CN108596334A - The judgement of data correspondence, generation method and system based on two-way deep learning - Google Patents

Abstract

The invention discloses a kind of, and the data correspondence based on two-way deep learning judges, generation method and system, the data correspondence judgment method include：Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；Judge whether have default correspondence between two class testing input datas using the two-way deep learning neural network of foundation；The data creation method includes：Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；Using the two-way deep learning neural network of foundation, another kind of input data is generated according to a kind of input data；The data correspondence judges that system includes the first two-way deep learning neural network module and data correspondence judgment module；The data generation system includes the second two-way deep learning neural network module and data generation module.

Description

The judgement of data correspondence, generation method and system based on two-way deep learning

Technical field

The present invention relates to a kind of judgement of data correspondence, generation method and system based on two-way deep learning, belong to Data correspondence judges and generation technique field.

Background technology

The data resolution of acquisition is often influenced by environment and collecting device and demand, such as photo is not achieved in resolution ratio Due to factors such as weather haze, light, resolution of video camera, resolution ratio is not high, such as scans the picture come due to scanner The not high scanned picture resolution ratio caused of resolution ratio it is not high, be required for improving the resolution ratio of data in the case of these, it is existing The mode for being improved data resolution is interpolation or fitting, and interpolation or fitting are usually to be weighted according to adjacent given data value Averagely obtain needing the unknown data value be inserted into, but the distribution of real data is not fixed, but change, and interpolation or The mode of fitting can only use fixed interpolation or fitting formula, therefore be inserted in a fixed manner to the data of changes in distribution Value or fitting will obviously cause the distortion of data while improving resolution ratio.Another kind is differentiated on earth from high-resolution Rate, the prior art realize that the high-resolution data that may result in is discontinuous, another often through some data are deleted Kind mode is to count or be fitted by high-resolution data weighted average to obtain high-resolution data, but the distribution of real data is simultaneously It is not fixed, variation, and the mode for counting or being fitted can only use fixed statistics or fitting formula, therefore with solid Fixed mode is counted or is fitted to the data of changes in distribution will obviously cause data while reducing resolution ratio Distortion.The mankind tend to be envisioned as high-resolution data according to the data of low resolution, also can be according to high-resolution data It is envisioned as the data of low resolution.The brain of people is similar to neural network.Depth learning technology be from nerual network technique development and Come.

Existing depth learning technology can be obtained output label by input data and (such as obtain the people's identity card by head portrait Number, the people's identification card number is for another example obtained by voice), but had to through tape label number in the top-down supervised training stage According to supervised training (such as head portrait with identification card number label for another example carries the voice of identification card number label).

If being provided simultaneously with the head portrait with identification card number label and the voice with identification card number label, can pass through Deep learning neural network respectively obtains the corresponding identification card number of a head portrait and the corresponding identification card number of a voice, then sentences Whether disconnected one corresponding identification card number of head portrait and the corresponding identification card number of one voice are identical, if the same sentence One head portrait that breaks corresponding with one voice is same people, judges one head portrait and described one if different Corresponding a voice is not same people.But if it is not provided simultaneously with the head portrait with identification card number label and carries identity card The voice of number label, then can not judge a head portrait it is corresponding with a voice whether be same people, can not because without label Complete deep learning.If without the output label layer of deep learning neural network, since different types of input data passes through What deep learning neural network recognized out is different types of feature, and cannot achieve matching (example between different type feature As be between characteristics of image and sound characteristic cannot achieve it is matched).So if only head portrait and voice and part head portrait and Correspondence (corresponding same people) between voice, then by existing depth learning technology, can not judge a head portrait and one Whether corresponding a voice is same people.

If being provided simultaneously with the head portrait with identification card number label and the voice with identification card number label, can pass through Deep learning neural network respectively obtains the corresponding identification card number of head portrait voice corresponding with this identification card number, to To the corresponding voice of this image.But if it is not provided simultaneously with the head portrait with identification card number label and carries identification card number The voice of label can not complete deep learning then without label.If without the output label layer of deep learning neural network, Recognize out by deep learning neural network due to different types of input data is different types of feature, and inhomogeneity It cannot achieve conversion (such as being that cannot achieve conversion between characteristics of image and sound characteristic) between type feature.If only head Correspondence (corresponding same people) between picture and voice and part head portrait and voice, then by existing depth learning technology, Voice can not be obtained according to head portrait or image is obtained according to voice.

In addition, in existing depth learning technology, another kind of excessively complicated label can not be obtained by inputting a kind of data Data export, because only that fairly simple label data (such as class label) could be used for existing depth learning technology, because For excessively complicated label data output can make the computation complexity of deep learning increase severely and can not within the effective time it is complete At calculating, therefore excessively complicated label data can not be as the output of existing deep learning neural network, therefore, by existing Depth learning technology can not be achieved input high-resolution data and obtain output or the input high-resolution number of high-resolution data According to obtaining the output of high-resolution data.Meanwhile the cognitive process of deep learning neural network can only recognize out input data Feature, rather than the data of more high-low resolution or more low resolution than input data can not be recognized out from input data, institute With by the cognitive process of deep learning neural network also cannot achieve from high-resolution data obtain high-resolution data or from High-resolution data obtains high-resolution data.

Invention content

First of the present invention is designed to provide a kind of data correspondence judgment method based on two-way deep learning, This method can establish the default correspondence between two class input datas by two-way deep learning, and then may determine that two Class testing input data whether there is correspondence.

Second object of the present invention is to provide a kind of data creation method based on two-way deep learning, and this method can To establish the default correspondence between two class input datas by two-way deep learning, and then number can be inputted according to one kind According to the another kind of input data of generation.

Third object of the present invention is to provide a kind of data correspondence based on two-way deep learning and judges system.

Fourth object of the present invention is to provide a kind of data generation system based on two-way deep learning.

First purpose of the present invention can be reached by adopting the following technical scheme that：

Data correspondence judgment method based on two-way deep learning, the method includes：

Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；

Judge whether have default pair between two class testing input datas using the two-way deep learning neural network of foundation It should be related to.

Further, the default correspondence established two-way deep learning neural network, learn between two class input datas Relationship specifically includes：

First kind input data, the second class input data are obtained, and has the first kind input number of default correspondence According to the second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

Using the second class input data as the input of the second deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

The first kind input data of default correspondence be will be provided with the second class input data respectively as the first depth The input for practising the input and the second deep learning neural network of neural network, the output for obtaining the first deep learning neural network are made For the first output, the output of the second deep learning neural network is obtained as the second output, in first kind input data, the second class It is established between input data, the first output, the second output and presets correspondence；

First kind input data and the second output that will be provided with default correspondence are neural respectively as the first deep learning The input data and output label of network, carry out top-down supervised training；

The second class input data and the first output that will be provided with default correspondence are neural respectively as the second deep learning The input data and output label of network, carry out top-down supervised training；

When the amount of the first output and the amount and identical data of the variance data of the second output that have a default correspondence When ratio is less than the first predetermined threshold value, stop top-down supervised training.

Further, the two-way deep learning neural network that the utilization is established judges It is no to have default correspondence, it specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, it is deep that deep learning obtains first Spend the third output of learning neural network；

Using the second class testing input data as the input of the second deep learning neural network, it is deep that deep learning obtains second Spend the 4th output of learning neural network；

Compare the amount of third output and the amount and identical data of the variance data of the 4th output, if third output and the The ratio of the amount of the variance data of four outputs and the amount of identical data is less than the second predetermined threshold value, then the first class testing input data Have default correspondence between the second class testing input data, does not otherwise have default correspondence.

Second object of the present invention can be reached by adopting the following technical scheme that：

Based on the data creation method of two-way deep learning, the method includes：

Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；

Using the two-way deep learning neural network of foundation, another kind of input data is generated according to a kind of input data.

