CN111291810A - Information processing model generation method based on target attribute decoupling and related equipment - Google Patents
Information processing model generation method based on target attribute decoupling and related equipment Download PDFInfo
- Publication number
- CN111291810A CN111291810A CN202010080705.4A CN202010080705A CN111291810A CN 111291810 A CN111291810 A CN 111291810A CN 202010080705 A CN202010080705 A CN 202010080705A CN 111291810 A CN111291810 A CN 111291810A
- Authority
- CN
- China
- Prior art keywords
- loss
- feature
- feature map
- model
- information processing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides an information processing model generation method based on target attribute decoupling and related equipment, which are characterized in that a feature graph output by a hidden layer is obtained, and the feature graph is encoded by utilizing Hash encoding to obtain coordinate values corresponding to each feature graph; clustering each feature graph according to the coordinate values to obtain feature graph groups, respectively calculating orthogonal loss and/or suppression loss corresponding to the feature graphs in each feature graph group, obtaining a model total loss value according to the calculated orthogonal loss and/or suppression loss, adjusting model parameters by using the model total loss value, and repeating the steps until the training is completed to obtain the generated information processing model. The method provided by the embodiment reduces attribute coupling by mining semantic attributes of the potential layer and constructing the orthogonal loss of the clustering group, and reduces cross coupling according to attributes by performing intersection suppression on the feature maps in the intersection region, so that attribute coupling between the feature maps is reduced, and the generalization capability of the network is improved.
Description
Technical Field
The invention relates to the technical field of neural networks, in particular to an information processing model generation method based on target attribute decoupling and related equipment.
Background
Since Goodfellow proposed generating a confrontational network (GAN), it has achieved impressive performance in various attribute editions, such as expression editions, style migration, hair color transformation, gender transformation, age transformation, and the like. However, semantic dependencies resulting from feature map interactions in the GAN's generated network may compromise the generalization capability of the generated network.
Feature maps are easily highly coupled to each other when GAN is learned from a large set of training data. However, highly coupled information learned from training data sets is difficult to apply to test data sets because the coupling between different identities and data sets varies widely. Meanwhile, the coupling information makes it difficult to edit one attribute independently because the coupling easily affects the generation of other attributes at the same time, resulting in generation of much noise at the time of attribute editing, affecting the generation effect.
To reduce the correlation between different profiles, various network structures and loss functions have been proposed. For example, Gram-Schmidt orthogonalization is adopted to reduce the interference between nodes so as to improve the network generalization capability, but for a network with a large number of nodes or feature maps, the decoupling according to the nodes or the feature maps is inefficient.
Therefore, the prior art is subject to further improvement.
Disclosure of Invention
In view of the defects in the prior art, the invention aims to provide a target attribute decoupling-based information processing model generation method and related equipment for a user, and overcome the defect that the generalization capability of a generated network is poor due to the fact that the generation effect is influenced by the coupling of feature maps when a network structure in the prior art is edited.
The technical scheme adopted by the invention for solving the technical problem is as follows:
in a first aspect, the present embodiment provides an information processing method based on target attribute decoupling, including the steps of:
inputting sample information to be processed into a preset generated confrontation network model, and acquiring a feature diagram output by a hidden layer of the generated confrontation network model;
encoding each feature map by utilizing Hash codes to obtain coordinate values corresponding to each feature map;
clustering each feature map according to the coordinate values to obtain at least two feature map groups;
respectively calculating orthogonal loss and/or suppression loss corresponding to the feature maps in each feature map group, and obtaining a model total loss value according to the calculated orthogonal loss and/or suppression loss;
and adjusting the model parameters of the generated countermeasure network model according to the total model loss value, and repeating the steps of inputting the to-be-processed sample information to a preset generated countermeasure network model to obtain the total model loss value according to the calculated orthogonal loss and/or suppression loss until the calculated total network loss value meets the preset condition, thereby obtaining the trained information processing model.
