CN113610105A - Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning - Google Patents

Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning Download PDF

Info

Publication number
CN113610105A
CN113610105A CN202110743656.2A CN202110743656A CN113610105A CN 113610105 A CN113610105 A CN 113610105A CN 202110743656 A CN202110743656 A CN 202110743656A CN 113610105 A CN113610105 A CN 113610105A
Authority
CN
China
Prior art keywords
domain
calculating
alignment
loss
learning
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.)
Pending
Application number
CN202110743656.2A
Other languages
Chinese (zh)
Inventor
田青
杨宏
朱雅喃
薛晓妹
储奕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
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 Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202110743656.2A priority Critical patent/CN113610105A/en
Publication of CN113610105A publication Critical patent/CN113610105A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
    • 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

Abstract

The invention discloses an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning, and belongs to the field of domain adaptation in transfer learning. Comprising the steps of 1) sample weighting; step 2) constructing a dynamic balance factor: calculating the data distribution alignment degree of a source domain and a target domain by adopting Maximum Mean Difference (MMD), calculating the discriminability of the source domain and the target domain by adopting Linear Discriminant Analysis (LDA), carrying out normalization processing on the two estimated values, and calculating a balance factor tau; step 3) calculating the domain alignment loss: putting the domain alignment task in meta-training, calculating domain alignment loss, and updating network parameters; and 4) calculating the classification loss and updating the model parameters. According to the method, the optimization consistency between the domain alignment task and the classification task is promoted by weighting the samples, dynamically adjusting the weights of the domain alignment loss and the classification loss, and calculating the domain alignment loss and the classification loss through meta-learning to optimize the network model parameters.