It is described to establish two-way deep learning neural network as a kind of embodiment, learn between two class input datas Default correspondence, specifically includes：

First kind input data, the second class input data are obtained, and has the first kind input number of default correspondence According to the second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

Using the second class input data as the input of the second deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

The first kind input data of default correspondence be will be provided with the second class input data respectively as the first depth The input for practising the input and the second deep learning neural network of neural network, the output for obtaining the first deep learning neural network are made For the first output, the output of the second deep learning neural network is obtained as the second output, in first kind input data, the second class It is established between input data, the first output, the second output and presets correspondence；

First kind input data and the second output that will be provided with default correspondence are neural respectively as the first deep learning The input data and output label of network, carry out top-down supervised training；

The second class input data and the first output that will be provided with default correspondence are neural respectively as the second deep learning The input data and output label of network, carry out top-down supervised training；

When the amount of the first output and the amount and identical data of the variance data of the second output that have a default correspondence When ratio is less than the first predetermined threshold value, stop top-down supervised training.

Further, described using the two-way deep learning neural network established, it is generated according to a kind of input data another Class input data, specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, it is deep that deep learning obtains first Spend the third output of learning neural network；

Third is exported into the output label as the second deep learning neural network, reversed generate obtains the second deep learning Second class input data of neural network；

Or

Using the second class testing input data as the input of the second deep learning neural network, it is deep that deep learning obtains second Spend the 4th output of learning neural network；

Output label by the 4th output as the first deep learning neural network, reversed generate obtain the first deep learning The first kind input data of neural network.

It is described to establish two-way deep learning neural network as another embodiment, learn between two class input datas Default correspondence, specifically include：

First kind input data, the second class input data are obtained, and has the first kind input number of default correspondence According to the second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

Using the second class input data as the input of the second deep learning neural network, unsupervised instruction from bottom to top is carried out Practice；

The first kind input data of default correspondence be will be provided with the second class input data respectively as the first depth It is deep to obtain first by the cognitive process of deep learning for the input for practising the input and the second deep learning neural network of neural network The top layer concept of learning neural network is spent as the first top layer concept, and the second depth is obtained by the cognitive process of deep learning The top layer concept of neural network is practised as the second top layer concept, in first kind input data, the second class input data, the first top layer It is established between concept, the second top layer concept and presets correspondence, and calculate the first top layer concept therein and the second top layer concept Variance data and identical data ratio；

The first kind input data of default correspondence be will be provided with the second top layer concept respectively as the first deep learning The input data of neural network and expected top layer concept, carry out top-down supervised training；

The second class input data of default correspondence be will be provided with the first top layer concept respectively as the second deep learning The input data of neural network and expected top layer concept, carry out top-down supervised training；

When the variance data and identical data of the first top layer concept for having default correspondence and the second top layer concept When ratio is less than default discrepancy threshold, stop top-down supervised training.

Further, described using the two-way deep learning neural network established, it is generated according to a kind of input data another Class input data, specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, pass through the cognition of deep learning Process obtains the third top layer concept of the first deep learning neural network；

Using third top layer concept as the top layer concept of the second deep learning neural network, generation obtains the second deep learning Second class input data of neural network, as the corresponding second class input data of the first class testing input data；

Or

Using the second class testing input data as the input of the second deep learning neural network, cognition obtains the second depth Practise the 4th top layer concept of neural network；

Using the 4th top layer concept as the top layer concept of the first deep learning neural network, generation obtains the first deep learning The first kind input data of neural network.

Third object of the present invention can be reached by adopting the following technical scheme that：

Data correspondence based on two-way deep learning judges system, the system comprises：

First two-way deep learning neural network module, for establishing two-way deep learning neural network, study two Default correspondence between class input data；

Data correspondence judgment module, for judging that two class testings are defeated using the two-way deep learning neural network established Enter between data and whether has default correspondence.

Further, the described first two-way deep learning neural network module, specifically includes：

First acquisition unit for obtaining first kind input data, the second class input data, and has default corresponding close The first kind input data of system and the second class input data；

First unsupervised training unit, for using first kind input data as the defeated of the first deep learning neural network Enter, carries out unsupervised training from bottom to top；

Second unsupervised training unit, for using the second class input data as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top；

First default correspondence relationship establishing unit, the first kind input data and second for will be provided with default correspondence Class input data is obtained respectively as the input of the input and the second deep learning neural network of the first deep learning neural network The output of first deep learning neural network obtains the output of the second deep learning neural network as second as the first output Output establishes between first kind input data, the second class input data, the first output, the second output and presets correspondence；

First supervised training unit, the first kind input data for will be provided with default correspondence are distinguished with the second output As the input data and output label of the first deep learning neural network, top-down supervised training is carried out；

Second supervised training unit, the second class input data for will be provided with default correspondence are distinguished with the first output As the input data and output label of the second deep learning neural network, top-down supervised training is carried out；

First stops supervised training unit, for the difference when the first output and the second output that have default correspondence When the ratio of the amount of data and the amount of identical data is less than the first predetermined threshold value, stop top-down supervised training.

Further, the data correspondence judgment module, specifically includes：

First deep learning unit, for using the first class testing input data as the defeated of the first deep learning neural network Enter, deep learning obtains the third output of the first deep learning neural network；

Second deep learning unit, for using the second class testing input data as the defeated of the second deep learning neural network Enter, deep learning obtains the 4th output of the second deep learning neural network；

Comparing unit, the amount for comparing third output and the amount and identical data of the variance data of the 4th output, if The ratio that the third exports the amount with the amount of the variance data of the 4th output and identical data is less than the second predetermined threshold value, then and first Have default correspondence between class testing input data and the second class testing input data, does not otherwise have default corresponding pass System.

Fourth object of the present invention can be reached by adopting the following technical scheme that：

Based on the data generation system of two-way deep learning, the system comprises：

Second two-way deep learning neural network module, for establishing two-way deep learning neural network, study two Default correspondence between class input data；

Data generation module, for using the two-way deep learning neural network established, being generated according to a kind of input data Another kind of input data.

As a kind of embodiment, the second two-way deep learning neural network module specifically includes：

Second acquisition unit for obtaining first kind input data, the second class input data, and has default corresponding close The first kind input data of system and the second class input data；

The unsupervised training unit of third, for using first kind input data as the defeated of the first deep learning neural network Enter, carries out unsupervised training from bottom to top；

4th unsupervised training unit, for using the second class input data as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top；

Second default correspondence relationship establishing unit, the first kind input data and second for will be provided with default correspondence Class input data is obtained respectively as the input of the input and the second deep learning neural network of the first deep learning neural network The output of first deep learning neural network obtains the output of the second deep learning neural network as second as the first output Output establishes between first kind input data, the second class input data, the first output, the second output and presets correspondence；

Third supervised training unit, the first kind input data for will be provided with default correspondence are distinguished with the second output As the input data and output label of the first deep learning neural network, top-down supervised training is carried out；

4th supervised training unit, the second class input data for will be provided with default correspondence are distinguished with the first output As the input data and output label of the second deep learning neural network, top-down supervised training is carried out；

Second stops supervised training unit, for the difference when the first output and the second output that have default correspondence When the ratio of the amount of data and the amount of identical data is less than the first predetermined threshold value, stop top-down supervised training.

Further, the data generation module, specifically includes：

Third deep learning unit, for using the first class testing input data as the defeated of the first deep learning neural network Enter, deep learning obtains the third output of the first deep learning neural network；

First reversed generation unit, for third to be exported to the output label as the second deep learning neural network, instead The second class input data of the second deep learning neural network is obtained to generation；

Or

4th deep learning unit, for using the second class testing input data as the defeated of the second deep learning neural network Enter, deep learning obtains the 4th output of the second deep learning neural network；

Second reversed generation unit, for the output label by the 4th output as the first deep learning neural network, instead The first kind input data of the first deep learning neural network is obtained to generation.

