WO2019237240A1 - 一种增强型生成式对抗网络以及目标样本识别方法 - Google Patents

一种增强型生成式对抗网络以及目标样本识别方法 Download PDF

Info

Publication number
WO2019237240A1
WO2019237240A1 PCT/CN2018/090761 CN2018090761W WO2019237240A1 WO 2019237240 A1 WO2019237240 A1 WO 2019237240A1 CN 2018090761 W CN2018090761 W CN 2018090761W WO 2019237240 A1 WO2019237240 A1 WO 2019237240A1
Authority
WO
WIPO (PCT)
Prior art keywords
enhanced
data
discriminator
generative adversarial
generator
Prior art date
Application number
PCT/CN2018/090761
Other languages
English (en)
French (fr)
Inventor
王书强
申妍燕
张文勇
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2018/090761 priority Critical patent/WO2019237240A1/zh
Publication of WO2019237240A1 publication Critical patent/WO2019237240A1/zh
Priority to US16/999,118 priority patent/US20200380366A1/en

Links

Images

Classifications

    • 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
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2431—Multiple classes
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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

Definitions

  • the present invention relates to the field of computer application technology, and in particular, to an enhanced generative adversarial network and a method for identifying target samples.
  • Generative Adversarial Networks have received extensive attention and application in complex distributed unsupervised learning problems.
  • Generative Adversarial Network is a deep learning model.
  • the model includes two modules: a generator model (G) and a discriminator model (D).
  • GANs generate fairly good outputs through the mutual game learning of generators and discriminators.
  • Generators and discriminators usually consist of a multilayer network containing convolutional and / or fully connected layers. Generators and discriminators must be differentiable, but not necessarily directly reversible.
  • the training goals of GAN are to obtain the parameters that maximize the accuracy of the discriminator classification, and to obtain the generator parameters that maximize the deception discriminator.
  • the other party During the training process, one party is fixed, the other party ’s network weights are updated, and alternately, iteratively optimizes its own network to form a competitive confrontation until the two parties reach a dynamic equilibrium (Nash equilibrium). At this time, the generated model G resumes training
  • the data distribution that is, the generated data distribution is highly similar to the real data, and the discrimination model D is difficult to identify.
  • the Deep Convolutional Neural Network has the following improvements to the traditional GAN: (1) adding constraints on the basis of the convolutional GAN topology; (2) applying the trained discriminative network to image classification tasks; (3) ) Visualize the convolution kernel; (4) Make full use of the vector operation attributes of the generated model, and easily control the multi-semantic characteristics of the generated samples.
  • the prior art mainly has the following disadvantages: 1 Deep learning models, especially the widely used deep convolutional neural network models, need to rely on a large number of training samples to show their advantages. However, in some fields, there are fewer high-quality labeled samples. For example, in the field of medical imaging, a large number of high-quality labeled samples are scarce for any disease. Existing deep convolutional neural network models cannot obtain sufficient training without sufficient training samples, which greatly limits their application in the field of medical imaging. 2 The semi-supervised learning model is a feasible solution to the above problems, especially the semi-supervised learning model based on generative adversarial network (GAN), which can learn the overall distribution of data using limited labeled samples.
  • GAN generative adversarial network
  • GAN generative adversarial network
  • the present invention provides an enhanced generative adversarial network and a method for identifying target samples, which aims to solve at least one of the above-mentioned technical problems in the prior art.
  • the present invention provides the following technical solutions:
  • An enhanced generative adversarial network includes at least one enhanced generator and at least one enhanced discriminator.
  • the enhanced generator processes the obtained initial data to obtain generated data, and provides the generated data to the enhanced discriminator.
  • the enhanced discriminator processes the generated data and feeds back the classification results to the enhanced generator.
  • the enhanced discriminator includes a volume base layer, a basic capsule layer, a convolution capsule layer, and a classification capsule layer. , The basic capsule layer, the convolution capsule layer and the classification capsule layer are connected in order.
  • the technical solution adopted in the embodiment of the present invention further includes: the enhanced generator obtains the generated data through a de-pooling layer, a linear correction, and a filtering layer.
  • the technical solution adopted in the embodiment of the present invention further includes: the number of layers and the structural parameters of the roll base layer, the base capsule layer, the convolution capsule layer and the classification capsule layer are set according to the characteristics of the target sample.
  • the technical solution adopted in the embodiment of the present invention further includes: the number of the enhanced generators is greater than one, the number of the enhanced discriminator arrays is greater than one, and the plurality of enhanced generators generate new sample data by category
  • the plurality of enhanced discriminators form an enhanced discriminator array to perform class prediction on unlabeled samples.
  • Another technical solution adopted by the embodiment of the present invention is: a method for identifying a target sample, which is characterized by including:
  • Step a Construct an enhanced generative adversarial network, wherein the constructed enhanced generative adversarial network includes at least one enhanced generator and at least one enhanced discriminator
  • Step b construct a multi-channel generative adversarial network based on the class characteristics of the constructed enhanced generative adversarial network and target samples, use the trained multi-channel generative adversarial network to make label predictions on unlabeled data, and generate according to the enhanced generator The samples of the corresponding category are accurately identified by the enhanced discriminator.
  • the technical solution adopted in the embodiment of the present invention further includes: in the step a, the enhanced generator processes the obtained initial data to obtain generated data, and provides the generated data to the enhanced discriminator, and the enhanced discrimination The generator processes the generated data and feeds back the classification results to the enhanced generator.
  • the enhanced discriminator includes: a roll base layer, a base capsule layer, a convolution capsule layer and a classification capsule layer, the roll base layer and the base capsule layer
  • the convolution capsule layer and the classification capsule layer are connected in sequence.
  • the technical solution adopted by the embodiment of the present invention further includes: the step a includes: designing an enhanced discriminator based on the capsule mechanism by using a pattern of vectorized expression of capsule features; and designing “generating” based on the Nash equilibrium capability of the enhanced generation confrontation model. -Discriminant "alternative optimization scheme; use labeled and unlabeled samples to design the objective function of the model.
  • the technical solution adopted in the embodiment of the present invention further includes: the step b includes: first classifying the original labeled data into categories, performing data enhancement operations on each type of data separately, and training the enhanced discriminator network; Network training on the enhanced generator; input noise data, generate new labeled data through the enhanced generator; classify unlabeled data through the enhanced discriminator; target based on the discriminator in the multi-channel generative adversarial network Samples are classified.