Description

Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning
Technical Field
The invention relates to an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning, and belongs to the field of domain adaptation in transfer learning.
Background
The domain adaptation mainly solves the problem of having the same characteristics and categories, but the data characteristics are distributed in different source domains and target domains. The target domain task is solved by migrating the knowledge of the source domain into the target domain. Currently, domain adaptation has been applied and succeeded in many fields. For example, in the aspect of pedestrian re-identification, the traditional pedestrian re-identification utilizes a labeled pedestrian image data set as a training set, so as to realize the pedestrian re-identification problem in the real world. However, it is time-consuming and labor-consuming to collect the pedestrian pictures and manually give labels, so researchers introduce domain adaptation methods, use pedestrian pictures of other scenes as training sets (the distribution of the pedestrian pictures is similar to but different from that of the images of the tasks to be completed), and obtain good effects.
The purpose of unsupervised domain adaptation is to pass the knowledge learned in the labeled source domain exemplars to the unlabeled target domain exemplars. Unsupervised domain adaptation can utilize existing source domain tagged data and network models and associated target domain untagged data to learn to derive a network model suitable for target domain data classification. Conventional unsupervised domain adaptation methods typically use measures such as correlation distance metrics to align the data distribution of the source domain and the target domain output by the deep network. In recent years, many adversarial domain adaptation methods have been proposed and achieved remarkable results, and most of these methods are based on generation of an adversarial network. The method mainly comprises the steps of training a discriminator to discriminate whether a sampling feature is from a source domain or a target domain, and simultaneously training a feature extractor to deceive the discriminator, so that the feature distributions of the source domain and the target domain are aligned and cannot be distinguished. Most of these methods focus on measuring domain differences at the domain level, without distinguishing whether samples from two domains are aligned according to the category to which they belong. Even if the global statistics are completely mixed up, the difference between the source domain and the target domain is not necessarily reduced, and even different types of samples are mixed together, so that the classification effect is still to be improved.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides an unsupervised domain adaptation image classification method based on dynamic weighted learning and meta learning, so as to solve the problem of optimization inconsistency between a domain alignment task and a classification task in the existing domain adaptation method.
The technical scheme of the invention is as follows:
the invention provides an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning. Next, a balance factor τ between the domain alignment and the class discrimination is calculated to perform dynamic weight learning. Finally, the domain alignment is placed under a meta-training task, the object classification task is placed under a meta-testing task, and the consistency of the optimized network parameters of the object classification task is enhanced by using the strategy based on meta-optimization. The model has the advantages of strong generalization capability, high classification accuracy and the like.
The technical scheme is as follows: the invention relates to an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning, which comprises the following steps of:
step 1) sample weighting: weighting each sample of the source domain and the target domain, wherein the weight of the sample in each domain is inversely proportional to the proportion of the total sample amount of the samples in the two domains;
step 2) constructing a dynamic balance factor: calculating the data distribution alignment degree of a source domain and a target domain by adopting Maximum Mean Difference (MMD), calculating the discriminability of the source domain and the target domain by adopting Linear Discriminant Analysis (LDA), carrying out normalization processing on the two estimated values, and calculating a balance factor tau;
step 3) calculating the domain alignment loss: putting the domain alignment task in meta-training, calculating domain alignment loss, and updating network parameters;
step 4), calculating classification loss: and (4) putting the classification task in a meta test, calculating the classification loss, calculating the total loss of the model through meta optimization, and updating the model parameters.
The unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning is further designed in the step 1) of the source domain and target domain samplesThe method for weighting comprises the following steps: sample weighting is carried out according to the formula (1) to obtain weighted samples
Figure BDA0003143609960000021
And
Figure BDA0003143609960000022
the sample weighting algorithm is as follows:
Figure BDA0003143609960000023
wherein, alpha is (0, 1)]Is a hyper-parameter controlling the degree of sample weighting, nsAnd ntRespectively representing the number of samples of the source domain and the target domain,
Figure BDA0003143609960000024
and
Figure BDA0003143609960000025
respectively representing samples of the source domain and the target domain.
The unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning is further designed in that the step 2) of constructing the dynamic balance factor specifically comprises the following steps:
step 2-1) calculating the data distribution alignment degree between the source domain and the target domain according to the formula (2):
Figure BDA0003143609960000031
wherein the content of the first and second substances,
Figure BDA0003143609960000032
respectively, extracting features of the source and target domains, smaller MMD (D)s,Dt) Meaning a better degree of alignment.
Step 2-2) calculating identifiability of the source domain and the target domain according to the formula (3):
Figure BDA0003143609960000033
wherein S isbIs an inter-class scattering matrix, SwIs an intra-class scattering matrix, with larger j (w) meaning better discriminability.
Step 2-3) carrying out normalization processing on the formula (2) and the formula (3) according to the formula (4):
Figure BDA0003143609960000034
where the values calculated by equations (2) and (3) are usually not an order of magnitude, we vary them linearly and map the results into a [0,1] range in order to reasonably normalize them.
Step 2-4) calculating a dynamic balance factor tau according to the formula (5):
Figure BDA0003143609960000035
of which the smaller one
Figure BDA0003143609960000036
Indicating better domain alignment, smaller
Figure BDA0003143609960000037
Indicating better discriminability. When the degree of alignment is far superior to discriminability,
Figure BDA0003143609960000038
close to the value of 0 (c) and,
Figure BDA0003143609960000039
close to 1 and τ close to 0. When the degree of alignment is much less than discernable,
Figure BDA00031436099600000310
close to the position of the light source at 1,
Figure BDA00031436099600000311
close to 0 and τ close to 1.
The unsupervised domain-adapted image classification method based on dynamic weighted learning and meta learning is further designed in that the domain alignment loss is calculated in the step 3), and the network parameter theta is updated according to the formula (6):
Figure BDA00031436099600000312
wherein, thetamRepresents mthE { 1.., M } set of parameters.
The unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning is further designed in that classification loss is calculated in the step 4), model total loss is calculated through meta-optimization, and model parameters are updated.
Advantageous effects
The invention provides an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning, which is a brand-new MetaAlign based on dynamic weighted learning and meta-optimization strategies. First, by weighting the samples of the two domains, the model bias caused by the imbalance of the sample sizes of the two domains is prevented. Second, a balance factor τ between domain alignment and class identification is calculated for dynamic weight learning. Finally, the domain alignment is placed under a meta-training task, the object classification task is placed under a meta-testing task, and the consistency of the optimized network parameters of the object classification task is enhanced by using the strategy based on meta-optimization. By the method, the two types of learning, namely domain alignment learning and classification learning, can be in an effective learning state, and the two types of learning can be simultaneously propelled to a good direction (namely, the domain distance is small, and the discriminability is large);
the method comprises the steps of respectively calculating domain alignment loss and classification loss under meta-training and meta-testing by domain alignment learning and classification learning in unsupervised domain adaptation, updating a network parameter theta through the calculated domain alignment loss, and using the theta for calculating the classification loss so as to strengthen the consistency of optimized network parameters. The generalization capability of the model is stronger, and the classification accuracy is higher.
Drawings
FIG. 1 is a frame diagram of an unsupervised domain adapted image classification method based on dynamic weighted learning and meta-learning;
FIG. 2 is a model diagram of a network of unsupervised domain adapted image classification methods based on dynamic weighted learning;
FIG. 3 is a network model diagram of classification tasks in a network of unsupervised domain adaptive image classification methods based on meta-learning.
Detailed Description
In order to make researchers in this field understand the technical problems and technical solutions in the present application better and achieve the technical effects that can be achieved by the application, a completely new unsupervised domain-adapted image classification method framework based on dynamic weighted learning and meta learning is further described in detail below with reference to fig. 1-3 and the detailed description.
The unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning provided by the invention comprises the following steps, and the flow is shown in figure 1:
step 1: the source domain and target domain samples are weighted.
In the unsupervised domain adaptation problem, when the sample size of one domain is very different from that of another domain, in particular, a domain with a large sample size is equivalent to having a large weight in the training process. This imbalance easily leads to model bias during training and to poor domain alignment and classification, resulting in negative migration.
In order to avoid the problem of unbalanced sample size of the two domains, the invention performs intuitive weighting on the samples of the source domain and the target domain. The weight of each domain sample is inversely proportional to their proportion in the total sample size of the two domains. Specifically, sample weighting is performed according to equation (1) to obtain weighted samples
Figure BDA0003143609960000051
And
Figure BDA0003143609960000052
Figure BDA0003143609960000053
wherein, alpha is (0, 1)]Is a hyper-parameter controlling the degree of sample weighting, nsAnd ntRepresenting the number of samples of the source domain and the target domain respectively,