As another embodiment, the second two-way deep learning neural network module specifically includes：

Third acquiring unit for obtaining first kind input data, the second class input data, and has default corresponding close The first kind input data of system and the second class input data；

5th unsupervised training unit, for using first kind input data as the defeated of the first deep learning neural network Enter, carries out unsupervised training from bottom to top；

6th unsupervised training unit, for using the second class input data as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top；

Third presets correspondence relationship establishing unit, the first kind input data and second for will be provided with default correspondence Class input data passes through respectively as the input of the input and the second deep learning neural network of the first deep learning neural network The cognitive process of deep learning obtains the top layer concept of the first deep learning neural network as the first top layer concept, passes through depth The cognitive process of study obtains the top layer concept of the second deep learning neural network as the second top layer concept, is inputted in the first kind It is established between data, the second class input data, the first top layer concept, the second top layer concept and presets correspondence, and calculated wherein The first top layer concept and the second top layer concept variance data and identical data ratio；

5th supervised training unit, first kind input data and the second top layer concept for will be provided with default correspondence Input data respectively as the first deep learning neural network and expected top layer concept, carry out top-down supervised training；

6th supervised training unit, the second class input data and the first top layer concept for will be provided with default correspondence Input data respectively as the second deep learning neural network and expected top layer concept, carry out top-down supervised training；

Third stops supervised training unit, general for working as the first top layer concept for having default correspondence and the second top layer When the variance data of thought and the ratio of identical data are less than default discrepancy threshold, stop top-down supervised training.

Further, the data generation module, specifically includes：

First deep learning recognizes unit, for using the first class testing input data as the first deep learning neural network Input, the third top layer concept of the first deep learning neural network is obtained by the cognitive process of deep learning；

First generation unit is generated using third top layer concept as the top layer concept of the second deep learning neural network To the second class input data of the second deep learning neural network, as the corresponding second class input of the first class testing input data Data；

Or

Second deep learning recognizes unit, for using the second class testing input data as the second deep learning neural network Input, cognition obtain the 4th top layer concept of the second deep learning neural network；

Second generation unit, for using the 4th top layer concept as the top layer concept of the first deep learning neural network, life At obtaining the first kind input data of the first deep learning neural network.

The present invention has following advantageous effect compared with the existing technology：

1, data correspondence judgment method of the invention can establish two class input datas by two-way deep learning Between default correspondence, and then may determine that two class testing input datas whether there is correspondence, for example, by two-way Deep learning establishes the correspondence between head image data and voice data (corresponding same people), and progress can be in head image data A certain image whether correspond to same people with a certain audio in voice data.

2, data creation method of the invention can be established pre- between two class input datas by two-way deep learning If correspondence, and then another kind of input data can be generated according to a kind of input data, for example, by two-way deep learning come The correspondence (corresponding same people) between head image data and voice data is established, and then can be according to a certain in head image data Head portrait generates corresponding audio or generates corresponding head portrait according to a certain audio in voice data.

3, data creation method of the invention can be established pre- between two class input datas by two-way deep learning If correspondence, and then another kind of input data can be generated according to a kind of input data, for example, by two-way deep learning come It establishes low resolution or has incomplete image data with high-resolution or (corresponding same without the correspondence between incomplete image data Head portrait), and then by low resolution or there is incomplete head portrait to generate high-resolution or without incomplete image, or pass through high-resolution or nothing Incomplete head portrait generates low resolution or has incomplete image.

Description of the drawings

Fig. 1 is the flow chart of the data correspondence judgment method of the embodiment of the present invention 1.

Fig. 2 is the flow chart for establishing two-way deep learning neural network of the embodiment of the present invention 1.

Fig. 3 is the stream for judging whether to have default correspondence between two class testing input datas of the embodiment of the present invention 1 Cheng Tu.

Fig. 4 is the flow chart of the data creation method of the embodiment of the present invention 2.

Fig. 5 is the flow chart that the second class input data is generated according to the first class testing input data of the embodiment of the present invention 2.

Fig. 6 is the flow chart that first kind input data is generated according to the second class testing input data of the embodiment of the present invention 2.

Fig. 7 is the flow chart for establishing two-way deep learning neural network of the embodiment of the present invention 3.

Fig. 8 is the flow chart that the second class input data is generated according to the first class testing input data of the embodiment of the present invention 3.

Fig. 9 is the flow chart that first kind input data is generated according to the second class testing input data of the embodiment of the present invention 3.

Figure 10 is that the data correspondence of the embodiment of the present invention 4 judges the structure diagram of system.

Figure 11 is the structure diagram of the first two-way deep learning neural network module of the embodiment of the present invention 4.

Figure 12 is the structure diagram of the data correspondence judgment module of the embodiment of the present invention 4.

Figure 13 is the structure diagram of the data generation system of the embodiment of the present invention 5.

Figure 14 is the structure diagram of the second two-way deep learning neural network module of the embodiment of the present invention 5.

Figure 15 is that the data generation module one of which of the embodiment of the present invention 5 generates the structure diagram of scheme.

Figure 16 is that the data generation module another kind of the embodiment of the present invention 5 generates the structure diagram of scheme.

Figure 17 is the structure diagram of the two-way deep learning neural network module of the embodiment of the present invention 6.

Figure 18 is that the data generation module one of which of the embodiment of the present invention 6 generates the structure diagram of scheme.

Figure 19 is that the data generation module another kind of the embodiment of the present invention 6 generates the structure diagram of scheme.

Specific implementation mode

Present invention will now be described in further detail with reference to the embodiments and the accompanying drawings, but embodiments of the present invention are unlimited In this.

Embodiment 1：

The explanation of deep learning and to establish process as follows：

Generated from an input the involved calculating of output can by a flow graph (flow graph) come It indicates：Flow graph be it is a kind of can indicate calculate figure, in this figure each node indicate a basic calculating and The value of one calculating, the result of calculating are applied to the value of the child node of this node.Consider such a set of computations, it can To be allowed in each node and possible graph structure, and define a family of functions.Input node does not have father node, defeated Egress does not have child node.

The special attribute of one of this flow graph is depth (depth)：The longest path of an output is input to from one Length.

Regard learning structure as a network, then the core ideas of deep learning is as follows：

The first step：Using unsupervised training from bottom to top

1) successively structure monolayer neuronal is first.

2) tuning is carried out using wake-sleep algorithms every layer.One layer is only adjusted every time, is successively adjusted.

This process can be regarded as the process of a feature learning, be to distinguish maximum with traditional neural network Part.

Wake-sleep algorithms are as follows：

1) the wake stages：Cognitive process is weighed by the input feature vector (Input) of lower layer and upward cognition (Encoder) Each layer of abstract representation (Code) is generated again, then a reconstruction information is generated by current generation (Decoder) weight (Reconstruction), input feature vector and reconstruction information residual error are calculated, the downlink for being declined modification interlayer using gradient is generated (Decoder) weight.Namely " if reality imagines different with me, change the east that my generation weight makes me imagine West becomes as reality ".

2) the sleep stages：Generating process is generated by Upper Concept (Code) and downward generation (Decoder) weight The state of lower layer recycles cognition (Encoder) weight to generate an abstract scene.Using initial upper layer concept and create abstract The residual error of scene declines upward cognition (Encoder) weight of modification interlayer using gradient.Namely " if the scene in dream It is not the corresponding concepts in my brain, it is exactly this concept to change my cognition weight to make this scene in my view ".

Specifically, first train first layer with no nominal data, when training first learns the parameter of first layer, and (this layer can be seen Work is to obtain one so that exporting and inputting the hidden layer of the three-layer neural network of difference minimum), due to the limit of model capacity System and sparsity constraints so that obtained model can learn the structure to data itself, to obtain having more than input The feature of expression ability；After study obtains (n-1)th layer, the input by n-1 layers of output as n-th layer, training n-th layer, by This respectively obtains the parameter of each layer.

Second step：Top-down supervised training

This step be first step study obtain each layer parameter on the basis of, add a classification in the coding layer most pushed up Device (such as Rogers spy's recurrence, SVM etc.) goes to finely tune then by the supervised training of tape label data using gradient descent method Whole network parameter.

The first step of deep learning is substantially a network parameter initialization procedure.It is different from traditional neural network initial value Random initializtion, deep learning neural network are obtained by the structure of unsupervised trained input data, thus this initial value Closer to global optimum, so as to obtain better effect.

Judge whether there is correspondence to have application value realistic between two class input datas, for example, according to certain in video The head portrait and voice of people judge whether its head portrait is this human head picture, or judge whether its voice is to dub, to differentiate that video is No forgery is tampered.For another example, the cry of corresponding animal, or the cry according to different animals are generated according to the photo of different animals Generate the photo of corresponding animal.

As shown in Figure 1, a kind of data correspondence judgment method based on two-way deep learning is present embodiments provided, it should Method includes the following steps：

S101, two-way deep learning neural network is established, learns the default correspondence between two class input datas.