  • the technical solution adopted in the embodiment of the present invention further includes: performing the class prediction on the unlabeled data by the enhanced discriminator includes: arbitrarily extracting a piece of data from the original unlabeled disease data set and inputting the data to each class of the discriminator, Each discriminator will judge this data category and output a number between 0 and 1. The closer the number is to 1, the higher the confidence of the class being judged; if there are multiple output values close to 1, it means that the The training of the device has not reached the optimal state, and it needs to continue to return to training.
  • the technical solution adopted in the embodiment of the present invention further includes: the three-dimensional wireframe reconstruction module reconstructing the three-dimensional wireframe specifically includes: performing initial matching of the two-dimensional line segments extracted from different images according to the geometrical constraints of the epipolar line, and using three-dimensional point clouds to perform matching.
  • the three-dimensional model building module constructs three-dimensional
  • the model includes: using the RANSAC method to generate a large number of candidate planes for the line segments in the three-dimensional wire frame, and combining the point cloud to optimize the selection to obtain a closed three-dimensional model.
  • the embodiment of the present invention has the beneficial effects that the enhanced generative adversarial network and the target sample recognition method of the embodiment of the present invention can effectively solve the problem of pattern recognition under the condition of a small number of labeled samples, and achieve the accuracy of the target sample.
  • Identification; the enhanced generative adversarial network and target sample identification method of the embodiment of the present invention use multiple enhanced generators in the multi-channel generative adversarial network to generate new sample data by category; at the same time, multiple channels can also be used
  • Multiple enhanced discriminators in the generative adversarial network form an enhanced discriminator array, and perform class prediction on unlabeled samples. As a result, the number of labeled data is increased, which makes it easier to train a complex deep convolutional neural network classifier.
  • the enhanced generative adversarial network and target sample recognition method of the embodiment of the present invention have the following significant features and Positive effect:
  • multiple enhanced generators learn the respective sub-distributions of different categories of data in the original database, which makes it easier for network training to reach the Nash balance and makes it easier to train the generative adversarial network .
  • the amount of labeled data can be increased significantly, thereby solving the problem of obtaining a large amount of high-quality labeled data in some fields, such as the field of medical imaging. Difficult problems can be more convenient to train classification models based on deep convolutional neural networks.
  • FIG. 1 is a schematic structural diagram of an enhanced generative adversarial network according to a first embodiment of the present invention
  • FIG. 2 is a schematic structural diagram of an enhanced generative adversarial network according to a second embodiment of the present invention.
  • FIG. 3 is a flowchart of a target sample recognition method according to an embodiment of the present invention.
  • FIG. 1 is a schematic structural diagram of an enhanced generative adversarial network according to a first embodiment of the present invention.
  • the enhanced generative adversarial network of the embodiment of the present invention includes an enhanced generator and an enhanced discriminator.
  • the enhanced generator will obtain the initial data through the de-pooling layer, linear correction, and filtering layer to obtain the generated data, and provide Generate the data to the enhanced discriminator.
  • the enhanced discriminator processes the data and feeds back the classification results to the enhanced generator.
  • the generated data can be generated pictures or other types of data.
  • the enhanced discriminator includes: a roll base layer, a base capsule layer, a convolution capsule layer and a classification capsule layer, a roll base layer, a base capsule layer, a convolution capsule layer and a classification capsule layer are connected in order.
  • the number of layers and structural parameters of the volume base layer, the basic capsule layer, the convolution capsule layer and the classification capsule layer depend on the characteristics of the target sample.
  • the enhanced generative adversarial network of the embodiment of the present invention proposes an enhanced discriminator based on the capsule mechanism on the basis of the existing deep convolutional generative adversarial network, and expresses the directional information features and hierarchical information features of the samples through vectorization to improve
  • the discriminator's sample feature expression and feature recognition capabilities enhance the generator's ability to learn the overall distribution of real data.
  • FIG. 2 is a schematic structural diagram of an enhanced generative adversarial network according to a second embodiment of the present invention.
  • the enhanced generative adversarial network of the embodiment of the present invention uses multiple enhanced generators in the multi-channel generative adversarial network to generate new sample data by category; at the same time, it is also possible to use multiple channels in the multi-channel generative adversarial network.
  • the enhanced discriminators form an enhanced discriminator array and perform class prediction on unlabeled samples. This increases the amount of labeled data and makes it easier to train complex deep convolutional neural network classifiers.
  • FIG. 3 is a flowchart of a target sample recognition method according to an embodiment of the present invention.
  • a method for identifying a target sample according to an embodiment of the present invention includes:
  • Step 100 construct an enhanced generative adversarial network based on the capsule mechanism
  • step 100 specifically includes: first, designing an enhanced discriminator based on the capsule mechanism, using the vectorized expression pattern of capsule features to improve the ability of feature extraction and feature recognition of the discriminator, and further spreading the gradient information to improve generator learning The ability of the sample to distribute as a whole, thereby enhancing the discrimination and generating ability of the entire generative adversarial model; secondly, designing a suitable "generating-discriminating" alternate optimization scheme to improve the ability of the enhanced generative adversarial model to reach the Nash equilibrium and improve the robustness of the model Finally, the objective function of the model is designed so that the model can make full use of labeled samples and unlabeled samples, improving the model's ability to quickly learn the overall distribution of data under the condition of limited labeled samples, and achieve accurate identification of samples.
  • this solution proposes an enhanced discriminator based on the capsule mechanism.
  • the vector information is used to express the directional information features and hierarchical information features of the sample to improve the sample feature expression and feature recognition of the discriminator. Capabilities, thereby enhancing the generator's ability to learn the overall distribution of real data.
  • FIG. 1 is a schematic diagram of an enhanced generative adversarial network structure.
  • the enhanced generative adversarial network includes an enhanced generator and an enhanced discriminator.
  • the enhanced generator will obtain the initial data through the de-pooling layer, linear correction, and filtering layer to obtain the generated image, and provide the generated image to the enhanced type.
  • Discriminator, enhanced discriminator processes the picture and feeds back the classification result to the enhanced generator.
  • the enhanced discriminator includes: a roll base layer, a base capsule layer, a convolution capsule layer and a classification capsule layer, a roll base layer, a base capsule layer, a convolution capsule layer and a classification capsule layer are connected in order.
  • the number of layers and the structural parameters of each layer depend on the characteristics of the target sample.
  • Step 200 Construct a multi-channel generative adversarial network, use the trained multi-channel generative adversarial network to perform label prediction on unlabeled data, generate samples of corresponding categories according to the optimized generator, and use the optimized discriminator to target samples Perform precise identification.