Figure BDA0003143609960000054
and
Figure BDA0003143609960000055
samples representing the source domain and the target domain, respectively.
Step 2: a dynamic balance factor is constructed.
In order to avoid excessive domain alignment or independent discriminability in the unsupervised domain adaptive image classification task and enable the domain alignment and the independent discriminability to simultaneously develop towards a good direction, the invention provides a method for measuring the domain alignment degree and the discriminability of each iteration in real time and then constructing a dynamic balance factor tau to control the weights of domain alignment loss and class discriminability loss.
Maximum Mean Difference (MMD) and Linear Difference Analysis (LDA) are used to measure the degree of domain alignment and class discriminability, respectively, of the current cross-domain feature representation.
And (3) calculating the alignment degree of the data distribution between the source domain and the target domain according to the formula (2).
Figure BDA0003143609960000056
Wherein the content of the first and second substances,
Figure BDA0003143609960000057
respectively, extracting features of the source and target domains, smaller MMD (D)s,Dt) Meaning a better degree of domain alignment.
The identifiability of the source and target domains is calculated according to equation (3):
Figure BDA0003143609960000058
wherein S isbIs an inter-class scattering matrix, SwIs an intra-class scattering matrix, with larger j (w) meaning better discriminability.
And (4) carrying out normalization processing on the formula (2) and the formula (3).
Figure BDA0003143609960000059
Where the values calculated by equations (2) and (3) are usually not an order of magnitude, we vary them linearly and map the results into a [0,1] range in order to reasonably normalize them.
The dynamic balance factor τ is calculated according to equation (5):
Figure BDA0003143609960000061
of which the smaller one
Figure BDA0003143609960000062
Indicating better domain alignment, smaller
Figure BDA0003143609960000063
Indicating better discriminability. When the degree of alignment is far superior to discriminability,
Figure BDA0003143609960000064
close to the value of 0 (c) and,
Figure BDA0003143609960000065
close to 1 and τ close to 0. When the degree of alignment is much less than discernable,
Figure BDA0003143609960000066
close to the position of the light source at 1,
Figure BDA0003143609960000067
close to 0 and τ close to 1.
And step 3: a domain alignment penalty is calculated.
Antagonistic learning has successfully introduced domain alignment by learning domain-invariant feature representations. The invention inputs weighted samples for obtaining a domain-invariant feature representation
Figure BDA0003143609960000068
And
Figure BDA0003143609960000069
the features are extracted by a feature extractor G. Training the parameter θ of the feature extractor G by optimizing the domain alignment loss of equation (6)gParameter θ of sum-field discriminator Dd. Putting the domain alignment task into meta-training, calculating the domain alignment loss, and updating the parameter theta of the feature extractor G in the network by the formula (7):
Figure BDA00031436099600000610
Figure BDA00031436099600000611
wherein, thetamRepresents mthE { 1.., M } set of parameters. And passing the parameter theta of the feature extractor G obtained by the domain alignment network in the meta-training to the classification network in the meta-test through the meta-learning. Antagonistic learning can effectively achieve domain alignment.
And 4, step 4: the classification loss is calculated.
By making the classifier C1,C2To maximize the difference to obtain strong discriminative features. Optimizing class discrimination loss through formula (8), training parameter theta of feature extractor G, classifiers C and C1,C2Parameter theta ofc,θc1,θc2。
Figure BDA00031436099600000612
Referring to fig. 2, the unlabeled target domain image and the labeled source domain image are input into the domain alignment network in the meta training after sample weighting processing. The domain alignment network consists of a feature extractor and a discriminator. The feature extractor G is used for extracting the features of the source domain and the target domain samples, and the discriminator D is used for discriminating the source domain image and the target domain image. The loss resulting from the domain aligned network in meta-training is shown in equation (9) below:
Figure BDA0003143609960000071
in addition, the unlabeled target domain image and the labeled source domain image are subjected to sample weighting processing and then input into the network in the meta-training. The classification network consists of one feature extractor and three classifiers, as shown in fig. 3. The feature extractor G is used to extract features of the source domain and target domain samples, which are then input into three classifiers trained from source domain data. By maximizing C1,C2The difference between them, to obtain strong discriminability. The loss resulting from the classification network in meta-training is shown in equation (10) below:
Figure BDA0003143609960000072
therefore, the total model loss of the unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning is defined as follows:
Figure BDA0003143609960000073
wherein, thetamRepresents mthE { 1.., M } set of parameters. To avoid loss of computational complexity, we set β ═ βm}mAdd 1 norm optimization:
Figure BDA0003143609960000074
b is a super parameter.
The present invention provides an unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning, and a plurality of methods and approaches for implementing the technical scheme, and the above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. All the components not specified in the present embodiment can be realized by the prior art.