The step is as shown in Fig. 2, specifically include：

S1011, first kind input data, the second class input data are obtained, and has the first kind of default correspondence Input data and the second class input data.

First example of the present embodiment, first kind input data are head portrait, and the second class input data is voice, is had pre- If correspondence is " corresponding same people ", obtain multiple head portraits, multiple voices, the same people of multipair correspondence head portrait and voice.

Second example of the present embodiment, first kind input data are low-resolution image, and the second class input data is height Image in different resolution, it is " corresponding same image " to have default correspondence.Obtain multiple low-resolution images, multiple high-resolution Image, the same people of multipair correspondence low-resolution image and high-definition picture.

S1012, using first kind input data as the input of the first deep learning neural network, carry out nothing from bottom to top Supervised training.

First example of the present embodiment is carried out using each head portrait as the input of the first deep learning neural network Unsupervised training from bottom to top.

Second example of the present embodiment, using each low-resolution image as the defeated of the first deep learning neural network Enter, carries out unsupervised training from bottom to top.

S1013, using the second class input data as the input of the second deep learning neural network, carry out nothing from bottom to top Supervised training.

First example of the present embodiment is carried out using each voice as the input of the second deep learning neural network Unsupervised training from bottom to top.

Second example of the present embodiment, using each high-definition picture as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top.

S1014, the first kind input data that will be provided with default correspondence and the second class input data are respectively as first The input of the input and the second deep learning neural network of deep learning neural network, obtains the first deep learning neural network Output obtains the output of the second deep learning neural network as the second output as the first output, first kind input data, It is established between second class input data, the first output, the second output and presets correspondence.

First example of the present embodiment, using the head portrait of every a pair of corresponding same people and voice as the first depth The input for practising the input and the second deep learning neural network of neural network, the output for obtaining the first deep learning neural network are made For the first output, obtain the output of the second deep learning neural network as the second output, the head portrait, the voice, this first It is established between output, second output and presets correspondence " corresponding same people ".

Second example of the present embodiment, by the low-resolution image and high-definition picture point of every a pair of corresponding same people Input not as the input and the second deep learning neural network of the first deep learning neural network, obtains the first deep learning As the first output, the output for obtaining the second deep learning neural network is exported as second, low at this for the output of neural network Establishing default correspondence between image in different resolution, the high-definition picture, first output, second output, " correspondence is same Image ".

S1015, the first kind input data that will be provided with default correspondence are with the second output respectively as the first depth The input data and output label for practising neural network, carry out top-down supervised training.

First example of the present embodiment, the head portrait of every a pair of corresponding same people and the second output is deep as first The input data and output label for spending learning neural network, carry out top-down supervised training.

Second example of the present embodiment distinguishes the low-resolution image of every a pair of corresponding same image and the second output As the input data and output label of the first deep learning neural network, top-down supervised training is carried out.

S1016, the second class input data that will be provided with default correspondence are with the first output respectively as the second depth The input data and output label for practising neural network, carry out top-down supervised training.

First example of the present embodiment, the voice of every a pair of corresponding same people and the first output is deep as first The input data and output label for spending learning neural network, carry out top-down supervised training.

S1017, judgement have the first output and the amount of the variance data of the second output and the identical number of default correspondence According to amount ratio whether be less than the first predetermined threshold value, if so, stopping top-down supervised training, if it is not, repeating step Rapid S1014~S1016.

First example of the present embodiment judges the difference number of the first output and the second output per a pair of corresponding same people According to amount and the ratio of amount of identical data whether be less than the first predetermined threshold value, if so, stop top-down supervised training, if It is no, repeat step S1014~S1016.

Second example of the present embodiment judges the difference of the first output and the second output per a pair of corresponding same image Whether the ratio of the amount of data and the amount of identical data is less than the first predetermined threshold value, if so, stop top-down supervised training, If it is not, repeating step S1014~S1016.

S102, judge whether to have between two class testing input datas using the two-way deep learning neural network of foundation it is pre- If correspondence.

Step S102 is as shown in figure 3, specifically include：

S1021, using the first class testing input data as the input of the first deep learning neural network, deep learning obtains The third of first deep learning neural network exports.

First example of the present embodiment, using test head portrait P as the input of the first deep learning neural network, depth The third of acquistion to the first deep learning neural network exports.

Second example of the present embodiment, using for the low-resolution image P of test as the first deep learning nerve net The input of network, deep learning obtain the third output of the first deep learning neural network.

S1022, using the second class testing input data as the input of the second deep learning neural network, deep learning obtains 4th output of the second deep learning neural network.

First example of the present embodiment, using tested speech Q as the input of the second deep learning neural network, depth Fourth output of the acquistion to the second deep learning neural network.

Second example of the present embodiment, will be refreshing as the second deep learning for the test high-definition picture Q of test Input through network, deep learning obtain the 4th output of the second deep learning neural network.

S1023, the amount for comparing third output and the amount and identical data of the variance data of the 4th output, if the third is defeated The ratio for going out the amount with the amount of the variance data of the 4th output and identical data is less than the second predetermined threshold value, then the first class testing is defeated Enter between data and the second class testing input data and have default correspondence, does not otherwise have default correspondence.

First example of the present embodiment compares the similarities and differences of third output and the 4th output, if third output and the The ratio of the amount of the variance data of four outputs and the amount of identical data is less than the second predetermined threshold value, then image P is corresponding with voice Q same One people, otherwise corresponds to different people；It is identical as the first predetermined threshold value under second predetermined threshold value default situations.

Second example of the present embodiment compares the similarities and differences of third output and the 4th output, if third output and the The ratio of the amount of the variance data of four outputs and the amount of identical data is less than the second predetermined threshold value, then low-resolution image P and height Image in different resolution Q corresponds to same image, otherwise corresponds to different images；Under second predetermined threshold value default situations with the first predetermined threshold value It is identical.

Embodiment 2：

The data creation method based on two-way deep learning, important in inhibiting are present embodiments provided, such as have Suspect (such as the suspect swindled by phone) may be recorded, but nobody met its head portrait, if can be according to it Its head portrait of speech production, to seize the suspect play the role of it is particularly important.For another example, it is generated and is corresponded to according to the photo of different animals The cry of animal, or the photo for corresponding to animal is generated according to the cry of different animals.

As shown in figure 4, the data creation method based on two-way deep learning of the present embodiment includes the following steps：

S401, two-way deep learning neural network is established, learns the default correspondence between two class input datas.

The detailed process of the step is with embodiment 1, and details are not described herein.

S402, the two-way deep learning neural network using foundation generate another kind of input number according to a kind of input data According to.

Step S402 is as shown in figure 5, specifically include：

S4021, using the first class testing input data as the input of the first deep learning neural network, deep learning obtains The third of first deep learning neural network exports.

First example of the present embodiment, using test head portrait as the input of the first deep learning neural network, depth The third of acquistion to the first deep learning neural network exports.

Second example of the present embodiment, using for the low-resolution image of test as the first deep learning neural network Input, deep learning obtain the first deep learning neural network third output.

S4022, third is exported to the output label as the second deep learning neural network, reversed generate obtains second deeply Spend the second class input data of learning neural network.

Reversed generate obtains the second class input data of the second deep learning neural network, can be real in the following manner It is existing：

L1, all and the second deep learning neural network output label is matched from the second deep learning neural network Matched output label.

L2, the corresponding second class input data of output label for obtaining successful match, as the second deep learning nerve net Second class input data of network.

Reversed generate obtains the second class input data of the second deep learning neural network, can also be real in the following manner It is existing：

L1, using all top layer concepts in top layer concept big data or Portions of top layer concept is selected as the current of genetic algorithm Population.

L2, using each top layer concept in current population as the input of grader in the second deep learning neural network, lead to It crosses grader and the corresponding output label of this top layer concept, the wherein encoding and decoding of deep learning neural network most pushed up is calculated Layer is a grader (such as Rogers spy's recurrence, SVM etc.), and top layer concept is the output of the last one hidden layer, and classification The input of device.

L3, by the similarity between the corresponding output label of each top layer concept and the output label of the second class input data As the fitness of each top layer concept in current population, wherein the corresponding output label of each top layer concept is logical in previous step What the coding layer (grader) most pushed up crossed in depth learning neural network obtained.

