element of degree of correlation that correlation is embodied, by putting in order corresponding to the result vector element
Contrast sample's generic on position is determined as sample generic to be identified；
Repeatedly export classification identification method 2.：By machine learning model f1The result vector of each output adds up, and is added up
Result vector so as to the correlation that each result vector element is characterized in accumulation result vector described in Statistical Comparison, finds out correlation
The highest result vector element of degree of correlation that property is embodied, by the position that puts in order corresponding to the result vector element
On contrast sample's generic be determined as sample generic to be identified.
8. the machine learning recognition methods according to claim 4 based on double contrast's study, which is characterized in that the comparison
Sample is chosen from preset multi-medium data sample database, choose every time to as machine learning model f1Input
Contrast sample's categorical measure L is less than the classification of the multi-medium data of known class included in the multi-medium data sample database
Quantity S, L and S are the integer more than 1, and need repeatedly to choose contrast sample's difference from the multi-medium data sample database
As machine learning model f1Input, multiple classification identifying processing is carried out to the sample identified, to ensure the choosing of contrast sample
Each multi-medium data classification included in the traversal multi-medium data sample database is taken, and for multi-medium data sample database
In each multi-medium data classification be performed both by contrast sample's selection operation of at least K times, K is that the identification with setting is chosen time
Number threshold value；
Then, Statistical Comparison machine learning model f1Each result vector in the result vector that each secondary classification identifying processing is exported
Element finds out the highest result vector element of degree of correlation that correlation is embodied, will be corresponding to the result vector element
Put in order on position contrast sample's generic be determined as sample generic to be identified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810128018.8A CN108229692B (en) | 2018-02-08 | 2018-02-08 | Machine learning identification method based on dual contrast learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810128018.8A CN108229692B (en) | 2018-02-08 | 2018-02-08 | Machine learning identification method based on dual contrast learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108229692A true CN108229692A (en) | 2018-06-29 |
| CN108229692B CN108229692B (en) | 2020-04-07 |
Family
ID=62669913
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810128018.8A Active CN108229692B (en) | 2018-02-08 | 2018-02-08 | Machine learning identification method based on dual contrast learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108229692B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109255374A (en) * | 2018-08-27 | 2019-01-22 | 中共中央办公厅电子科技学院 | A kind of aesthetic properties evaluation method based on intensive convolutional network and multitask network |
| CN109344661A (en) * | 2018-09-06 | 2019-02-15 | 南京聚铭网络科技有限公司 | A kind of webpage integrity assurance of the micro code based on machine learning |
| CN110852376A (en) * | 2019-11-11 | 2020-02-28 | 杭州睿琪软件有限公司 | Method and system for identifying biological species |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
| CN104657745A (en) * | 2015-01-29 | 2015-05-27 | 中国科学院信息工程研究所 | Labelled sample maintaining method and two-way learning interactive classification method |
| US20160253466A1 (en) * | 2013-10-10 | 2016-09-01 | Board Of Regents, The University Of Texas System | Systems and methods for quantitative analysis of histopathology images using multiclassifier ensemble schemes |
| CN106022392A (en) * | 2016-06-02 | 2016-10-12 | 华南理工大学 | Deep neural network sample automatic accepting and rejecting training method |
| CN106845421A (en) * | 2017-01-22 | 2017-06-13 | 北京飞搜科技有限公司 | Face characteristic recognition methods and system based on multi-region feature and metric learning |
| CN106897746A (en) * | 2017-02-28 | 2017-06-27 | 北京京东尚科信息技术有限公司 | Data classification model training method and device |
| CN107193836A (en) * | 2016-03-15 | 2017-09-22 | 腾讯科技（深圳）有限公司 | A kind of recognition methods and device |
-
2018
- 2018-02-08 CN CN201810128018.8A patent/CN108229692B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101826166A (en) * | 2010-04-27 | 2010-09-08 | 青岛大学 | Novel recognition method of neural network patterns |
| US20160253466A1 (en) * | 2013-10-10 | 2016-09-01 | Board Of Regents, The University Of Texas System | Systems and methods for quantitative analysis of histopathology images using multiclassifier ensemble schemes |