Optionally, the step of calculating the orthogonal loss and/or the suppression loss corresponding to the feature maps in each feature map group, and obtaining a total model loss value according to the calculated orthogonal loss and/or the calculated suppression loss includes:
calculating orthogonal loss among feature graphs in each feature group, and obtaining a model total loss value according to the orthogonal loss;
and/or calculating the inhibition loss corresponding to the feature maps at the intersection in each feature map group, and obtaining the total model loss value according to the inhibition loss.
Optionally, the step of encoding each feature map by using hash coding to obtain coordinate values corresponding to each feature map includes:
and carrying out hash coding on the average characteristic diagram obtained by averaging the areas of the characteristic diagrams to obtain the coordinate value of each characteristic diagram.
Optionally, the step of obtaining the coordinate values of each feature map by performing hash coding on the average feature map obtained by averaging the regions of each feature map includes:
and selecting the first N average characteristic graphs with the strongest response in the average characteristic graphs corresponding to the characteristic graphs for carrying out Hash coding to obtain coordinate values corresponding to the characteristic graphs, wherein N is a positive integer.
Optionally, the step of clustering each feature map according to the coordinate values to obtain at least two feature map groups includes:
and clustering each feature map by using a k-means clustering algorithm according to the coordinate values to obtain at least two feature map groups.
Optionally, the orthogonal loss function expression corresponding to the orthogonal loss is as follows:
where m is the number of feature map groups, gi，gjIs the vector vectorized by the average feature map of the ith and jth groups, and m, i and j are positive integers.
Optionally, the generating a confrontation network model includes generating a network module and discriminating the network module;
the generating of the total loss function of the confrontation network model comprises generating a loss function of the network module, judging the loss function of the network module, an orthogonal loss function and a restraining loss function.
Optionally, an expression of a total loss function corresponding to the model total loss value is as follows:
wherein E represents data distribution, G represents a generation network module, and G (x, c') is used for generation by the generation network moduleInput image x and target domain label c' conditioned image, GsupRepresenting a generating network module with feature suppression and c is an original domain label; l issupThe method comprises the steps that a preset inhibition loss is obtained, and the influence deviation of used and unused characteristic inhibition on a characteristic diagram for counteracting the inhibited discard in the countermeasure network is represented; l isrecFor application to generating reconstruction losses in network modules, LOriGAnd LOriDRepresenting a loss function, λ, of the generating network module and the discriminating network modulesup、λrecAnd λGOIs a hyperparameter, LGAnd LDRespectively, generating the objective functions of the network module and the discriminating network module D.
In a second aspect, the present embodiment also discloses an electronic device, which includes a processor, and a storage medium communicatively connected to the processor, wherein the storage medium is adapted to store a plurality of instructions; the processor is adapted to invoke instructions in the storage medium to perform the steps of implementing the information processing based on the target property decoupling.
In a third aspect, a computer readable storage medium stores one or more programs, which are executable by one or more processors, for implementing the steps of the target property decoupling based information processing model generation method.
The invention has the beneficial effects that the invention provides an information processing model generation method based on target attribute decoupling and related equipment, firstly, the characteristic graphs output by all hidden layers are obtained, and each characteristic graph is coded by utilizing Hash codes to obtain the coordinate value corresponding to each characteristic graph; clustering and grouping the characteristic graphs according to the coordinate values, respectively calculating orthogonal loss and/or suppression loss corresponding to the characteristic graphs in each characteristic graph group, obtaining a model total loss value according to the calculated orthogonal loss and/or suppression loss, adjusting model parameters by using the model total loss value, and repeating the steps until the training is completed to obtain the generated information processing model. The embodiment reduces attribute coupling between feature maps and improves generalization capability of a generated network by mining semantic attributes of a potential layer and constructing orthogonal loss of a cluster group to reduce attribute coupling, and by performing intersection suppression on feature maps in intersection areas of different groups to reduce cross coupling according to attributes.