Claims (5)

1. An unsupervised domain adaptive image classification method based on dynamic weighted learning and meta learning is characterized by comprising the following steps of:
step 1) sample weighting: weighting each sample of the source domain and the target domain, the weight of the sample being inversely proportional to the proportion of their total sample size in the two domains;
step 2) constructing a dynamic balance factor: calculating the data distribution alignment degree of a source domain and a target domain by adopting maximum average difference MMD, calculating the discriminability of the source domain and the target domain by adopting linear discriminant analysis LDA, carrying out normalization processing on the two estimated values, and calculating a balance factor tau;
step 3) calculating the domain alignment loss: putting the domain alignment task in meta-training, calculating domain alignment loss, and updating network parameters;
step 4), calculating classification loss: and (4) putting the classification task in a meta test, calculating the classification loss, calculating the total loss of the model through meta optimization, and updating the model parameters.
2. The unsupervised domain-adapted image classification method according to claim 1, wherein in step 1), the weighting process is performed on each sample of the source domain and the target domain as follows: sample weighting is carried out according to the formula (1) to obtain weighted samples
Figure FDA0003143609950000011
And
Figure FDA0003143609950000012
the sample weighting algorithm is as follows:
Figure FDA0003143609950000013
wherein, alpha is (0, 1)]Is a hyper-parameter controlling the degree of sample weighting, nsAnd ntRepresenting the number of samples of the source domain and the target domain respectively,
Figure FDA0003143609950000014
and
Figure FDA0003143609950000015
samples representing the source domain and the target domain, respectively.
3. The unsupervised domain-adapted image classification method according to claim 2, wherein in step 2), the constructing of the dynamic balance factor specifically comprises the steps of:
step 2-1) calculating the data distribution alignment degree between the source domain and the target domain according to the formula (2):
Figure FDA0003143609950000016
wherein the content of the first and second substances,
Figure FDA0003143609950000017
respectively extracting the features of the source domain and the target domain, and the small maximum mean difference MMD (D)s,Dt) Meaning a better degree of alignment;
step 2-2) calculating the discriminativity of the source domain and the target domain according to the formula (3):
Figure FDA0003143609950000018
wherein S isbIs an inter-class scattering matrix, SwIs an intra-class scattering matrix, the value of J (w) is that the inter-class scattering matrix is larger than the above intra-class scattering matrix, the larger the value of J (w) is, the larger the distance between each two classes is compared with the intra-class distance of each class, i.e. the larger identifiability is provided for each class, so the larger J (w) means the better discriminability.
Step 2-3) carrying out normalization processing on the formula (2) and the formula (3) according to the formula (4):
Figure FDA0003143609950000021
wherein the values calculated by the equations (2) and (3) are usually not in an order of magnitude, and in order to reasonably normalize the two values, we change them linearly and map the result into the range of [0,1 ];
step 2-4) calculating a dynamic balance factor tau according to the formula (5):
Figure FDA0003143609950000022
of which the smaller one
Figure FDA0003143609950000023
Indicating better domain alignment, smaller
Figure FDA0003143609950000024
Indicating a better discriminability; when the degree of alignment is far better than the discriminability,
Figure FDA0003143609950000025
close to the value of 0 (c) and,
Figure FDA0003143609950000026
close to 1, τ is close to 0; when the degree of alignment is much lower than the discriminability,
Figure FDA0003143609950000027
close to the position of the light source at 1,
Figure FDA0003143609950000028
close to 0 and τ close to 1.
4. The unsupervised domain-adapted image classification method according to claim 3, wherein in step 3), the domain alignment loss is calculated, and the network parameter θ is updated according to equation (6):
Figure FDA0003143609950000029
wherein, thetamRepresents mthE { 1.., M } set of parameters.
5. The unsupervised domain-adapted image classification method according to claim 4, wherein in step 4), the classification loss is calculated, and model parameters are updated by calculating model total loss through meta-optimization.
CN202110743656.2A 2021-07-01 2021-07-01 Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning Pending CN113610105A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110743656.2A CN113610105A (en) 2021-07-01 2021-07-01 Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110743656.2A CN113610105A (en) 2021-07-01 2021-07-01 Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning

Publications (1)

Publication Number Publication Date
CN113610105A true CN113610105A (en) 2021-11-05

Family

ID=78337112

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110743656.2A Pending CN113610105A (en) 2021-07-01 2021-07-01 Unsupervised domain adaptive image classification method based on dynamic weighted learning and meta-learning

Country Status (1)

Country Link
CN (1) CN113610105A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114500325A (en) * 2022-01-27 2022-05-13 重庆邮电大学 SDN controller fault self-adaptive intelligent detection method based on unsupervised transfer learning

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180165554A1 (en) * 2016-12-09 2018-06-14 The Research Foundation For The State University Of New York Semisupervised autoencoder for sentiment analysis
CN110348579A (en) * 2019-05-28 2019-10-18 北京理工大学 A kind of domain-adaptive migration feature method and system
CN111275092A (en) * 2020-01-17 2020-06-12 电子科技大学 Image classification method based on unsupervised domain adaptation
CN111695209A (en) * 2020-05-13 2020-09-22 东南大学 Rotary machine small sample health assessment method driven by meta-depth learning
CN112308158A (en) * 2020-11-05 2021-02-02 电子科技大学 Multi-source field self-adaptive model and method based on partial feature alignment
CN112396097A (en) * 2020-11-09 2021-02-23 中山大学 Unsupervised domain self-adaptive visual target detection method and system based on weighted optimal transmission and storage medium
CN112541458A (en) * 2020-12-21 2021-03-23 中国科学院自动化研究所 Domain-adaptive face recognition method, system and device based on meta-learning

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180165554A1 (en) * 2016-12-09 2018-06-14 The Research Foundation For The State University Of New York Semisupervised autoencoder for sentiment analysis
CN110348579A (en) * 2019-05-28 2019-10-18 北京理工大学 A kind of domain-adaptive migration feature method and system
CN111275092A (en) * 2020-01-17 2020-06-12 电子科技大学 Image classification method based on unsupervised domain adaptation
CN111695209A (en) * 2020-05-13 2020-09-22 东南大学 Rotary machine small sample health assessment method driven by meta-depth learning
CN112308158A (en) * 2020-11-05 2021-02-02 电子科技大学 Multi-source field self-adaptive model and method based on partial feature alignment
CN112396097A (en) * 2020-11-09 2021-02-23 中山大学 Unsupervised domain self-adaptive visual target detection method and system based on weighted optimal transmission and storage medium
CN112541458A (en) * 2020-12-21 2021-03-23 中国科学院自动化研究所 Domain-adaptive face recognition method, system and device based on meta-learning

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
NI XIAO等: "Dynamic Weighted Learning for Unsupervised Domain Adaptation", 《PROCEEDINGS OF THE IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》, pages 15242 - 15251 *
QIANRU SUN等: "Meta-Transfer Learning for Few-Shot Learning", 《PROCEEDINGS OF THE IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》, pages 403 - 412 *
曹猛: "基于关系挖掘的人脸年龄估计研究", 《中国优秀硕士学位论文全文数据库：信息科技辑》, no. 2, pages 138 - 851 *
陈迪: "基于对抗和正则化方法的域适应算法研究", 《中国优秀硕士学位论文全文数据库：信息科技辑》, no. 7, pages 138 - 527 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114500325A (en) * 2022-01-27 2022-05-13 重庆邮电大学 SDN controller fault self-adaptive intelligent detection method based on unsupervised transfer learning
CN114500325B (en) * 2022-01-27 2023-07-18 重庆邮电大学 SDN controller fault self-adaptive intelligent detection method based on unsupervised transfer learning

Similar Documents

Publication Publication Date Title
CN108388927B (en) Small sample polarization SAR terrain classification method based on deep convolution twin network
CN104992184B (en) A kind of multiclass image classification method based on semi-supervised extreme learning machine
CN110163033B (en) Positive sample acquisition method, pedestrian detection model generation method and pedestrian detection method
CN101814149B (en) Self-adaptive cascade classifier training method based on online learning
CN110263697A (en) Pedestrian based on unsupervised learning recognition methods, device and medium again
CN111553193A (en) Visual SLAM closed-loop detection method based on lightweight deep neural network
CN108171184A (en) Method for distinguishing is known based on Siamese networks again for pedestrian
Wu et al. End-to-end chromosome Karyotyping with data augmentation using GAN
CN105718866B (en) A kind of detection of sensation target and recognition methods
CN104598924A (en) Target matching detection method
CN110807760B (en) Tobacco leaf grading method and system
CN104463196A (en) Video-based weather phenomenon recognition method
CN108985360A (en) Hyperspectral classification method based on expanding morphology and Active Learning
CN107392241A (en) A kind of image object sorting technique that sampling XGBoost is arranged based on weighting
CN109344856B (en) Offline signature identification method based on multilayer discriminant feature learning
CN106650667A (en) Pedestrian detection method and system based on support vector machine
CN111368660A (en) Single-stage semi-supervised image human body target detection method
CN113177612B (en) Agricultural pest image identification method based on CNN few samples
CN103839078A (en) Hyperspectral image classifying method based on active learning
CN109977887A (en) A kind of face identification method of anti-age interference
CN105512675B (en) A kind of feature selection approach based on the search of Memorability multiple point crossover gravitation
CN106874929A (en) A kind of pearl sorting technique based on deep learning
CN109214444B (en) Game anti-addiction determination system and method based on twin neural network and GMM
CN113489685A (en) Secondary feature extraction and malicious attack identification method based on kernel principal component analysis
CN110414587A (en) Depth convolutional neural networks training method and system based on progressive learning

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