  • step 200 first, based on the enhanced generative adversarial network, a multi-channel enhanced generative adversarial network is designed based on the characteristics of the target sample, and different types of samples are used to train the generative adversarial network of each channel separately; Second, use the trained multi-channel generative adversarial network to make label predictions on unlabeled data, and use the optimized generator to generate samples of the corresponding category, so as to obtain a more adequate labeled sample. Finally, using the optimized discriminator, Achieve accurate identification of target samples.
  • this scheme assumes that the number of sample categories is N.
  • N-type data set x ⁇ X 1 , X 2 , ... X N ⁇
  • the overall distribution of the data set x follows the distribution x ⁇ ⁇ ;
  • Each type of data subset can be represented as Obey distribution
  • Step 200 specifically includes:
  • Step 210 perform data enhancement on the labeled samples
  • step 220 the original labeled data is classified by category first, and common data enhancement operations are performed on each type of data.
  • the data enhancement operations include rotation, flipping, scaling, etc., and the expanded data set is sufficient for training.
  • Corresponding multi-channel confrontation generation network Corresponding multi-channel confrontation generation network.
  • Step 220 training the enhanced discriminator network
  • step 220 for the samples containing N classes, during the model training process, N adversarial generation networks are trained one by one.
  • N adversarial generation networks are trained one by one.
  • the i-th channel confrontation generation network only the corresponding type i data is input.
  • the corresponding optimized discriminator training objective function is:
  • Loss (G) represents the Loss Function of the generator G. It measures the difference between the real data (P data (x)) and the generated data (P G (x)).
  • the parameters ⁇ D of the discriminator network D are initialized and the number of iterations n D of the discriminator training is set.
  • n D of the discriminator training is set.
  • ⁇ x 1 , x 2 , x 3 , ... x m ⁇ are sampled from the real data.
  • Sample some data ⁇ z 1 , z 2 , z 3 , ... z m ⁇ (normally or uniformly) from the noise data.
  • Calculate some penalty data ⁇ is a random number between 0 and 1; then the parameters can be updated:
  • the enhanced discriminator can obtain better performance.
  • the remaining N-1 enhanced discriminators are trained to form a group of N enhanced discriminator groups.
  • Step 230 Perform network training on the enhanced generator
  • step 230 for a group of multi-channel adversarial generation networks, N enhanced generators are also included.
  • the training data is divided into N categories by category, and the corresponding enhanced generator is trained using the data of each category.
  • the optimized generator training objective function is:
  • V (G, D) is the same as that in step 2.
  • the parameters ⁇ G of the generator network G need to be initialized and the number of generator training iterations n G is set . Then sample some data ⁇ z 1 , z 2 , z 3 , ... z m ⁇ (normal distribution or uniform distribution) from the noise data; then the parameters can be updated:
  • Step 240 input the noise data, and generate new label data through the enhanced generator
  • step 240 after obtaining a set of trained multi-channel generative adversarial networks, remove all enhanced generators For each enhanced network, inputting a set of noise data (Gaussian distribution, uniform distribution, etc.) will generate a new set of labeled data, whose labeling category is the category corresponding to the generator:
  • noise data Gausian distribution, uniform distribution, etc.
  • Representative data newly generated class i, Z i represents the i class initialization noise.
  • the same operation on a generator can generate new label data. This can increase the amount of each type of data and also increase the size of the data set.
  • Step 250 Perform class prediction on the unlabeled data through an enhanced discriminator
  • step 250 after obtaining a set of trained multi-channel generative adversarial networks, remove all the generators Take an arbitrary copy of the data from the original unlabeled disease dataset Entered separately into each class of discriminator.
  • Each discriminator will judge this data category (y 1 , y 2 , ... y N ), and output a number between 0-1. The closer the number is to 1, the higher the confidence level of the class. Ideally, only one of the discriminators will have an output value close to 1, and the other output values will be close to 0. In this way, the category of the input data can be determined to be the category with the largest output value. If there are multiple output values close to 1, it means that the training of the generator has not reached the optimal state, and it needs to continue to return to training.
  • Step 260 Classify the target samples based on the discriminator in the multi-channel generative adversarial network.
  • the N enhanced discriminators correspond to N categories of disease data, respectively.
  • These enhanced discriminators can directly perform classification prediction on test samples.
  • an arbitrary test sample (d k ) is taken and input into each enhanced discriminator ().
  • Each enhanced discriminator (D s ) will hard output a number between 0 and 1 for the input sample (d k ). The larger the number, the higher the confidence that the discriminator judges the test sample as the corresponding category.
  • the enhanced generative adversarial network and target sample recognition method propose an enhanced generative adversarial network incorporating a capsule mechanism.
  • the network uses vectorized patterns to express image features and significantly enhances the discriminator.
  • the ability to recognize real samples and generated samples further improves the generator's ability to learn the overall distribution of real data, making it easier for the generation of adversarial models to reach the Nash equilibrium.
  • the enhanced generative adversarial network and the target sample identification method further propose a multi-channel generative adversarial model.
  • N-type data set x ⁇ X 1 , X 2 , ... X N ⁇
  • the overall distribution of the data set x follows the distribution x ⁇ ⁇ ; each type of data subset can be expressed as Obey distribution
  • N only needs to learn the true sample distribution of the corresponding category
  • the method of this project makes it easier for N enhanced generators to reach the Nash equilibrium state, thereby reducing the problem of difficult training of generative adversarial networks.
  • the multi-channel optimization generator can generate labeled high-quality data by category:
  • the enhanced generative adversarial network and the target sample recognition method in the embodiments of the present invention can effectively solve the problem of pattern recognition under the condition of a small number of labeled samples, and achieve accurate recognition of target samples.
  • the enhanced generative adversarial network and the target sample identification method of the embodiment of the present invention use multiple enhanced generators in the multi-channel generative adversarial network to generate new sample data by category; at the same time, a multi-channel generative method can also be used Multiple enhanced discriminators in the adversarial network form an enhanced discriminator array, and perform class prediction on unlabeled samples. This increases the amount of labeled data and makes it easier to train complex deep convolutional neural network classifiers.
  • the enhanced generative adversarial network and the method for identifying target samples according to the embodiments of the present invention have the following significant features and positive effects:
  • By training a multi-channel generative adversarial network multiple enhanced generators learn the original
  • the respective sub-distributions of different categories of data in the database make it easier for network training to reach Nash equilibrium and make it easier to train generative adversarial networks.
  • the amount of labeled data can be increased significantly, thereby solving the problem of obtaining a large amount of high-quality labeled data in some fields, such as the field of medical imaging Difficult problems can be more convenient to train classification models based on deep convolutional neural networks.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

