Enhanced generative adversarial network and target sample recognition method

Patent metadata

Field	Value
Publication number	US-12154036-B2
Application number	US-202016999118-A
Country	US
Kind code	B2
Filing date	Aug 21, 2020
Priority date	Jun 12, 2018
Publication date	Nov 26, 2024
Grant date	Nov 26, 2024

Abstract

The present disclosure relates to an enhanced generative adversarial network and a target sample recognition method. The enhanced generative adversarial network in the present disclosure includes at least one enhanced generator and at least one enhanced discriminator, where the enhanced generator obtains generated data by processing initial data, and provides the generated data to the enhanced discriminator; the enhanced discriminator processes the generated data and feeds back a classification result to the enhanced generator; the enhanced discriminator includes: a convolution layer, a basic capsule layer, a convolution capsule layer, and a classification capsule layer, and the convolution layer, the basic capsule layer, the convolution capsule layer, and the classification capsule layer are sequentially connected to each other.

First claim

What is claimed is:

1. An enhanced generative adversarial network, comprising at least one enhanced generator and at least one enhanced discriminator, wherein: the enhanced generator obtains generated data by processing initial data, and provides the generated data to the enhanced discriminator; the enhanced discriminator processes the generated data and feeds back a classification result to the enhanced generator; the enhanced discriminator comprises: a convolution layer, a basic capsule layer, a convolution capsule layer, and a classification capsule layer, and the convolution layer, the basic capsule layer, the convolution capsule layer, and the classification capsule layer are sequentially connected to each other; wherein quantities and structure parameters of the convolution layer, the basic capsule layer, the convolution capsule layer, and the classification capsule layer are set according to a feature of a target sample;
wherein there is more than one enhanced generator, there is more than one enhanced discriminator, the more than one enhanced generator generates new sample data by classification, and the more than one enhanced discriminator performs classification prediction on an unlabeled sample by forming an enhanced discriminator array.

2. The enhanced generative adversarial network according to claim 1 , wherein the enhanced generator obtains the generated data by using an unpooling layer, linear rectification, and a filtering layer to process the initial data.

3. A target sample recognition method using the enhanced generative adversarial network, comprising:
step a: constructing an enhanced generative adversarial network, wherein the constructed enhanced generative adversarial network comprises at least one enhanced generator and at least one enhanced discriminator; and
step b: constructing a multi-channel generative adversarial network based on the constructed enhanced generative adversarial network and a class feature of a target sample, performing label prediction on unlabeled data by using a trained multi-channel generative adversarial network, generating a sample of a corresponding class based on the enhanced generator, and accurately recognizing the target sample by utilizing the enhanced discriminator,
wherein step b comprises: classifying labeled raw data by class, and performing a data augmentation operation on each class of data; training an enhanced discriminator network; performing network training on the enhanced generator; inputting noise data, and generating new labeled data by using the enhanced generator; performing class prediction on the unlabeled data by using the enhanced discriminator; and classifying the target sample based on the discriminator in the multi-channel generative adversarial network.

4. The target sample recognition method according to claim 3 , wherein in step a, the enhanced generator obtains generated data by processing initial data, and provides the generated data to the enhanced discriminator, and the enhanced discriminator processes the generated data and feeds back a classification result to the enhanced generator.

5. The target sample recognition method according to claim 3 , wherein in step a, the enhanced discriminator comprises: a convolution layer, a basic capsule layer, a convolution capsule layer, and a classification capsule layer, and the convolution layer, the basic capsule layer, the convolution capsule layer, and the classification capsule layer are sequentially connected to each other.

6. The target sample recognition method according to claim 4 , wherein in step a, the enhanced discriminator comprises: a convolution layer, a basic capsule layer, a convolution capsule layer, and a classification capsule layer, and the convolution layer, the basic capsule layer, the convolution capsule layer, and the classification capsule layer are sequentially connected to each other.

7. The target sample recognition method according to claim 3 , wherein step a comprises: designing the enhanced discriminator based on a capsule mechanism by utilizing a capsule feature vectorization expression pattern; designing a “generation-discrimination” alternating optimization solution based on a Nash equilibrium capability of an enhanced generative adversarial model; and designing an objective function of the model by utilizing labeled and unlabeled samples.

8. The target sample recognition method according to claim 3 , wherein the performing class prediction on the unlabeled data by using the enhanced discriminator comprises:
selecting all the discriminators (D i opt , i=1, . . . , N), selecting any piece of data ({tilde over (x)}) from an unlabeled raw disease data set, and inputting the any piece of data ({tilde over (x)}) into each class of discriminator, wherein each discriminator determines a class (v 1 , V 2 , . . . . V N ) of the data, and outputs a number between 0 and 1, a number closer to 1 means a higher confidence level indicating that the class is determined, and if there are multiple output values close to 1, it indicates that training of the generator does not reach an optimal state, and training needs to continue,

0
≤

D
i
opt

(

x
~

)

≤
1

,

i
-

1
⁢

…
⁢

N

⁢

y
k

=

arg

max
i

{

D
1
opt

(

x
~

)

,

…
⁢

D
i
opt

(

x
~

)

,

…
⁢

D
N
opt

(

x
~

)

Assignees

• Shenzhen Inst Of Adv Tech Cas

Inventors

• Wang Shuqiang
• Shen Yanyan
• Zhang Wenyong

Classifications

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/045

  Combinations of networks · CPC title