Drawings
FIG. 1 is a flowchart illustrating steps of a method for generating an information processing model for a countermeasure network based on target attribute decoupling according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating a principle of hash encoding a feature map according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of the principle of orthogonal suppression and intersection suppression in an embodiment of the present invention;
FIG. 4 is a schematic diagram illustrating a method for processing an intersection suppression loss pair feature map in a training process according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer and clearer, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Various non-limiting embodiments of the present invention are described in detail below with reference to the accompanying drawings.
Since existing decoupling algorithms are inefficient for networks with a large number of nodes or signature graphs, decoupling by node or by signature graph is inefficient. Meanwhile, semantic attributes of an object are often encoded by using a stack of network hidden outputs, and it is difficult to capture information about the semantic attributes by the operation of a feature mapping mode. Semantic attributes are often mined and pruned, and their decoupling can enable independent editing of each attribute and better generalization capability on test datasets. Therefore, the method disclosed by the embodiment of the invention mainly aims at the decoupling of semantic attributes to obtain a better decoupling effect. The semantic attribute represents the semantics of which attribute is represented by the feature in the network, for example, a feature represents that the GAN network is controlled to generate the attribute corresponding to the hair, and a feature represents that the GAN network is controlled to generate the attribute corresponding to the gender of the person.
In order to explore semantic attributes based on feature maps, Klemmer adopts local correlation of feature neighborhoods in the condition generation process. Wu and He uniformly divide the profile into groups and perform group-by-group normalization to simulate regularization by attribute.
In order to reduce the correlation between different feature maps in the generated countermeasure network GAN, namely, attribute coupling between the feature maps, the embodiment of the invention introduces semantic decomposition in the generation of the countermeasure network GAN to reduce the attribute correlation. The semantic decomposition is to classify and process the features representing different semantics, and comprises two methods of clustering group feature orthogonality and intersection inhibition. In this embodiment, feature decomposition is performed on the feature map to obtain a feature map group, and then decoupling processing is performed on the feature maps in different clustered feature map groups.
Exemplary method
The embodiment provides an information processing model generation method based on target attribute decoupling, as shown in fig. 1, including the steps of:
and step S1, inputting the sample information to be processed into a preset generated confrontation network model, and acquiring the characteristic diagram output by the generated confrontation network hidden layer.
The method of the embodiment of the invention performs decoupling processing aiming at the characteristic diagram output by the hidden layer of the generated countermeasure network. The generation of a competing network comprises two "competing" networks: the generation network module is used for capturing data distribution, and the discrimination network module is used for estimating the probability that one sample comes from real data instead of the generated sample. The generation network module and the judgment network module compete with each other so as to learn the required model. The hidden layer is other layers except the input layer and the output layer, and the method is suitable for target attribute decoupling of all feature graphs output by the hidden layer. Therefore, in this step, the feature maps output by the hidden layers are first obtained, and after being decoupled, the feature maps are input to the next hidden layer or the output layer.
In step S2, each feature map is encoded by hash encoding, and coordinate values corresponding to the respective feature maps are obtained.
Hash coding is the encoding of features using a hash algorithm, also called hash algorithm, which generally satisfies the relationship: and f, inputting data with any length, and outputting a fixed-length data key after the data are processed by a hash algorithm. Therefore, in this step, hash codes are used to encode each feature map, and coordinate values corresponding to each feature map are obtained.
Specifically, the step of encoding each feature map by using hash coding to obtain coordinate values corresponding to each feature map includes:
and carrying out hash coding on the average characteristic diagram obtained by averaging the areas of the characteristic diagrams to obtain the coordinate value of each characteristic diagram.
Examples are: referring to fig. 2, assuming that the hidden layer has 256 feature maps, each feature map is first divided into several patch blocks according to the area average, such as: firstly, dividing a hidden layer feature map with the size of 32 x 32 into 8 x 8 block patches, namely each block patch comprises 4 x 4 pixels, extracting the strongest response of topN (N is 2 in an experiment) in the block patches, splicing the N strongest responses to obtain an average feature map, carrying out hash coding on the average feature map, and identifying the obtained coordinates as coordinate values corresponding to the feature map. I.e., the signature is encoded by 256 x 32 to 256 x (N x 2), (N is required to be 2 because there are two values (x, y) per coordinate), and then the hash-coded coordinate values corresponding to the 256 signatures are obtained.