本发明涉及计算机应用技术领域，特别涉及一种增强型生成式对抗网络以及目标样本识别方法。本发明增强型生成式对抗网络包括至少一个增强型生成器和至少一个增强型判别器，所述增强型生成器将获得的初始数据进行处理得到生成数据，并提供生成数据给增强型判别器，所述增强型判别器对生成数据进行处理，并反馈分类结果给增强型生成器，所述增强型判别器包括卷基层、基础胶囊层、卷积胶囊层和分类胶囊层，所述卷基层、基础胶囊层、卷积胶囊层和分类胶囊层依次相连。

Description

一种增强型生成式对抗网络以及目标样本识别方法 技术领域
本发明涉及计算机应用技术领域，特别涉及一种增强型生成式对抗网络以及目标样本识别方法。
背景技术
生成式对抗网络(GAN，Generative Adversarial Networks)近年来在复杂分布的无监督学习问题上得到广泛关注和应用。生成式对抗网络(GAN)是一种深度学习模型。模型包括两个模块：生成器模型(G)和判别器模型(D)，GAN通过生成器和判别器的互相博弈学习产生相当好的输出。生成器和判别器通常由包含卷积和(或)全连接层的多层网络构成。生成器和判别器必须是可微的，但并不必要是直接可逆的。GAN的训练目标是获得令判别器分类准确率最大化的参数，以及获得最大化欺骗判别器的生成器参数。在训练的过程中固定一方，更新另一方的网络权重，交替进行，分别迭代优化自己的网络，从而形成竞争对抗，直到双方达到一个动态的平衡(纳什均衡)，此时生成模型G恢复了训练数据的分布，即生成的数据分布与真实数据高度相似，判别模型D难以识别。
深度卷积神经网络(DCGAN)对传统的GAN主要有如下改进：(1)在卷积GAN拓扑结构的基础上增加约束条件；(2)将训练后的判别网络应用于图像分类任务；(3)对卷积核进行了可视化；(4)充分利用了生成模型的向量运算属性，容易控制生成样本的多语义特性。
现有技术主要有如下缺点：①深度学习模型，尤其是应用广泛的深度卷积神经网络模型，需要依靠大量的训练样本，才能显示出优势。但是，在一些 领域中，高质量的标注样本较少，譬如，在医学影像领域，大量高质量标注样本对任何一种疾病都是稀缺的。现有的深度卷积神经网络模型在缺少足够训练样本的条件下，无法获得充分训练，这极大地限制了其在医学影像领域的应用。②半监督学习模型是解决以上问题的一种可行方案，尤其是基于生成式对抗网络(GAN)的半监督学习模型，它可以利用有限的标注样本学习数据的整体分布。但是现有生成式对抗网络(GAN)模型通常存在如下问题：(a)GAN难训练，尤其是分布复杂的多类别数据条件下；(b)GAN易引起模型崩溃，从而导致生模型失效。
发明内容
本发明提供了一种增强型生成式对抗网络以及目标样本识别方法，旨在至少在一定程度上解决现有技术中的上述技术问题之一。
为了解决上述问题，本发明提供了如下技术方案：
一种增强型生成式对抗网络，包括至少一个增强型生成器和至少一个增强型判别器，所述增强型生成器将获得的初始数据进行处理得到生成数据，并提供生成数据给增强型判别器，所述增强型判别器对生成数据进行处理，并反馈分类结果给增强型生成器，所述增强型判别器包括卷基层、基础胶囊层、卷积胶囊层和分类胶囊层，所述卷基层、基础胶囊层、卷积胶囊层和分类胶囊层依次相连。
本发明实施例采取的技术方案还包括：所述增强型生成器将获得的初始数据通过反池化层、线性修正、过滤层得到生成数据。
本发明实施例采取的技术方案还包括：所述卷基层、基础胶囊层，卷积胶囊层和分类胶囊层的层数和结构参数根据目标样本的特征进行设定。
本发明实施例采取的技术方案还包括：所述增强型生成器的数量大于1个，所述增强型判别器阵列数量大于1个，所述多个增强型生成器按类别生成 新的样本数据，所述多个增强型判别器组成增强型判别器阵列，对未标注样本进行类别预测。
本发明实施例采取的另一技术方案为：一种目标样本识别方法，其特征在于，包括：
步骤a：构建增强型生成式对抗网络，其中，构建的增强型生成式对抗网络包括至少一个增强型生成器和至少一个增强型判别器
步骤b：根据构建的增强型生成式对抗网络和目标样本的类别特点构建多通道生成式对抗网络，利用训练过的多通道生成式对抗网络对无标注数据进行标签预测，根据增强型生成器生成相应类别的样本，利用增强型判别器对目标样本进行精准识别。
本发明实施例采取的技术方案还包括：在所述步骤a中，所述增强型生成器将获得的初始数据进行处理得到生成数据，并提供生成数据给增强型判别器，所述增强型判别器对生成数据进行处理，并反馈分类结果给增强型生成器。
本发明实施例采取的技术方案还包括：在所述步骤a中，所述增强型判别器包括：卷基层、基础胶囊层，卷积胶囊层和分类胶囊层，所述卷基层、基础胶囊层，卷积胶囊层和分类胶囊层依次相连。
本发明实施例采取的技术方案还包括：所述步骤a包括：利用胶囊特征向量化表达的模式，设计基于胶囊机制的增强型判别器；根据增强型生成对抗模型的纳什平衡能力，设计“生成-判别”交替优化方案；利用有标注样本和无标注样本，设计模型的目标函数。
本发明实施例采取的技术方案还包括：所述步骤b包括：对于原始带标注的数据先进行按类别分类，分别对每一类的数据进行数据增强操作；对增强型判别器网络进行训练；对增强型生成器进行网络训练；输入噪声数据，通过增强型生成器生成新的标注数据；通过增强型判别器对无标注数据进行类别预测；基于多通道生成式对抗网络中的判别器对目标样本进行分类。
本发明实施例采取的技术方案还包括：所述通过增强型判别器对无标注数据进行类别预测包括：从原始未标注疾病数据集中任意取出一份数据分别输入到每一类的判别器当中，每一个判别器都会对此数据类别进行判断输出一个0到1之间的数字，数字越接近1意味着被判定为该类的置信度越高；如果存在多个接近1的输出值，说明生成器的训练并没有达到最优化的状态，需要继续返回训练。
本发明实施例采取的技术方案还包括：所述三维线框重建模块重建三维线框具体包括：将不同图像上提取的二维线段根据极线几何约束进行初始匹配，使用三维点云对匹配进行验证，根据相机位置和旋转量将每一对匹配好的二维线段重建出一条三维线段，对所有重建出的三维线段进行聚类合并得到最终的三维线框；所述三维模型构建模块构建三维模型包括：对三维线框中的线段使用RANSAC方法产生大量候选平面，结合点云进行优化选择得到封闭的三维模型。
相对于现有技术，本发明实施例产生的有益效果在于：本发明实施例的增强型生成式对抗网络以及目标样本识别方法可以有效解决少量标注样本条件下的模式识别问题，实现目标样本的精准识别；本发明实施例的增强型生成式对抗网络以及目标样本识别方法使用多通道生成式对抗网络中的多个增强型生成器进行按类别地生成新的样本数据；同时，也可以使用多通道生成式对抗网络中的多个增强型判别器组成一个增强型判别器阵列，并对未标注样本进行类别预测。从而增加了标注数据的数量，使得训练复杂的深度卷积神经网络分类器更加容易；本发明实施例的增强型生成式对抗网络以及目标样本识别方法与现有技术相比，具有以下显著特点和积极效果：通过训练多通道生成式对抗网络使得多个增强型生成器分别学习到原始数据库中不同类别数据各自的子分布，这使得网络训练更容易达到纳什平衡，使得生成式对抗网络更容易训练。通过多通道生成式对抗网络中的多个增强型生成器和相应的增强型判别器阵列，可以大量地增加标注数据量，从而解决了在一些领域，譬如医学影像领域， 大量高质量标注数据获取困难的问题，可以更加方便训练基于深度卷积神经网络的分类模型。
附图说明
图1是本发明第一实施例的增强型生成式对抗网络的结构示意图；