L4, bred based on current population (including filial generation mutation) obtain it is new work as pre-group kind, return to step 1) It repeats, until (preset condition is for example, the number repeated has been more than preset times or kind for the preset condition that meets stopping The number of individuals that fitness is more than preset value in group has been more than default number of individuals) because the individual in current population is top layer concept, So being bred (including filial generation mutation) using based on current population obtains the new individual worked as in pre-group kind also as top layer Concept；The process of breeding is divided into three steps：Selection intersects, variation.Current population obtains after selection, intersection, mutation operator New works as pre-group kind.For example, top layer concept be animal picture feature image, then from current population according to fitness from greatly to The feature image of small selection predetermined number, the feature image chosen is weighted two-by-two and averagely obtains new multiple features A few pixels value in certain feature images in new multiple feature images is carried out complementary operation, obtained new institute by picture There is individual that new current population is added.

L5, the maximum top layer concept of fitness in the current population finally obtained or fitness are more than each of preset value Top layer concept of the top layer concept (may have multiple) as the second deep learning neural network, generation obtain the second class input number According to the second class input data as the second deep learning neural network.

Third is exported the output label as the second deep learning neural network, instead by first example of the present embodiment The input voice of the second deep learning neural network is obtained to generation, as the corresponding voice of test head portrait.

Third is exported the output label as the second deep learning neural network, instead by second example of the present embodiment The input data of the second deep learning neural network is obtained to generation, as the corresponding high score of low-resolution image for test Resolution image.

Or

Step S402 is as shown in fig. 6, specifically include following steps：

S4023, using the second class testing input data as the input of the second deep learning neural network, deep learning obtains 4th output of the second deep learning neural network.

First example of the present embodiment, using tested speech as the input of the second deep learning neural network, depth Fourth output of the acquistion to the second deep learning neural network.

Second example of the present embodiment, will be neural as the second deep learning for the test high-definition picture of test The input of network, deep learning obtain the 4th output of the second deep learning neural network.

S4024, the output label by the 4th output as the first deep learning neural network, reversed generate obtain first deeply Spend the first kind input data of learning neural network.

First example of the present embodiment, the output mark by the 4th output as the first deep learning neural network Label, reversed generate obtains the input head portrait of the first deep learning neural network, as the corresponding head portrait of tested speech.

Second example of the present embodiment, the output mark by the 4th output as the first deep learning neural network Label, reversed generate obtain the input picture of the first deep learning neural network, as the high-definition picture correspondence for test Low-resolution image.

Embodiment 3：

By two-way generation depth learning technology, first resolution input data can be established by two-way deep learning Default correspondence between second resolution input data, and then can be according to first in first resolution input data Input data generates the second input data in second resolution input data or according to the in second resolution input data Two input datas generate the first input data in first resolution input data.But it is not often corresponded in reality High-resolution sample data and low resolution sample data or only high-resolution data sample or only low resolution Rate data sample, therefore (mutual corresponding high-resolution can not be must have by existing two-way generation depth learning technology Sample data and low resolution sample data) realize that input high-resolution data obtains output or the input height of high-resolution data Resolution data obtains the output of high-resolution data.Meanwhile in two-way generation depth learning technology, first according to a kind of defeated Enter the input data in data and obtain output label, the input number in another kind of input data is then obtained by output label According to.Different output labels are generally always fewer than different input datas, so from the cognitive process for inputting data into output label It is convergence (multipair few), and the generating process from output label to input data is diverging (few to more), therefore can cause Cognitive process from high-resolution data to output label is convergence (multipair few), and from output label to high-resolution life It is diverging (few to more) at process, can causes, input the low-resolution image of a pig, obtained output label is pig mark Label, then the high-definition picture of a pig is obtained by pig label, but the pig in the two images is not same pig, because not Although same pig is all pig label, appearance has the difference in details.Therefore, pass through existing two-way generation deep learning nerve Network although high-resolution data can be obtained by high-resolution data, or obtains low resolution number by high-resolution data According to, and the corresponding output label of the two is consistent, but can not ensure that the data details of the two are also consistent, this obviously cannot be satisfied height Demand when resolution data is converted to data details consistency.Similarly, the prior art also cannot achieve no incomplete data and have Conversion between incomplete data.

The present embodiment based on the data creation method general steps of two-way deep learning with above-described embodiment 2, difference It is in the detailed process of step S401 and S402, important in inhibiting is applied, such as photo, remote sensing images, voice, point cloud The data resolutions such as data increase or decrease, and check details or overall picture convenient for people and be conducive to computer more refining or more saving Handle to computing resource.It for another example repairs incomplete or has a data such as the photo blocked, remote sensing images, voice, point cloud data.

The step S401 of the present embodiment is as shown in fig. 7, specifically include：

S4011, first kind input data, the second class input data are obtained, and has the first kind of default correspondence Input data and the second class input data.

First example of the present embodiment, first kind input data are head portrait；Second class input data is voice；Have pre- If correspondence is " corresponding same people ".Obtain multiple head portraits, multiple voices, the same people of multipair correspondence head portrait and voice.

Second example of the present embodiment, first kind input data are low-resolution image；Second class input data is height Image in different resolution；It is " corresponding same image " to have default correspondence.Obtain multiple low-resolution images, multiple high-resolution Image, the same people of multipair correspondence low-resolution image and high-definition picture.

The third example of the present embodiment, first kind input data are to have incomplete data；Second class input data is without residual Lack data；It is " corresponding same data " to have default correspondence, and obtaining multiple has incomplete data, multiple no incomplete datas, more There are incomplete data and no incomplete data to the same data of correspondence.

S4012, using first kind input data as the input of the first deep learning neural network, carry out nothing from bottom to top Supervised training.

First example of the present embodiment is carried out using each head portrait as the input of the first deep learning neural network Unsupervised training from bottom to top.

Second example of the present embodiment, using each low-resolution image as the defeated of the first deep learning neural network Enter, carries out unsupervised training from bottom to top.

Each is had incomplete data as the defeated of the first deep learning neural network by the third example of the present embodiment Enter, carries out unsupervised training from bottom to top.

S4013, using the second class input data as the input of the second deep learning neural network, carry out nothing from bottom to top Supervised training.

First example of the present embodiment is carried out using each voice as the input of the second deep learning neural network Unsupervised training from bottom to top.

Second example of the present embodiment, using each high-definition picture as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top.

The third example of the present embodiment, using each without incomplete data as the defeated of the second deep learning neural network Enter, carries out unsupervised training from bottom to top.

S4014, the first kind input data that will be provided with default correspondence and the second class input data are respectively as first The input of the input and the second deep learning neural network of deep learning neural network, is obtained by the cognitive process of deep learning The top layer concept of first deep learning neural network obtains second as the first top layer concept, by the cognitive process of deep learning The top layer concept of deep learning neural network is as the second top layer concept, in first kind input data, the second class input data, It is established between one top layer concept, the second top layer concept and presets correspondence, and calculate the first top layer concept therein and the second top The layer variance data of concept and the ratio of identical data.

First example of the present embodiment, using the head portrait of every a pair of corresponding same people and voice as the first depth It is deep to obtain first by the cognitive process of deep learning for the input for practising the input and the second deep learning neural network of neural network The top layer concept of learning neural network is spent as the first top layer concept, and the second depth is obtained by the cognitive process of deep learning The top layer concept of neural network is practised as the second top layer concept, on the head portrait, the voice, the first top layer concept, second top It is established between layer concept and presets correspondence " corresponding same people ", and calculate the first top layer concept therein and the second top layer concept Variance data and identical data ratio.

Second example of the present embodiment, by the low-resolution image and high-definition picture point of every a pair of corresponding same people Input not as the input and the second deep learning neural network of the first deep learning neural network, passes through recognizing for deep learning Know that process obtains the top layer concept of the first deep learning neural network as the first top layer concept, passes through the cognition of deep learning Journey obtains the top layer concept of the second deep learning neural network as the second top layer concept, in the low-resolution image, the high score It is established between resolution image, the first top layer concept, the second top layer concept and presets correspondence " corresponding same image ", and counted Calculate the ratio of the variance data and identical data of the first top layer concept therein and the second top layer concept.