| CN104657745A (en) * | 2015-01-29 | 2015-05-27 | 中国科学院信息工程研究所 | Labelled sample maintaining method and two-way learning interactive classification method |
| CN107193836A (en) * | 2016-03-15 | 2017-09-22 | 腾讯科技（深圳）有限公司 | A kind of recognition methods and device |
| CN106022392A (en) * | 2016-06-02 | 2016-10-12 | 华南理工大学 | Deep neural network sample automatic accepting and rejecting training method |
| CN106845421A (en) * | 2017-01-22 | 2017-06-13 | 北京飞搜科技有限公司 | Face characteristic recognition methods and system based on multi-region feature and metric learning |
| CN106897746A (en) * | 2017-02-28 | 2017-06-27 | 北京京东尚科信息技术有限公司 | Data classification model training method and device |
Non-Patent Citations (3)
| Title |
|---|
| KAIMING HE等: "Deep residual learning for image recognition", 《IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 张利军: "大规模机器学习理论研究与应用", 《中国博士学位论文全文数据库 信息科技辑》 * |
| 王少波 等: "神经网络学习样本点的选取方法比较", 《郑州大学学报（工学版）》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109255374A (en) * | 2018-08-27 | 2019-01-22 | 中共中央办公厅电子科技学院 | A kind of aesthetic properties evaluation method based on intensive convolutional network and multitask network |
| CN109344661A (en) * | 2018-09-06 | 2019-02-15 | 南京聚铭网络科技有限公司 | A kind of webpage integrity assurance of the micro code based on machine learning |
| CN110852376A (en) * | 2019-11-11 | 2020-02-28 | 杭州睿琪软件有限公司 | Method and system for identifying biological species |
| CN110852376B (en) * | 2019-11-11 | 2023-05-26 | 杭州睿琪软件有限公司 | Method and system for identifying biological species |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108229692B (en) | 2020-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108229588A (en) | A kind of machine learning recognition methods based on deep learning | |
| Han et al. | Competition-driven multimodal multiobjective optimization and its application to feature selection for credit card fraud detection | |
| Xue et al. | Particle swarm optimization for feature selection in classification: A multi-objective approach | |
| CN107563431A (en) | A kind of image abnormity detection method of combination CNN transfer learnings and SVDD | |
| CN108229692A (en) | A kind of machine learning recognition methods based on double contrast's study | |
| Zhang et al. | Improved biogeography-based optimization algorithm and its application to clustering optimization and medical image segmentation | |
| Brancati et al. | Gigapixel histopathological image analysis using attention-based neural networks | |
| CN109543741A (en) | A kind of FCM algorithm optimization method based on improvement artificial bee colony | |
| CN108537168A (en) | Human facial expression recognition method based on transfer learning technology | |
| CN107506350A (en) | A kind of method and apparatus of identification information | |
| Dhal et al. | An overview on nature-inspired optimization algorithms and their possible application in image processing domain | |
| CN108345942A (en) | A kind of machine learning recognition methods based on embedded coding study | |
| Xue et al. | Evolutionary sequential transfer optimization for objective-heterogeneous problems | |
| CN110287985A (en) | A kind of deep neural network image-recognizing method based on the primary topology with Mutation Particle Swarm Optimizer | |
| CN104463221A (en) | Imbalance sample weighting method suitable for training of support vector machine | |
| CN109816030A (en) | A kind of image classification method and device based on limited Boltzmann machine | |
| Lyu et al. | A novel multi-task optimization algorithm based on the brainstorming process | |
| CN108229693A (en) | A kind of machine learning identification device and method based on comparison study | |
| CN108345943A (en) | A kind of machine learning recognition methods based on embedded coding with comparison study | |
| CN109583492A (en) | A kind of method and terminal identifying antagonism image | |
| CN117459570A (en) | Client selection and self-adaptive model aggregation method and system based on reinforcement learning in federal learning | |
| Liu et al. | A improved NSGA-II algorithm based on sub-regional search | |
| Zhang et al. | Uncertainty measures and feature selection based on composite entropy for generalized multigranulation fuzzy neighborhood rough set | |
| Singh et al. | Transfer Learning Approach on Bacteria Classification from Microscopic Images | |
| Mishra et al. | Parallel multi-criterion genetic algorithms: review and comprehensive study |
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 | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220119 Address after: 408299 No. 7, floor 1, building B, No. 16, Second Branch Road, Pingdu East Road, Sanhe street, Fengdu County, Chongqing Patentee after: Chongqing Maoqiao Technology Co.,Ltd. Address before: No. 69 lijiatuo Chongqing District of Banan City Road 400054 red Patentee before: Chongqing University of Technology |