图2是本发明第二实施例的增强型生成式对抗网络的结构示意图；
图3是本发明实施例的目标样本识别方法的流程图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
请参阅图1，图1是本发明第一实施例的增强型生成式对抗网络的结构示意图。本发明实施例的增强型生成式对抗网络包括增强型生成器和增强型判别器，增强型生成器将获得初始数据通过反池化层、线性修正、过滤(Filtering)层得到生成数据，并提供生成数据给增强型判别器，增强型判别器对数据进行处理后反馈分类结果给增强型生成器，其中，生成数据可以是生成图片或者其他类型的数据。增强型判别器包括：卷基层、基础胶囊层，卷积胶囊层和分类胶囊层，卷基层、基础胶囊层，卷积胶囊层和分类胶囊层依次相连。卷基层、基础胶囊层，卷积胶囊层和分类胶囊层的层数的多少和结构参数根据目标样本的特征而定。
本发明实施例的增强型生成式对抗网络在现有深度卷积生成式对抗网络 的基础上，提出基于胶囊机制的增强型判别器，通过向量化表达样本的方向信息特征、层次信息特征，提升判别器的样本特征表达和特征识别能力，从而增强生成器学习真实数据整体分布的能力。
请参阅图2，图2是本发明第二实施例的增强型生成式对抗网络的结构示意图。本发明实施例的增强型生成式对抗网络通过构建N个增强型生成式对抗网络组成多通道生成式对抗网络，使得每一个增强型生成器(G i，i＝1…N)只需要学习对应类别的真实样本分布
Figure PCTCN2018090761-appb-000001
相比于普通的生成器需要学习所有样本的整体分布(φ)而言，本发明实施例的增强型生成式对抗网络使得N个增强型生成器更容易达到纳什平衡状态，从而降低了生成式对抗网络难训练的问题，提升了生成式对抗网络的普适性。经过多次迭代训练，可以得到最优的判别器
Figure PCTCN2018090761-appb-000002
和生成器
Figure PCTCN2018090761-appb-000003
利用多通道最优化生成器可以按类别生成带标注高质量数据：
Figure PCTCN2018090761-appb-000004
其中，
Figure PCTCN2018090761-appb-000005
代表新生成第i类数据，Z i代表第i类初始化噪声。同时利用多通道最优化判别器可以对任意未标注数据
Figure PCTCN2018090761-appb-000006
进行标签预测：
Figure PCTCN2018090761-appb-000007
Figure PCTCN2018090761-appb-000008
本发明实施例的增强型生成式对抗网络使用多通道生成式对抗网络中的多个增强型生成器进行按类别地生成新的样本数据；同时，也可以使用多通道生成式对抗网络中的多个增强型判别器组成一个增强型判别器阵列，并对未标注样本进行类别预测。从而增加了标注数据的数量，使得训练复杂的深度卷积神经网络分类器更加容易。
请参阅图3，图3是本发明实施例的目标样本识别方法的流程图。本发明实施例的目标样本识别方法包括：
步骤100：构建基于胶囊机制的增强型生成式对抗网络；
其中，步骤100具体包括：首先，设计基于胶囊机制的增强型判别器，利 用胶囊特征向量化表达的模式，提升判别器特征提取和特征识别的能力，并进一步通过梯度信息传播，提升生成器学习样本整体分布的能力，从而增强整个生成对抗模型的判别能力和生成能力；其次，设计合适的“生成-判别”交替优化方案，提高增强型生成对抗模型达到纳什平衡的能力，提高模型的鲁棒性；最后，设计模型的目标函数，使模型能够充分利用有标注样本和无标注样本，提升模型在有限的标注样本条件下快速学习数据整体分布的能力，实现样本的精准识别。
本方案在现有深度卷积生成式对抗网络的基础上，提出基于胶囊机制的增强型判别器，通过向量化表达样本的方向信息特征、层次信息特征，提升判别器的样本特征表达和特征识别能力，从而增强生成器学习真实数据整体分布的能力。请一并参阅图1，图1是增强型生成式对抗网络结构示意图。增强型生成式对抗网络包括增强型生成器和增强型判别器，增强型生成器将获得初始数据通过反池化层、线性修正、过滤(Filtering)层得到生成图片，并提供生成图片给增强型判别器，增强型判别器对图片进行处理后反馈分类结果给增强型生成器。增强型判别器包括：卷基层、基础胶囊层，卷积胶囊层和分类胶囊层，卷基层、基础胶囊层，卷积胶囊层和分类胶囊层依次相连。各层的层数的多少和结构参数根据目标样本的特征而定。
步骤200：构建多通道生成式对抗网络，利用训练过的多通道生成式对抗网络对无标注数据进行标签预测，根据优化后的生成器生成相应类别的样本，利用优化后的判别器对目标样本进行精准识别。
在步骤200中，首先，在增强型生成式对抗网络的基础上，针对目标样本的类别特点，设计多通道增强型生成式对抗网络，利用不同类别的样本分别训练各个通道的生成式对抗网络；其次，利用训练过的多通道生成式对抗网络对无标注数据进行标签预测，并利用优化后的生成器生成相应类别的样本，从而获得更充分的标注样本；最后，利用优化后的判别器，实现目标样本的精准识别。
请一并参阅图3，为了方便方案描述，本方案假设样本类别数目为N，针对N类数据集x＝{X 1，X 2，…X N}，数据集x整体分布服从分布x～φ；其中每一类数据子集可表示为
Figure PCTCN2018090761-appb-000009
并服从分布
Figure PCTCN2018090761-appb-000010
通过构建N个增强型生成式对抗网络组成多通道生成式对抗网络，使得每一个增强型生成器(G i，i＝1…N)只需要学习对应类别的真实样本分布
Figure PCTCN2018090761-appb-000011
步骤200具体包括：
步骤210：对标注样本进行数据增强；
在步骤220中，对于原始带标注的数据先进行按类别分类，分别对每一类的数据进行常见的数据增强操作，数据增强操作包括旋转、翻转、缩放等等，扩充后的数据集足够训练对应的多通道对抗生成网络。
步骤220：对增强型判别器网络进行训练；
在步骤220中，对于包含N类的样本，在模型训练过程中会逐一训练包含N个对抗生成网络。对于第i通道对抗生成网络只输入对应的i类的数据。针对第i通道对抗生成网络，对应的最优化的判别器训练目标函数为：
Figure PCTCN2018090761-appb-000012
Figure PCTCN2018090761-appb-000013
其中，Loss(G)代表着生成器G的损失函数(Loss Function)。它要测量的是真实数据(P data(x))与生成数据(P G(x))之间的差异。初始化判别器网络D的参数θ D和设定判别器训练迭代次数n D。训练判别器D时，在每一次的迭代中，从真实数据中采样出若干数据{x 1，x 2，x 3，…x m}。从噪声数据中采样出若干数据{z 1，z 2，z 3，…z m}(正态分布或者均匀分布)。然后通过生成器获得生成若干生成数据
Figure PCTCN2018090761-appb-000014
计算出若干惩罚项数据
Figure PCTCN2018090761-appb-000015
Figure PCTCN2018090761-appb-000016
∈是一个0～1之间的随机数；然后可以进行参数更新：
Figure PCTCN2018090761-appb-000017