Every a pair of corresponding same people is had incomplete data to make respectively with no incomplete data by the third example of the present embodiment For the input of the input and the second deep learning neural network of the first deep learning neural network, pass through the cognition of deep learning Journey obtains the top layer concept of the first deep learning neural network as the first top layer concept, is obtained by the cognitive process of deep learning To the second deep learning neural network top layer concept as the second top layer concept, this have incomplete data, this without incomplete data, It is established between the first top layer concept, the second top layer concept and presets correspondence " corresponding same data ", and calculated therein The ratio of the variance data and identical data of first top layer concept and the second top layer concept.

The cognitive process of deep learning specifically includes：

The concept on upper layer is generated by the concept and upward cognition (Encoder) weight of lower layer, Upper Concept compares lower layer Concept is more abstract, and undermost concept is input data, and the concept of top layer is top layer concept.It is to recognize from input data first Know to obtain the concept of hidden layer first layer, is then recognized to obtain the concept of the hidden layer second layer by the concept of hidden layer first layer, it is such as such It pushes away, the concept i.e. top layer concept until obtaining hiding last layer.

S4015, the first kind input data that will be provided with default correspondence and the second top layer concept are deep respectively as first The input data of learning neural network and expected top layer concept are spent, top-down supervised training is carried out.

First example of the present embodiment, using the head portrait of every a pair of corresponding same people and the second top layer concept as the The input data of one deep learning neural network and expected top layer concept, carry out top-down supervised training；

Second example of the present embodiment, by the low-resolution image and the second top layer concept of every a pair of corresponding same image Input data respectively as the first deep learning neural network and expected top layer concept, carry out top-down supervised training；

Every a pair of corresponding same data are had incomplete data and the second top layer concept point by the third example of the present embodiment Not as the input data of the first deep learning neural network and expected top layer concept, top-down supervised training is carried out；

By input data and top layer concept exercise supervision trained detailed process with by input data and output label The trained process that exercises supervision is similar, unsupervised training obtain each layer parameter on the basis of, gone using gradient descent method micro- An adjustment network parameter, the difference is that input data and top layer concept exercise supervision, trained detailed process need not export Encoding-decoding process between label and top layer concept.

Include by an input data and its corresponding expected top layer the concept trained detailed process that exercises supervision：

L1, practical top layer concept is obtained by input data input deep learning neural computing.

L2, the residual error being expected between top layer concept and practical top layer concept is calculated, is gone using gradient descent method according to residual error Entire network parameter is finely tuned, the target of adjustment is to enable practical top layer concept is closer to be expected top layer concept.

L3, repetition L1, L2 preset threshold residual value until the residual error between expected top layer concept and practical top layer concept is less than, Then terminated by this input data and its corresponding expected top layer the concept trained process that exercises supervision.

S4016, the second class input data that will be provided with default correspondence and the first top layer concept are deep respectively as second The input data of learning neural network and expected top layer concept are spent, top-down supervised training is carried out.

First example of the present embodiment, using the voice of every a pair of corresponding same people and the first top layer concept as the The input data of one deep learning neural network and expected top layer concept, carry out top-down supervised training；

Second example of the present embodiment, by the high-definition picture and the first top layer concept of every a pair of corresponding same image Input data respectively as the first deep learning neural network and expected top layer concept, carry out top-down supervised training；

The third example of the present embodiment, by every a pair of corresponding same data without incomplete data and the first top layer concept point Not as the input data of the first deep learning neural network and expected top layer concept, top-down supervised training is carried out；

Detailed process is the same as above-mentioned steps S4015.

S4017, judgement have the variance data and phase of the first top layer concept and the second top layer concept of default correspondence Whether the ratio with data is less than default discrepancy threshold, if so, stopping top-down supervised training, if it is not, repeating step Rapid S4014~S4016.

The step S402 of the present embodiment is as shown in figure 8, specifically include：

S4021, using the first class testing input data as the input of the first deep learning neural network, pass through deep learning Cognitive process obtain the third top layer concept of the first deep learning neural network.

First example of the present embodiment passes through depth using test head portrait as the input of the first deep learning neural network The cognitive process of degree study obtains the third top layer concept of the first deep learning neural network.

Second example of the present embodiment, using for the low-resolution image of test as the first deep learning neural network Input, the third top layer concept of the first deep learning neural network is obtained by the cognitive process of deep learning.

The third example of the present embodiment has incomplete data as the first deep learning neural network using for test Input, the third top layer concept of the first deep learning neural network is obtained by the cognitive process of deep learning.

The cognitive process of deep learning specifically includes：

The concept on upper layer is generated by the concept and upward cognition (Encoder) weight of lower layer, Upper Concept compares lower layer Concept is more abstract, and undermost concept is input data, and the concept of top layer is top layer concept；It is to recognize from input data first Know to obtain the concept of hidden layer first layer, is then recognized to obtain the concept of the hidden layer second layer by the concept of hidden layer first layer, it is such as such It pushes away, the concept i.e. top layer concept until obtaining hiding last layer.

S4022, using third top layer concept as the top layer concept of the second deep learning neural network, it is deep that generation obtains second The the second class input data for spending learning neural network, as the corresponding second class input data of the first class testing input data.

First example of the present embodiment, by the top layer that third top layer concept is used as to the deep learning neural network Concept generates input data by the generating process of deep learning, as the corresponding voice of test head portrait.

Second example of the present embodiment, by the top layer that third top layer concept is used as to the deep learning neural network Concept generates input data, the corresponding high-resolution of low-resolution image as test by the generating process of deep learning Image.

The third example of the present embodiment, by the top layer that third top layer concept is used as to the deep learning neural network Concept generates input data by the generating process of deep learning, has the corresponding no incomplete data of incomplete data as test.

The generating process of deep learning specifically includes：

The concept of lower layer is generated by the concept and downward generation (Decoder) weight on upper layer, lower layer's concept compares upper layer Concept specifically, the concept of top layer is top layer concept, and undermost concept is input data.It is from top layer concept first (hiding last layer), which generates, obtains the concept of hidden layer layer second from the bottom, then recognizes to obtain by the concept of hidden layer layer second from the bottom The concept of hidden layer layer third from the bottom, so analogizes, until obtaining the concept i.e. input data of input layer.

Or

The step S402 of the present embodiment is as shown in figure 9, specifically include：

S4023, using the second class testing input data as the input of the second deep learning neural network, cognition obtains second 4th top layer concept of deep learning neural network.

First example of the present embodiment passes through depth using tested speech as the input of the second deep learning neural network The cognitive process of degree study obtains the 4th top layer concept of the second deep learning neural network.

Second example of the present embodiment, using for the high-definition picture of test as the second deep learning neural network Input, the 4th top layer concept of the second deep learning neural network is obtained by the cognitive process of deep learning.

The third example of the present embodiment, using for test without incomplete data as the second deep learning neural network Input, the 4th top layer concept of the second deep learning neural network is obtained by the cognitive process of deep learning.

The cognitive process of deep learning specifically includes：

The concept on upper layer is generated by the concept and upward cognition (Encoder) weight of lower layer, Upper Concept compares lower layer Concept is more abstract, and undermost concept is input data, and the concept of top layer is top layer concept.It is to recognize from input data first Know to obtain the concept of hidden layer first layer, is then recognized to obtain the concept of the hidden layer second layer by the concept of hidden layer first layer, it is such as such It pushes away, the concept i.e. top layer concept until obtaining hiding last layer.

S4024, using the 4th top layer concept as the top layer concept of the first deep learning neural network, it is deep that generation obtains first Spend the first kind input data of learning neural network.

First example of the present embodiment, by the top layer that the 4th top layer concept is used as to the deep learning neural network Concept generates input data, as the corresponding head portrait of tested speech by the generating process of deep learning.

Second example of the present embodiment, by the top layer that the 4th top layer concept is used as to the deep learning neural network Concept generates input data, the corresponding low resolution of high-definition picture as test by the generating process of deep learning Image.

The third example of the present embodiment, by the top layer that the 4th top layer concept is used as to the deep learning neural network Concept generates input data by the generating process of deep learning, has incomplete data without incomplete data is corresponding as test.

The generating process of deep learning specifically includes：

The concept of lower layer is generated by the concept and downward generation (Decoder) weight on upper layer, lower layer's concept compares upper layer Concept specifically, the concept of top layer is top layer concept, and undermost concept is input data.It is from top layer concept first (hiding last layer), which generates, obtains the concept of hidden layer layer second from the bottom, then recognizes to obtain by the concept of hidden layer layer second from the bottom The concept of hidden layer layer third from the bottom, so analogizes, until obtaining the concept i.e. input data of input layer.

It is understood that although data type is different, first kind input data and the second class for having correspondence are defeated Enter the top layer concept that data may be matched by two-way deep learning, because having this two classes input data of correspondence Only class is different, and so as to cause details difference, but its key feature is consistent, thus its top layer concept is possible to reach one It causes, therefore passes through two-way supervised training so that two kinds of input datas are matched top layer is conceptive, so that this right It should be related to and can be able to store and express by two-way deep learning neural network, and then can be applied to have correspondence Conversion between inhomogeneity data.

Embodiment 4：

As shown in Figure 10, it present embodiments provides a kind of data correspondence based on two-way deep learning and judges system, The system includes the first two-way deep learning neural network module 1001 and data correspondence judgment module 1002.

The first two-way deep learning neural network module 1001, for establishing two-way deep learning nerve net Network learns the default correspondence between two class input datas；The first two-way deep learning neural network module 1001 As shown in figure 11, it specifically includes：

First acquisition unit 10011 for obtaining first kind input data, the second class input data, and has default The first kind input data of correspondence and the second class input data.

First unsupervised training unit 10012, for using first kind input data as the first deep learning neural network Input, carry out unsupervised training from bottom to top.

Second unsupervised training unit 10013, for using the second class input data as the second deep learning neural network Input, carry out unsupervised training from bottom to top.

First default correspondence relationship establishing unit 10014, the first kind input data for will be provided with default correspondence It is defeated with the second deep learning neural network respectively as inputting for the first deep learning neural network with the second class input data Enter, obtains the output of the first deep learning neural network as the first output, obtain the output of the second deep learning neural network As the second output, default pair is established between first kind input data, the second class input data, the first output, the second output It should be related to.

First supervised training unit 10015, the first kind input data and second for will be provided with default correspondence are defeated Go out the input data and output label respectively as the first deep learning neural network, carries out top-down supervised training.

Second supervised training unit 10016, the second class input data and first for will be provided with default correspondence are defeated Go out the input data and output label respectively as the second deep learning neural network, carries out top-down supervised training.

First stops supervised training unit 10017, for when the first output for having default correspondence and the second output Variance data amount and identical data amount ratio be less than the first predetermined threshold value when, stop top-down supervised training.

The data correspondence judgment module 1002, for judging two using the two-way deep learning neural network established Whether default correspondence is had between class testing input data, the data correspondence judgment module 1002 is as shown in figure 12, It specifically includes：

First deep learning unit 10021, for using the first class testing input data as the first deep learning nerve net The input of network, deep learning obtain the third output of the first deep learning neural network；

Second deep learning unit 10022, for using the second class testing input data as the second deep learning nerve net The input of network, deep learning obtain the 4th output of the second deep learning neural network；

Comparing unit 10023, the amount for comparing third output and the amount and identical data of the variance data of the 4th output, If the ratio that the third exports the amount with the amount of the variance data of the 4th output and identical data is less than the second predetermined threshold value, Have default correspondence between first class testing input data and the second class testing input data, does not otherwise have default correspondence Relationship.

Embodiment 5：

As shown in figure 13, a kind of data generation system based on two-way deep learning is present embodiments provided, the system packet Include the second two-way deep learning neural network module 1301 and data generation module 1302.

The second two-way deep learning neural network module 1301, for establishing two-way deep learning nerve net Network learns the default correspondence between two class input datas；The second two-way deep learning neural network module 1301 As shown in figure 14, it specifically includes：

Second acquisition unit 13011 for obtaining first kind input data, the second class input data, and has default The first kind input data of correspondence and the second class input data；

The unsupervised training unit 13012 of third, for using first kind input data as the first deep learning neural network Input, carry out unsupervised training from bottom to top；

4th unsupervised training unit 13013, for using the second class input data as the second deep learning neural network Input, carry out unsupervised training from bottom to top；

Second default correspondence relationship establishing unit 13014, the first kind input data for will be provided with default correspondence It is defeated with the second deep learning neural network respectively as inputting for the first deep learning neural network with the second class input data Enter, obtains the output of the first deep learning neural network as the first output, obtain the output of the second deep learning neural network As the second output, default pair is established between first kind input data, the second class input data, the first output, the second output It should be related to；

Third supervised training unit 13015, the first kind input data and second for will be provided with default correspondence are defeated Go out the input data and output label respectively as the first deep learning neural network, carries out top-down supervised training；

4th supervised training unit 13016, the second class input data and first for will be provided with default correspondence are defeated Go out the input data and output label respectively as the second deep learning neural network, carries out top-down supervised training；

Second stops supervised training unit 13017, for when the first output for having default correspondence and the second output Variance data amount and identical data amount ratio be less than the first predetermined threshold value when, stop top-down supervised training.

The data generation module 1302, for using the two-way deep learning neural network established, being inputted according to one kind Data generate another kind of input data；

The data generation module 1302 is as shown in figure 15, specifically includes：

Third deep learning unit 13021, for using the first class testing input data as the first deep learning nerve net The input of network, deep learning obtain the third output of the first deep learning neural network；

First reversed generation unit 13022, for third to be exported to the output mark as the second deep learning neural network Label, reversed generate obtain the second class input data of the second deep learning neural network；

Or

The data generation module 1302 is as shown in figure 16, specifically includes：

4th deep learning unit 13023, for using the second class testing input data as the second deep learning nerve net The input of network, deep learning obtain the 4th output of the second deep learning neural network；

Second reversed generation unit 13024, for the output mark by the 4th output as the first deep learning neural network Label, reversed generate obtain the first kind input data of the first deep learning neural network.

Embodiment 6：

The present embodiment based on the data generation system overall structure of two-way deep learning with embodiment 5, difference is in In the concrete composition part of the second two-way deep learning neural network module 1301 and data generation module 1302.

The second two-way deep learning neural network module 1301 of the present embodiment as shown in figure 17, specifically includes：

Third acquiring unit 13011 for obtaining first kind input data, the second class input data, and has default The first kind input data of correspondence and the second class input data；

5th unsupervised training unit 13012, for using first kind input data as the first deep learning neural network Input, carry out unsupervised training from bottom to top；

6th unsupervised training unit 13013, for using the second class input data as the second deep learning neural network Input, carry out unsupervised training from bottom to top；

Third presets correspondence relationship establishing unit 13014, the first kind input data for will be provided with default correspondence It is defeated with the second deep learning neural network respectively as inputting for the first deep learning neural network with the second class input data Entering, the top layer concept that the first deep learning neural network is obtained by the cognitive process of deep learning is used as the first top layer concept, The top layer concept that the second deep learning neural network is obtained by the cognitive process of deep learning is used as the second top layer concept, It is established between a kind of input data, the second class input data, the first top layer concept, the second top layer concept and presets correspondence, and Calculate the ratio of the variance data and identical data of the first top layer concept therein and the second top layer concept；

5th supervised training unit 13015, the first kind input data for will be provided with default correspondence and the second top Input data and expected top layer concept of the layer concept respectively as the first deep learning neural network, carry out top-down supervision Training；

6th supervised training unit 13016, the second class input data for will be provided with default correspondence and the first top Input data and expected top layer concept of the layer concept respectively as the second deep learning neural network, carry out top-down supervision Training；

Third stops supervised training unit 13017, for when the first top layer concept and second for having default correspondence When the variance data of top layer concept and the ratio of identical data are less than default discrepancy threshold, stop top-down supervised training.