And step S3, clustering each feature map according to the coordinate values to obtain at least two feature map groups.
In this step, each feature map is clustered according to the coordinate values corresponding to each feature map, that is, each feature map is divided into at least two feature map groups according to the coordinate values. The clustering algorithm used in the step is a k-means clustering algorithm, namely, the k-means clustering algorithm is used for clustering each feature map according to the coordinate values to obtain at least two feature map groups.
The K-means clustering algorithm is characterized in that K objects are randomly selected as initial clustering centers at first, then Euclidean distances between each object and each clustering center are calculated, and each object is allocated to the clustering center closest to the object, so that each feature map group is obtained.
Further, the step of obtaining the coordinate values of each feature map by performing hash coding on the average feature map obtained by averaging the regions of each feature map includes:
and selecting the first N average characteristic graphs with the strongest response in the average characteristic graphs corresponding to the characteristic graphs for carrying out Hash coding to obtain coordinate values corresponding to the characteristic graphs, wherein N is a positive integer. In this embodiment, N is 2, that is, the number of feature map groups is 2.
Specifically, after each feature map is quantized, the obtained coordinate values are:
wherein (x)ij，yij) The patch coordinates (j is an arbitrary number from 1 to nc) in 8 × 8 patch of the jth maximum response from the large to the small in the plurality of average feature maps, i represents the ith feature map, and nc is set to 2. The average feature map refers to a feature map obtained by performing area average division on an original feature map output by a hidden layer, for example: the original feature map size of the hidden layer is 32 × 32, and directly taking the maximum response may be unstable and poor in robustness, so that the original feature map is averaged according to the region of 4 × 4, the size of the averaged feature map becomes 8 × 8, and then the averaged feature map is used for hash coding to obtain coordinate values corresponding to the original feature map.
And step S4, respectively calculating the orthogonal loss and/or the suppression loss corresponding to the feature maps in each feature map group, and obtaining a model total loss value according to the calculated orthogonal loss and/or the calculated suppression loss.
And the vectorized average feature maps are orthogonal pairwise to obtain an orthogonal result, and the vectorized average feature maps are data, so that pairwise orthogonality can be performed on the vectorized average feature maps to obtain orthogonal data.
In the step, two methods of clustering group feature orthogonality and intersection inhibition can be used, or one method is used for realizing the decoupling operation of the feature map.
Referring to fig. 3, a specific way of using the feature orthogonality of the cluster group is as follows:
and calculating the orthogonal loss among the characteristic graphs in each characteristic group, and obtaining the total loss value of the model according to the orthogonal loss.
Specifically, the orthogonal loss is added into a total loss function in the process of training the network, so that the characteristic diagram attribute coupling of the generated confrontation network model output after the training is finished is low.
Further, after clustering the feature maps, the method for calculating the orthogonal loss according to each feature map group obtained by clustering is to calculate the orthogonal loss by using a preset orthogonal loss function expression, specifically, the orthogonal loss function expression is as follows:
where m is the number of feature map groups, gi，gjIs the vector vectorized by the average feature map of the ith and jth groups, and m, i and j are positive integers.
And after clustering and grouping the characteristic graphs, calculating orthogonal loss values of the characteristic graphs in the group according to the grouped characteristic graphs by using the orthogonal loss function.
Further, the way of using intersection suppression is:
and calculating the inhibition loss corresponding to the feature maps at the intersection in each feature map group, obtaining a model total loss value according to the inhibition loss, and adding the inhibition loss to the model total loss value, so that decoupling of the output feature maps can be realized in the model training process.
In the specific implementation process, when only the orthogonal loss is considered, only the orthogonal loss is added to the model total loss value, if only the suppression loss is considered, only the suppression loss is added to the model total loss value, and if the optimal effect is to be obtained, the orthogonal loss and the suppression loss are added to the model total loss value at the same time, so that the target attribute of the feature map is decoupled. Because the distance between groups is increased by the orthogonality after clustering, and the distance inside the clustering group is reduced by the inhibition after clustering, the two can complement each other by combining, and therefore, a better attribute decoupling effect can be obtained.