经过以上n D次迭代训练就可以使得增强型判别器得到更好的性能。按照同样的方法，对剩下的N-1个增强型判别器进行训练，从而形成一组N个增强型判别器组。
步骤230：对增强型生成器进行网络训练；
在步骤230中，对于一组多通道对抗生成网络，也包含了N个增强型生成器。把训练数据按类别分成N类，并使用每一类的数据对相应的增强型生成器进行训练。最优化的生成器训练目标函数为：
Figure PCTCN2018090761-appb-000018
其中V(G，D)的定义与步骤二相同。对于寻找最优化生成器的过程，需要先初始化生成器网络G的参数θ G和设定生成器训练迭代次数n G。再从噪声数据中采样出若干数据{z 1，z 2，z 3，…z m}(正态分布或者均匀分布)；然后可以进行参数更新：
Figure PCTCN2018090761-appb-000019
然后再返回第一步进行新一轮的迭代，一共进行n G次的迭代更新，就可以使增强型型生成器达到最优的性能。按照同样的方法，对CSM其他类别的生成器进行训练，从而可以得到一组增强型生成器。
步骤240：输入噪声数据，通过增强型生成器生成新的标注数据；
在步骤240中，在得到一组训练好的多通道生成式对抗网络之后，取出所有的增强型生成器
Figure PCTCN2018090761-appb-000020
对于每一个增强型网络，分别输入一组噪声数据(高斯分布、均匀分布等)将会生成一组新的带标注数据，它的标注类别为该生成器所对应的类别：
Figure PCTCN2018090761-appb-000021
其中，
Figure PCTCN2018090761-appb-000022
代表新生成第i类数据，Z i代表第i类初始化噪声。以此类推， 对一个生成器进行相同的操作，都可以生成新的标注数据。这样便可以增加每类数据的数量，也扩充了数据集的大小。
步骤250：通过增强型判别器对无标注数据进行类别预测；
在步骤250中，在得到一组训练好的多通道生成式对抗网络之后，取出所有的生成器
Figure PCTCN2018090761-appb-000023
从原始未标注疾病数据集中任意取出一份数据
Figure PCTCN2018090761-appb-000024
分别输入到每一类的判别器当中。每一个判别器都会对此数据类别进行判断(y 1，y 2，…y N)，输出一个0-1之间的数字，数字越接近1意味着被判定为该类的置信度越高。理想状态下，所有的判别器中只会有一个输出值接近于1，其他的输出值都接近于0。这样就可以判定该输入数据的类别为输出值最大的那一类。如果存在多个接近1的输出值，说明生成器的训练并没有达到最优化的状态，需要继续返回训练。
Figure PCTCN2018090761-appb-000025
Figure PCTCN2018090761-appb-000026
以此类推，对原始数据集中的每一份未标注数据进行上述操作，会使未标注数据变成标注数据。从而增加了数据集中带标注的数据量，也扩充了数据集的大小。
步骤260：基于多通道生成式对抗网络中的判别器对目标样本进行分类。
在步骤260中，训练好的多通道生成式对抗网络中包含了N个增强型生成式对抗网络。把其中所有的增强型判别器取出来可以形成一组增强型判别器阵列(D s，s＝1…N)。这N个增强型判别器分别对应着疾病数据的N个类别。这些增强型判别器可以直接对测试样本进行分类预测。在测试阶段中，取任意测试样本(d k)分别输入到每一个增强型判别器()当中。每一个增强型判别器(D s)对于输入样本(d k)会对硬输出一个0～1之间的数字。该数字越大代表该判别器判定测试样本为所对应类别的置信度越高。正常情况下，一组N个增强型判别器对于一个测试样本得输出值中只有一个会接近1(例如D j)，其余的输出值都会接近0。根据以上分析可知，对于该测试样本(d k)的类别会被预测为某疾病 的第j类样本。按照以上方法，对于测试集中的每一个测试样本可以重复以上的操作即可对每个样本进行分类。
针对有限的标注样本，本发明实施例的增强型生成式对抗网络以及目标样本识别方法提出了融合胶囊机制的增强型生成式对抗网络，该网络利用向量化模式表达图像特征，显著增强了判别器对真实样本和生成样本的识别能力，进而提高了生成器学习真实数据整体分布的能力，使得生成对抗模型更容易达到纳什平衡。
在增强型生成式对抗网络的基础上，本发明实施例的增强型生成式对抗网络以及目标样本识别方法进一步提出了多通道生成对抗模型。针对N类数据集x＝{X 1，X 2，…X N}，数据集x整体分布服从分布x～φ；其中每一类数据子集可表示为
Figure PCTCN2018090761-appb-000027
并服从分布
Figure PCTCN2018090761-appb-000028
通过构建N个增强型生成式对抗网络组成多通道生成式对抗网络，使得每一个增强型生成器(G i，i＝1…N)只需要学习对应类别的真实样本分布
Figure PCTCN2018090761-appb-000029
相比于普通的生成器需要学习所有样本的整体分布(φ)而言，本项目的方法使得N个增强型生成器更容易达到纳什平衡状态，从而降低了生成式对抗网络难训练的问题，提升了生成式对抗网络的普适性。经过多次迭代训练，可以得到最优的判别器
Figure PCTCN2018090761-appb-000030
和生成器
Figure PCTCN2018090761-appb-000031
利用多通道最优化生成器可以按类别生成带标注高质量数据：
Figure PCTCN2018090761-appb-000032
其中，
Figure PCTCN2018090761-appb-000033
代表新生成第i类数据，Z i代表第i类初始化噪声。同时利用多通道最优化判别器可以对任意未标注数据
Figure PCTCN2018090761-appb-000034
进行标签预测：
Figure PCTCN2018090761-appb-000035
Figure PCTCN2018090761-appb-000036
本发明实施例的增强型生成式对抗网络以及目标样本识别方法可以有效解决少量标注样本条件下的模式识别问题，实现目标样本的精准识别。
本发明实施例的增强型生成式对抗网络以及目标样本识别方法使用多通道生成式对抗网络中的多个增强型生成器进行按类别地生成新的样本数据；同时， 也可以使用多通道生成式对抗网络中的多个增强型判别器组成一个增强型判别器阵列，并对未标注样本进行类别预测。从而增加了标注数据的数量，使得训练复杂的深度卷积神经网络分类器更加容易。
本发明实施例的增强型生成式对抗网络以及目标样本识别方法与现有技术相比，具有以下显著特点和积极效果：通过训练多通道生成式对抗网络使得多个增强型生成器分别学习到原始数据库中不同类别数据各自的子分布，这使得网络训练更容易达到纳什平衡，使得生成式对抗网络更容易训练。通过多通道生成式对抗网络中的多个增强型生成器和相应的增强型判别器阵列，可以大量地增加标注数据量，从而解决了在一些领域，譬如医学影像领域，大量高质量标注数据获取困难的问题，可以更加方便训练基于深度卷积神经网络的分类模型。
虽然本发明参照当前的较佳实施方式进行了描述，但本领域的技术人员应能理解，上述较佳实施方式仅用来说明本发明，并非用来限定本发明的保护范围，任何在本发明的精神和原则范围之内，所做的任何修饰、等效替换、改进等，均应包含在本发明的权利保护范围之内。