The data generation module 1302 of the present embodiment is as shown in figure 18, specifically includes：

First deep learning recognizes unit 13021, for using the first class testing input data as the first deep learning god Input through network obtains the third top layer concept of the first deep learning neural network by the cognitive process of deep learning；

First generation unit 13022, it is raw using third top layer concept as the top layer concept of the second deep learning neural network At the second class input data of the second deep learning neural network is obtained, as corresponding second class of the first class testing input data Input data；

Or

The data generation module 1302 of the present embodiment is as shown in figure 19, specifically includes：

Second deep learning recognizes unit 13023, for using the second class testing input data as the second deep learning god Input through network, cognition obtain the 4th top layer concept of the second deep learning neural network；

Second generation unit 13024, for the 4th top layer concept is general as the top layer of the first deep learning neural network It reads, generation obtains the first kind input data of the first deep learning neural network.

It is appreciated that term " first ", " second " etc. used in the system of the various embodiments described above can be used for describing various Unit, but these units should not be limited by these terms.These terms are only used to distinguish first unit and another unit.It lifts For example, without departing from the scope of the invention, the first unsupervised training unit can be referred to as and be known as second without prison Training unit is superintended and directed, and similarly, the second unsupervised training unit can be known as to the first unsupervised training unit, the first unsupervised instruction Practice unit and the second unsupervised training unit both unsupervised training units, but it is not same unsupervised training unit.

The above, patent preferred embodiment only of the present invention, but the protection domain of patent of the present invention is not limited to This, any one skilled in the art is in the range disclosed in patent of the present invention, according to the skill of patent of the present invention Art scheme and its inventive concept are subject to equivalent substitution or change, belong to the protection domain of patent of the present invention.

Claims (10)

1. the data correspondence judgment method based on two-way deep learning, it is characterised in that：The method includes：

Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；

Judge whether have default corresponding pass between two class testing input datas using the two-way deep learning neural network of foundation System.

2. the data correspondence judgment method according to claim 1 based on two-way deep learning, it is characterised in that：Institute It states and establishes two-way deep learning neural network, learn the default correspondence between two class input datas, specifically include：

Obtain first kind input data, the second class input data, and have the first kind input data of default correspondence with Second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised training from bottom to top is carried out；

Using the second class input data as the input of the second deep learning neural network, unsupervised training from bottom to top is carried out；

First kind input data and the second class input data that will be provided with default correspondence are refreshing respectively as the first deep learning Input through network and the input of the second deep learning neural network obtain the output of the first deep learning neural network as the One output obtains the output of the second deep learning neural network as the second output, is inputted in first kind input data, the second class It is established between data, the first output, the second output and presets correspondence；

The first kind input data and second that will be provided with default correspondence are exported respectively as the first deep learning neural network Input data and output label, carry out top-down supervised training；

The the second class input data and first that will be provided with default correspondence are exported respectively as the second deep learning neural network Input data and output label, carry out top-down supervised training；

When the ratio of the first output and the amount of the variance data of the second output and the amount of identical data that have default correspondence When less than the first predetermined threshold value, stop top-down supervised training.

3. the data correspondence judgment method according to claim 2 based on two-way deep learning, it is characterised in that：Institute It states and judges whether have default correspondence between two class testing input datas using the two-way deep learning neural network of foundation, It specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, deep learning obtains the first depth Practise the third output of neural network；

Using the second class testing input data as the input of the second deep learning neural network, deep learning obtains the second depth Practise the 4th output of neural network；

Compare the amount of third output and the amount and identical data of the variance data of the 4th output, if the third exports and the 4th is defeated The ratio of the amount of the variance data gone out and the amount of identical data is less than the second predetermined threshold value, then the first class testing input data and the Have default correspondence between two class testing input datas, does not otherwise have default correspondence.

4. the data creation method based on two-way deep learning, it is characterised in that：The method includes：

Two-way deep learning neural network is established, the default correspondence between two class input datas is learnt；

Using the two-way deep learning neural network of foundation, another kind of input data is generated according to a kind of input data.

5. the data creation method according to claim 4 based on two-way deep learning, it is characterised in that：The foundation is double To deep learning neural network, learns the default correspondence between two class input datas, specifically include：

Obtain first kind input data, the second class input data, and have the first kind input data of default correspondence with Second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised training from bottom to top is carried out；

Using the second class input data as the input of the second deep learning neural network, unsupervised training from bottom to top is carried out；

First kind input data and the second class input data that will be provided with default correspondence are refreshing respectively as the first deep learning Input through network and the input of the second deep learning neural network obtain the output of the first deep learning neural network as the One output obtains the output of the second deep learning neural network as the second output, is inputted in first kind input data, the second class It is established between data, the first output, the second output and presets correspondence；

The first kind input data and second that will be provided with default correspondence are exported respectively as the first deep learning neural network Input data and output label, carry out top-down supervised training；

The the second class input data and first that will be provided with default correspondence are exported respectively as the second deep learning neural network Input data and output label, carry out top-down supervised training；

When the ratio of the first output and the amount of the variance data of the second output and the amount of identical data that have default correspondence When less than the first predetermined threshold value, stop top-down supervised training.

6. the data creation method according to claim 5 based on two-way deep learning, it is characterised in that：Described utilize is built Vertical two-way deep learning neural network generates another kind of input data according to a kind of input data, specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, deep learning obtains the first depth Practise the third output of neural network；

Third is exported into the output label as the second deep learning neural network, reversed generate obtains the second deep learning nerve Second class input data of network；

Or

Using the second class testing input data as the input of the second deep learning neural network, deep learning obtains the second depth Practise the 4th output of neural network；

Output label by the 4th output as the first deep learning neural network, reversed generate obtain the first deep learning nerve The first kind input data of network.

7. the data creation method according to claim 4 based on two-way deep learning, it is characterised in that：The foundation is double To deep learning neural network, learns the default correspondence between two class input datas, specifically include：

Obtain first kind input data, the second class input data, and have the first kind input data of default correspondence with Second class input data；

Using first kind input data as the input of the first deep learning neural network, unsupervised training from bottom to top is carried out；

Using the second class input data as the input of the second deep learning neural network, unsupervised training from bottom to top is carried out；

First kind input data and the second class input data that will be provided with default correspondence are refreshing respectively as the first deep learning The input of input and the second deep learning neural network through network, the first depth is obtained by the cognitive process of deep learning The top layer concept of neural network is practised as the first top layer concept, the second deep learning god is obtained by the cognitive process of deep learning Top layer concept through network is general in first kind input data, the second class input data, the first top layer as the second top layer concept It reads, established between the second top layer concept and preset correspondence, and calculate the first top layer concept therein and the second top layer concept The ratio of variance data and identical data；

First kind input data and the second top layer concept that will be provided with default correspondence are neural respectively as the first deep learning The input data of network and expected top layer concept, carry out top-down supervised training；

The second class input data and the first top layer concept that will be provided with default correspondence are neural respectively as the second deep learning The input data of network and expected top layer concept, carry out top-down supervised training；

When the ratio of the variance data and identical data of the first top layer concept and the second top layer concept that have default correspondence When less than default discrepancy threshold, stop top-down supervised training.

8. the data creation method according to claim 7 based on two-way deep learning, it is characterised in that：Described utilize is built Vertical two-way deep learning neural network generates another kind of input data according to a kind of input data, specifically includes：

Using the first class testing input data as the input of the first deep learning neural network, pass through the cognitive process of deep learning Obtain the third top layer concept of the first deep learning neural network；

Using third top layer concept as the top layer concept of the second deep learning neural network, generation obtains the second deep learning nerve Second class input data of network, as the corresponding second class input data of the first class testing input data；

Or

Using the second class testing input data as the input of the second deep learning neural network, cognition obtains the second deep learning god The 4th top layer concept through network；

Using the 4th top layer concept as the top layer concept of the first deep learning neural network, generation obtains the first deep learning nerve The first kind input data of network.

9. the data correspondence based on two-way deep learning judges system, it is characterised in that：The system comprises：

First two-way deep learning neural network module, for establishing two-way deep learning neural network, two classes of study are defeated Enter the default correspondence between data；

Data correspondence judgment module, for judging that two class testings input number using the two-way deep learning neural network established Whether has default correspondence between.

10. the data generation system based on two-way deep learning, it is characterised in that：The system comprises：

Second two-way deep learning neural network module, for establishing two-way deep learning neural network, two classes of study are defeated Enter the default correspondence between data；

Data generation module, for using the two-way deep learning neural network established, being generated according to a kind of input data another Class input data.