Further, the generation of the countermeasure network comprises a generation network module and a discrimination network module;
the loss function comprises a loss value of the generation network module, a loss value of the discrimination network module, an orthogonal loss and a suppression loss. Wherein, the feature map orthogonality is gradually restrained in the training process through orthogonality loss; feature suppression is the suppression of loss constraints through intersection in the training process: and the inhibition loss is obtained by comparing the final effect graph generated by the feature graph at the intersection of the discarded hidden layers and the feature graph at the intersection of the unreleased hidden layers in the network during training. Both losses are increasingly constrained during the training process.
In one embodiment, the total loss function is expressed as:
wherein E represents data distribution, G represents generation of a network module, G (x, c ') generates an image conditioned on an input image x and a target domain label c' for the network module, GsupRepresenting a generating network module with feature suppression, and c is the original domain label, LsupIs a preset rejection loss, characterizing the influence deviation of used and unused characteristic rejection on counteracting rejected characteristic graphs in the reactive network, Lrec is the reconstruction loss applied in the generation network module, LOriGAnd LOriDRepresenting a loss function, λ, of the generating network module and the discriminating network modulesup、λrecAnd λGOIs a hyperparameter, LGAnd LDRespectively, generating the objective functions of the network module and the discriminating network module D.
Specifically, due to the feature map in each clustered feature map group, if the feature map is far away from the cluster center, the feature generation is inhibited, that is, the feature map at the intersection between the cluster groups is inhibited. As shown in fig. 3, feature maps at the intersection of two feature maps are discarded, the suppression manner adopted in this embodiment is to perform L1 distance loss constraint by discarding the feature map generated without discarding the feature map, and the smaller L1loss, the less feature maps at the intersection are generated, where L1 is a preset distance value away from the cluster center, and in a specific embodiment, a distance from the cluster center exceeding 1.5 times the average distance between each feature map in the group and the central feature map may be regarded as an abnormal distance.
Further, as shown in fig. 4, a method for processing an intersection suppression loss to a feature map in a training process is shown, where the suppression loss is obtained by comparing a final effect map generated by the feature map at the intersection of the discarded hidden layers in the network and the feature map at the intersection of the unrecessed hidden layers in the training process. The discarding mode is shown in fig. 4, and the features inside the circle (the radius is 1.5 times the average distance of each feature of the cluster group from the cluster center) are retained, and the feature outside the circle is considered as the feature map easily coupled at the intersection, so that the feature map is discarded. In the training process, the target attribute features in the two decoupled cluster groups of fig. 4 are slowly orthogonal to each other along with the training of the network, and the features at the intersection are less and less.
Step S5, adjusting the model parameters of the generated countermeasure network model according to the total model loss value, and repeating the steps from inputting the to-be-processed sample information to the preset generated countermeasure network model to obtaining the total model loss value according to the calculated orthogonal loss and/or suppression loss until the calculated total network loss value meets the preset condition, so as to obtain the trained information processing model.
The method comprises the following steps in each training process of generating the confrontation network model: inputting sample information to be processed into a preset generated confrontation network model, obtaining a feature diagram output by each hidden layer in the generated confrontation network model, respectively calculating orthogonal loss and suppression loss according to the output feature diagrams, respectively adding the orthogonal loss and the suppression loss into a model total loss value, correcting the model total loss value, and modifying a model parameter of the generated confrontation network model according to the corrected model total loss value.
Repeating the training process: inputting sample information to be processed into a generated confrontation network model, obtaining a characteristic diagram output from one or any selected target hidden layer, and repeatedly calculating orthogonal loss and suppression loss between the characteristic diagrams until a total network loss value of the generated confrontation network model corresponding to the modified model parameters meets a preset condition, thereby obtaining the trained information processing model.