Claims (10)

  1. 一种增强型生成式对抗网络，其特征在于，包括至少一个增强型生成器和至少一个增强型判别器，所述增强型生成器将获得的初始数据进行处理得到生成数据，并提供生成数据给增强型判别器，所述增强型判别器对生成数据进行处理，并反馈分类结果给增强型生成器，所述增强型判别器包括卷基层、基础胶囊层、卷积胶囊层和分类胶囊层，所述卷基层、基础胶囊层、卷积胶囊层和分类胶囊层依次相连。
  2. 根据权利要求1所述的增强型生成式对抗网络，其特征在于，所述增强型生成器将获得的初始数据通过反池化层、线性修正、过滤层得到生成数据。
  3. 根据权利要求1或2所述的增强型生成式对抗网络，其特征在于，所述卷基层、基础胶囊层、卷积胶囊层和分类胶囊层的层数和结构参数根据目标样本的特征进行设定。
  4. 根据权利要求1所述的增强型生成式对抗网络，其特征在于，所述增强型生成器的数量大于1个，所述增强型判别器阵列数量大于1个，所述多个增强型生成器按类别生成新的样本数据，所述多个增强型判别器组成增强型判别器阵列，对未标注样本进行类别预测。
  5. 一种利用权利要求1至4任一项所述的增强型生成式对抗网络的目标样本识别方法，其特征在于，包括：
    步骤a：构建增强型生成式对抗网络，其中，构建的增强型生成式对抗网络包括至少一个增强型生成器和至少一个增强型判别器；
    步骤b：根据构建的增强型生成式对抗网络和目标样本的类别特点构建多通道生成式对抗网络，利用训练过的多通道生成式对抗网络对无标注数据进行标签预测，根据增强型生成器生成相应类别的样本，利用增强型判别器对目标样本进行精准识别。
  6. 根据权利要求5所述的目标样本识别方法，其特征在于，在所述步骤a中，所述增强型生成器将获得的初始数据进行处理得到生成数据，并提供生成数据给增强型判别器，所述增强型判别器对生成数据进行处理，并反馈分类结果给增强型生成器。
  7. 根据权利要求5或6所述的目标样本识别方法，其特征在于，在所述步 骤a中，所述增强型判别器包括：卷基层、基础胶囊层，卷积胶囊层和分类胶囊层，所述卷基层、基础胶囊层，卷积胶囊层和分类胶囊层依次相连。
  8. 根据权利要求5所述的目标样本识别方法，其特征在于，所述步骤a包括：利用胶囊特征向量化表达的模式，设计基于胶囊机制的增强型判别器；根据增强型生成对抗模型的纳什平衡能力，设计“生成-判别”交替优化方案；利用有标注样本和无标注样本，设计模型的目标函数。
  9. 根据权利要求5所述的目标样本识别方法，其特征在于，所述步骤b包括：对于原始带标注的数据先进行按类别分类，分别对每一类的数据进行数据增强操作；对增强型判别器网络进行训练；对增强型生成器进行网络训练；输入噪声数据，通过增强型生成器生成新的标注数据；通过增强型判别器对无标注数据进行类别预测；基于多通道生成式对抗网络中的判别器对目标样本进行分类。
  10. 根据权利要求9所述的目标样本识别方法，其特征在于，所述通过增强型判别器对无标注数据进行类别预测包括：从原始未标注疾病数据集中任意取出一份数据分别输入到每一类的判别器当中，每一个判别器都会对此数据类别进行判断输出一个0到1之间的数字，数字越接近1意味着被判定为该类的置信度越高；如果存在多个接近1的输出值，说明生成器的训练并没有达到最优化的状态，需要继续返回训练。
PCT/CN2018/090761 2018-06-12 2018-06-12 一种增强型生成式对抗网络以及目标样本识别方法 WO2019237240A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2018/090761 WO2019237240A1 (zh) 2018-06-12 2018-06-12 一种增强型生成式对抗网络以及目标样本识别方法
US16/999,118 US20200380366A1 (en) 2018-06-12 2020-08-21 Enhanced generative adversarial network and target sample recognition method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/090761 WO2019237240A1 (zh) 2018-06-12 2018-06-12 一种增强型生成式对抗网络以及目标样本识别方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/999,118 Continuation US20200380366A1 (en) 2018-06-12 2020-08-21 Enhanced generative adversarial network and target sample recognition method

Publications (1)

Publication Number Publication Date
WO2019237240A1 true WO2019237240A1 (zh) 2019-12-19

Family

ID=68841793

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/090761 WO2019237240A1 (zh) 2018-06-12 2018-06-12 一种增强型生成式对抗网络以及目标样本识别方法

Country Status (2)

Country Link
US (1) US20200380366A1 (zh)
WO (1) WO2019237240A1 (zh)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111767949A (zh) * 2020-06-28 2020-10-13 华南师范大学 一种基于特征和样本对抗共生的多任务学习方法及其系统
CN112001480A (zh) * 2020-08-11 2020-11-27 中国石油天然气集团有限公司 基于生成对抗网络的滑动定向数据小样本扩增方法
CN112488294A (zh) * 2020-11-20 2021-03-12 北京邮电大学 一种基于生成对抗网络的数据增强系统、方法和介质
CN112733963A (zh) * 2021-02-01 2021-04-30 中国人民解放军海军航空大学航空作战勤务学院 一种通用图像目标分类方法及系统
CN113654818A (zh) * 2021-07-21 2021-11-16 广州大学 基于胶囊网络的设备故障检测方法、系统、装置及介质
CN113721113A (zh) * 2021-09-02 2021-11-30 广西大学 一种基于半监督生成对抗网络的故障选线方法
CN113744175A (zh) * 2021-09-16 2021-12-03 中国人民解放军火箭军工程大学 一种基于双向约束生成对抗网络的图像生成方法及系统
CN114469120A (zh) * 2022-01-12 2022-05-13 大连海事大学 一种基于相似度阈值迁移的多尺度Dtw-BiLstm-Gan心电信号生成方法
CN114936622A (zh) * 2022-04-09 2022-08-23 西北工业大学 基于循环生成对抗网络的水声目标识别方法及装置
WO2022193495A1 (en) * 2021-03-15 2022-09-22 Huawei Cloud Computing Technologies Co., Ltd. Methods and systems for semantic augmentation of images
CN117409192A (zh) * 2023-12-14 2024-01-16 武汉大学 一种基于数据增强的红外小目标检测方法及装置

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11868437B1 (en) * 2019-09-30 2024-01-09 Sighthound, Inc. Training set enhancement for neural networks
CN111126503B (zh) * 2019-12-27 2023-09-26 北京同邦卓益科技有限公司 一种训练样本的生成方法和装置
JP7446903B2 (ja) * 2020-04-23 2024-03-11 株式会社日立製作所 画像処理装置、画像処理方法及び画像処理システム
CN112565261B (zh) * 2020-12-04 2021-11-23 浙江大学 基于多生成器AugGAN的对抗动态恶意API序列生成方法
CN112560795B (zh) * 2020-12-30 2022-07-26 南昌航空大学 一种基于cn-gan与cnn的sar图像目标识别算法
KR102541462B1 (ko) * 2021-01-28 2023-06-12 한국과학기술원 이형 코어 아키텍처 기반의 캡슐 신경망 추론 장치 및 그 방법
CN113037750B (zh) * 2021-03-09 2022-08-02 成都信息工程大学 一种车辆检测数据增强训练方法、系统、车辆及存储介质
CN113066049B (zh) * 2021-03-10 2023-04-07 湖南珞佳智能科技有限公司 Mems传感器疵病种类识别方法及系统
CN113326737A (zh) * 2021-05-06 2021-08-31 西北工业大学 一种水中目标的数据增强方法
CN113177599B (zh) * 2021-05-10 2023-11-21 南京信息工程大学 一种基于gan的强化样本生成方法
CN113326873B (zh) * 2021-05-19 2024-07-23 云南电网有限责任公司电力科学研究院 一种基于数据增强的电力设备分合闸状态自动分类方法
CN113591917B (zh) * 2021-06-29 2024-04-09 深圳市捷顺科技实业股份有限公司 一种数据增强的方法及装置
CN113420870B (zh) * 2021-07-04 2023-12-22 西北工业大学 用于水声目标识别的U-Net结构生成对抗网络及方法
CN114268981A (zh) * 2021-09-10 2022-04-01 南京星航通信技术有限公司 网络故障检测与诊断方法及系统
CN114499923B (zh) * 2021-11-30 2023-11-10 北京天融信网络安全技术有限公司 一种icmp模拟报文的生成方法及装置
CN114858782B (zh) * 2022-07-05 2022-09-27 中国民航大学 基于拉曼高光谱对抗判别模型的奶粉掺杂非定向检测方法
US11895344B1 (en) 2022-12-09 2024-02-06 International Business Machines Corporation Distribution of media content enhancement with generative adversarial network migration