The feature maps output by the hidden layers in the countermeasure network model are preset to be generated, and the feature maps are firstly grouped by using an effective clustering algorithm, wherein the algorithm is used for mining semantic attributes of the potential layers and constructing the orthogonal loss of the clustering groups so as to reduce attribute coupling. Meanwhile, the feature maps in the intersection areas of different groups are further suppressed to reduce cross coupling according to attributes, the target attributes of the feature maps obtained in the trained information processing model are decoupled, the feature maps of different target attributes are mutually orthogonal, the distribution of the feature maps of different attributes is obvious, and mutual intersection of the feature maps rarely occurs, so that the trained information processing model can have better generalization capability, and can be applied to most networks or data sets to obtain good information processing effect.
Exemplary device
The embodiment also discloses an electronic device, which comprises a processor and a storage medium in communication connection with the processor, wherein the storage medium is suitable for storing a plurality of instructions; the processor is adapted to invoke instructions in the storage medium to perform the steps of implementing the information handling model generation method based on target attribute decoupling.
The embodiment also discloses a computer readable storage medium, wherein the computer readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to realize the steps of the information processing model generation method based on target attribute decoupling.
The invention provides an information processing model generation method based on target attribute decoupling and related equipment, which are characterized in that a feature graph generated and output by a hidden layer of a countermeasure network model is obtained, and the feature graph is encoded by utilizing Hash encoding to obtain coordinate values corresponding to each feature graph; clustering each feature graph according to the coordinate values to obtain feature graph groups, respectively calculating orthogonal loss and/or suppression loss corresponding to the feature graphs in each feature graph group, obtaining a model total loss value according to the calculated orthogonal loss and/or suppression loss, adjusting model parameters by using the model total loss value, and repeating the steps until the training is completed to obtain the generated information processing model. The method for generating the information processing model provided by this embodiment reduces attribute coupling by mining semantic attributes of the potential layer and constructing orthogonal loss of the cluster group, and reduces cross coupling by attributes by performing intersection suppression on the feature maps in the intersection region, thereby reducing attribute coupling between the feature maps and improving generalization capability of the generated information processing model.
It should be understood that equivalents and modifications of the technical solution and inventive concept thereof may occur to those skilled in the art, and all such modifications and alterations should fall within the scope of the appended claims.
Claims (10)
1. An information processing model generation method based on target attribute decoupling is characterized by comprising the following steps:
inputting sample information to be processed into a preset generated confrontation network model, and acquiring a feature diagram output by a hidden layer of the generated confrontation network model;
encoding each feature map by utilizing Hash codes to obtain coordinate values corresponding to each feature map;
clustering each feature map according to the coordinate values to obtain at least two feature map groups;
respectively calculating orthogonal loss and/or suppression loss corresponding to the feature maps in each feature map group, and obtaining a model total loss value according to the calculated orthogonal loss and/or suppression loss;
and adjusting the model parameters of the generated countermeasure network model according to the total model loss value, and repeating the steps of inputting the to-be-processed sample information to a preset generated countermeasure network model to obtain the total model loss value according to the calculated orthogonal loss and/or suppression loss until the calculated total network loss value meets the preset condition, thereby obtaining the trained information processing model.
2. The method according to claim 1, wherein the step of calculating the orthogonality loss and/or the suppression loss corresponding to the feature maps in each feature map group, and obtaining the total model loss value according to the calculated orthogonality loss and/or suppression loss comprises:
calculating orthogonal loss among feature graphs in each feature group, and obtaining a model total loss value according to the orthogonal loss;
and/or calculating the inhibition loss corresponding to the feature maps at the intersection in each feature map group, and obtaining the total model loss value according to the inhibition loss.
3. The method for generating an information processing model based on target attribute decoupling according to claim 1, wherein the step of encoding each feature map by using hash encoding to obtain coordinate values corresponding to each feature map comprises:
and carrying out hash coding on the average characteristic diagram obtained by averaging the areas of the characteristic diagrams to obtain the coordinate value of each characteristic diagram.
4. The method according to claim 3, wherein the step of obtaining the coordinate values of the feature maps by performing hash coding on the average feature map obtained by averaging the regions of the feature maps comprises:
and selecting the first N average characteristic graphs with the strongest response in the average characteristic graphs corresponding to the characteristic graphs for carrying out Hash coding to obtain coordinate values corresponding to the characteristic graphs, wherein N is a positive integer.
5. The method for generating an information processing model based on target attribute decoupling according to claim 1, wherein the step of clustering each feature map according to the coordinate values to obtain at least two feature map groups comprises:
and clustering each feature map by using a k-means clustering algorithm according to the coordinate values to obtain at least two feature map groups.
6. The method for generating an information processing model based on target attribute decoupling according to claim 2, wherein the orthogonal loss function expression corresponding to the orthogonal loss is as follows:
where m is the number of feature map groups, gi，gjIs the vector after the vectorization of the average feature map of the ith and jth groups, and m, i and j are positive integers.
7. The method for generating an information processing model based on target attribute decoupling according to claim 2, wherein the generating a confrontation network model comprises generating a network module and discriminating the network module;
the generating of the total loss function of the confrontation network model comprises generating a loss function of the network module, judging the loss function of the network module, an orthogonal loss function and a restraining loss function.
8. The method for generating an information processing model based on target attribute decoupling according to claim 7, wherein an expression of a total loss function corresponding to the model total loss value is as follows:
wherein E represents data distribution, G represents a generation network module, G (x, c ') is used for generating an image with the condition of an input image x and a target domain label c', GsupRepresenting a generating network module with feature suppression and c is an original domain label; l issupThe method comprises the steps that a preset inhibition loss is obtained, and the influence deviation of used and unused characteristic inhibition on a characteristic diagram for counteracting the inhibited discard in the countermeasure network is represented; l isrecFor application to generating reconstruction losses in network modules, LOriGAnd LOriDRepresenting a loss function, λ, of the generating network module and the discriminating network modulesup、λrecAnd λGOIs a hyperparameter, LGAnd LDRespectively, generating the objective functions of the network module and the discriminating network module D.
9. An electronic device comprising a processor, a storage medium communicatively coupled to the processor, the storage medium adapted to store a plurality of instructions; the processor is adapted to call instructions in the storage medium to perform the steps of implementing the information processing model generation method based on target attribute decoupling of any of claims 1-8.
10. A computer readable storage medium, storing one or more programs, which are executable by one or more processors, to implement the steps of the information processing model generation method based on target attribute decoupling according to any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010080705.4A CN111291810B (en) | 2020-02-05 | 2020-02-05 | Information processing model generation method based on target attribute decoupling and related equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010080705.4A CN111291810B (en) | 2020-02-05 | 2020-02-05 | Information processing model generation method based on target attribute decoupling and related equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111291810A true CN111291810A (en) | 2020-06-16 |
| CN111291810B CN111291810B (en) | 2023-03-14 |
Family
ID=71024439
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010080705.4A Active CN111291810B (en) | 2020-02-05 | 2020-02-05 | Information processing model generation method based on target attribute decoupling and related equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111291810B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112396129A (en) * | 2020-12-08 | 2021-02-23 | 中山大学 | Countermeasure sample detection method and general countermeasure attack defense system |
| CN113408673A (en) * | 2021-08-19 | 2021-09-17 | 联想新视界(南昌)人工智能工研院有限公司 | Generation countermeasure network subspace decoupling and generation editing method, system and computer |