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107180392A (zh) * 2017-05-18 2017-09-19 北京科技大学 一种电力企业电费回收数据模拟方法
CN107292336A (zh) * 2017-06-12 2017-10-24 西安电子科技大学 一种基于dcgan的极化sar图像分类方法
CN107944370A (zh) * 2017-11-17 2018-04-20 西安电子科技大学 基于dccgan模型的极化sar图像分类方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3685316B1 (en) * 2017-10-27 2023-08-09 Google LLC Capsule neural networks
US10592779B2 (en) * 2017-12-21 2020-03-17 International Business Machines Corporation Generative adversarial network medical image generation for training of a classifier
US20190303742A1 (en) * 2018-04-02 2019-10-03 Ca, Inc. Extension of the capsule network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107180392A (zh) * 2017-05-18 2017-09-19 北京科技大学 一种电力企业电费回收数据模拟方法
CN107292336A (zh) * 2017-06-12 2017-10-24 西安电子科技大学 一种基于dcgan的极化sar图像分类方法
CN107944370A (zh) * 2017-11-17 2018-04-20 西安电子科技大学 基于dccgan模型的极化sar图像分类方法

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
"RoboCup 2008: RoboCup 2008: Robot Soccer World Cup XII; (Lecture Notes in Computer Science; Lect.Notes Computer])", 23 January 2019, article AYUSH JAISWAL ET AL.: "CapsuleGan: Generative Adversarial Capsule Network", pages: 1 - 10, XP047501137, 558, DOI: 10.1007/978-3-030-11015-4_38 *
MATHIJS PIETERS ET AL.: "Comparing Generative ADversarial Network Techniques for Image Creation and Modification", ARXIV.ORG, 24 March 2018 (2018-03-24), pages 1 - 20, XP080862541, ISSN: 2331-8422 *
YASH UPADHYAY ET AL.: "Generative Adversarial Network Architectures For Image Synthesis Using Capsule networks", ARXIV, 11 June 2018 (2018-06-11), pages 1 - 9, XP080889162 *

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111767949A (zh) * 2020-06-28 2020-10-13 华南师范大学 一种基于特征和样本对抗共生的多任务学习方法及其系统
CN112001480A (zh) * 2020-08-11 2020-11-27 中国石油天然气集团有限公司 基于生成对抗网络的滑动定向数据小样本扩增方法
CN112001480B (zh) * 2020-08-11 2024-01-26 中国石油天然气集团有限公司 基于生成对抗网络的滑动定向数据小样本扩增方法
CN112488294A (zh) * 2020-11-20 2021-03-12 北京邮电大学 一种基于生成对抗网络的数据增强系统、方法和介质
CN112733963A (zh) * 2021-02-01 2021-04-30 中国人民解放军海军航空大学航空作战勤务学院 一种通用图像目标分类方法及系统
CN112733963B (zh) * 2021-02-01 2023-02-21 中国人民解放军海军航空大学航空作战勤务学院 一种通用图像目标分类方法及系统
WO2022193495A1 (en) * 2021-03-15 2022-09-22 Huawei Cloud Computing Technologies Co., Ltd. Methods and systems for semantic augmentation of images
US11593945B2 (en) 2021-03-15 2023-02-28 Huawei Cloud Computing Technologies Co., Ltd. Methods and systems for semantic augmentation of images
CN113654818A (zh) * 2021-07-21 2021-11-16 广州大学 基于胶囊网络的设备故障检测方法、系统、装置及介质
CN113654818B (zh) * 2021-07-21 2022-09-16 广州大学 基于胶囊网络的设备故障检测方法、系统、装置及介质
CN113721113B (zh) * 2021-09-02 2023-10-24 广西大学 一种基于半监督生成对抗网络的故障选线方法
CN113721113A (zh) * 2021-09-02 2021-11-30 广西大学 一种基于半监督生成对抗网络的故障选线方法
CN113744175A (zh) * 2021-09-16 2021-12-03 中国人民解放军火箭军工程大学 一种基于双向约束生成对抗网络的图像生成方法及系统
CN113744175B (zh) * 2021-09-16 2024-01-19 中国人民解放军火箭军工程大学 一种基于双向约束生成对抗网络的图像生成方法及系统
CN114469120A (zh) * 2022-01-12 2022-05-13 大连海事大学 一种基于相似度阈值迁移的多尺度Dtw-BiLstm-Gan心电信号生成方法
CN114936622A (zh) * 2022-04-09 2022-08-23 西北工业大学 基于循环生成对抗网络的水声目标识别方法及装置
CN114936622B (zh) * 2022-04-09 2024-02-27 西北工业大学 基于循环生成对抗网络的水声目标识别方法及装置
CN117409192A (zh) * 2023-12-14 2024-01-16 武汉大学 一种基于数据增强的红外小目标检测方法及装置
CN117409192B (zh) * 2023-12-14 2024-03-08 武汉大学 一种基于数据增强的红外小目标检测方法及装置

Also Published As

Publication number Publication date
US20200380366A1 (en) 2020-12-03

Similar Documents

Publication Publication Date Title
WO2019237240A1 (zh) 一种增强型生成式对抗网络以及目标样本识别方法
CN109063724B (zh) 一种增强型生成式对抗网络以及目标样本识别方法
Sun et al. Acne: Attentive context normalization for robust permutation-equivariant learning
Jia et al. A semisupervised Siamese network for hyperspectral image classification
CN110689086B (zh) 基于生成式对抗网络的半监督高分遥感图像场景分类方法
Liu et al. Relation-shape convolutional neural network for point cloud analysis
Yuan et al. Scene recognition by manifold regularized deep learning architecture
CN110443143B (zh) 多分支卷积神经网络融合的遥感图像场景分类方法
Xu et al. Learning deep structured multi-scale features using attention-gated crfs for contour prediction
CN104966104B (zh) 一种基于三维卷积神经网络的视频分类方法
Liu et al. Scene classification using hierarchical Wasserstein CNN
Lin et al. Active-learning-incorporated deep transfer learning for hyperspectral image classification
EP4002161A1 (en) Image retrieval method and apparatus, storage medium, and device
Ren et al. 3d-a-nets: 3d deep dense descriptor for volumetric shapes with adversarial networks
Wu et al. Feedback weight convolutional neural network for gait recognition
CN106296734B (zh) 基于极限学习机和boosting多核学习的目标跟踪方法
Abdul-Rashid et al. Shrec’18 track: 2d image-based 3d scene retrieval
Bai et al. Coordinate CNNs and LSTMs to categorize scene images with multi-views and multi-levels of abstraction
Ning et al. Conditional generative adversarial networks based on the principle of homologycontinuity for face aging
Huang et al. Flexible gait recognition based on flow regulation of local features between key frames
Ribeiro et al. Learning with capsules: A survey
Cai et al. Rgb-d scene classification via multi-modal feature learning
Yuan et al. Few-shot scene classification with multi-attention deepemd network in remote sensing
Xie et al. Learning cycle-consistent cooperative networks via alternating MCMC teaching for unsupervised cross-domain translation
Peng et al. Motion boundary emphasised optical flow method for human action recognition

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18922790

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18922790

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 18922790

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 07/09/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18922790

Country of ref document: EP

Kind code of ref document: A1