| CN114354185A (en) * | 2021-12-29 | 2022-04-15 | 重庆邮电大学 | Gear box fault diagnosis method based on decoupling network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106353738A (en) * | 2016-08-26 | 2017-01-25 | 西安空间无线电技术研究所 | Novel robust self-adaptive wave beam forming method under DOA mismatch conditions |
| CN109255769A (en) * | 2018-10-25 | 2019-01-22 | 厦门美图之家科技有限公司 | The training method and training pattern and image enchancing method of image enhancement network |
| CN109978792A (en) * | 2019-03-28 | 2019-07-05 | 厦门美图之家科技有限公司 | A method of generating image enhancement model |
| CN110415184A (en) * | 2019-06-28 | 2019-11-05 | 南开大学 | A kind of multi-modality images Enhancement Method based on orthogonal first space |
-
2020
- 2020-02-05 CN CN202010080705.4A patent/CN111291810B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106353738A (en) * | 2016-08-26 | 2017-01-25 | 西安空间无线电技术研究所 | Novel robust self-adaptive wave beam forming method under DOA mismatch conditions |
| CN109255769A (en) * | 2018-10-25 | 2019-01-22 | 厦门美图之家科技有限公司 | The training method and training pattern and image enchancing method of image enhancement network |
| CN109978792A (en) * | 2019-03-28 | 2019-07-05 | 厦门美图之家科技有限公司 | A method of generating image enhancement model |
| CN110415184A (en) * | 2019-06-28 | 2019-11-05 | 南开大学 | A kind of multi-modality images Enhancement Method based on orthogonal first space |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112396129A (en) * | 2020-12-08 | 2021-02-23 | 中山大学 | Countermeasure sample detection method and general countermeasure attack defense system |
| CN112396129B (en) * | 2020-12-08 | 2023-09-05 | 中山大学 | Challenge sample detection method and universal challenge attack defense system |
| CN113408673A (en) * | 2021-08-19 | 2021-09-17 | 联想新视界(南昌)人工智能工研院有限公司 | Generation countermeasure network subspace decoupling and generation editing method, system and computer |
| CN114354185A (en) * | 2021-12-29 | 2022-04-15 | 重庆邮电大学 | Gear box fault diagnosis method based on decoupling network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111291810B (en) | 2023-03-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111291810B (en) | Information processing model generation method based on target attribute decoupling and related equipment | |
| CN111080513B (en) | Attention mechanism-based human face image super-resolution method | |
| CN108171663B (en) | Image filling system of convolutional neural network based on feature map nearest neighbor replacement | |
| CN110941794A (en) | Anti-attack defense method based on universal inverse disturbance defense matrix | |
| KR20080021592A (en) | Image comparison by metric embeddings | |
| CN113361606A (en) | Deep map attention confrontation variational automatic encoder training method and system | |
| Goh et al. | A hybrid evolutionary algorithm for feature and ensemble selection in image tampering detection | |
| CN112613479B (en) | Expression recognition method based on light-weight streaming network and attention mechanism | |
| CN111709523A (en) | Width learning method based on internal integration | |
| CN114332500A (en) | Image processing model training method and device, computer equipment and storage medium | |
| CN114841296A (en) | Device clustering method, terminal device and storage medium | |
| CN115496144A (en) | Power distribution network operation scene determining method and device, computer equipment and storage medium | |
| Chen et al. | A finger vein recognition algorithm based on deep learning | |
| CN112329918A (en) | Anti-regularization network embedding method based on attention mechanism | |
| CN113515519A (en) | Method, device and equipment for training graph structure estimation model and storage medium | |
| CN111950635A (en) | Robust feature learning method based on hierarchical feature alignment | |
| Ayadi et al. | MIGSOM: multilevel interior growing self-organizing maps for high dimensional data clustering | |
| CN116306951A (en) | Quantum computing method and device, electronic equipment and storage medium | |
| CN112488238B (en) | Hybrid anomaly detection method based on countermeasure self-encoder | |
| CN115131605A (en) | Structure perception graph comparison learning method based on self-adaptive sub-graph | |
| Li et al. | A graphical approach for filter pruning by exploring the similarity relation between feature maps | |
| Gorban et al. | Robust principal graphs for data approximation | |
| CN113191380B (en) | Image evidence obtaining method and system based on multi-view features | |
| CN111898421B (en) | Regularization method for video behavior recognition | |
| CN113205175A (en) | Multi-layer attribute network representation learning method based on mutual information maximization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |