CN116912593B - Domain countermeasure remote sensing image target classification method - Google Patents
Domain countermeasure remote sensing image target classification method Download PDFInfo
- Publication number
- CN116912593B CN116912593B CN202310950762.7A CN202310950762A CN116912593B CN 116912593 B CN116912593 B CN 116912593B CN 202310950762 A CN202310950762 A CN 202310950762A CN 116912593 B CN116912593 B CN 116912593B
- Authority
- CN
- China
- Prior art keywords
- domain
- encoder
- classifier
- class
- training
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 18
- 230000003042 antagnostic effect Effects 0.000 claims description 4
- 230000008485 antagonism Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 9
- 230000006870 function Effects 0.000 description 5
- 230000001364 causal effect Effects 0.000 description 3
- WDLRUFUQRNWCPK-UHFFFAOYSA-N Tetraxetan Chemical compound OC(=O)CN1CCN(CC(O)=O)CCN(CC(O)=O)CCN(CC(O)=O)CC1 WDLRUFUQRNWCPK-UHFFFAOYSA-N 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- 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
- G06N3/09—Supervised learning
-
- 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
- G06N3/094—Adversarial learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to the technical field of computer vision image processing, and discloses a domain countermeasure remote sensing image target classification method. Training a domain encoder and a domain classifier through multiple remote sensing image datasets and training losses; inputting a certain domain sample to a class encoder, and inputting another domain sample of the same class to the domain encoder; fixed domain encoders, domain classifiers, and generators training class encoders and classifiers by classifying loss, counterloss, and data enhancement loss. Training a domain encoder and a domain classifier well through the first step of training, and providing an auxiliary training network for the second step of training; the second step of training is to fight against the class encoder through the domain classifier trained in the first step, so that the domain generalization effect of the class encoder is improved, the domain encoder fully decouples the characteristics of the domain features through the domain encoder trained in the first step, and the class encoder fully decouples the coupled class features through a data enhancement training mode, so that the generalization capability is further improved.
Description
Technical Field
The invention relates to the technical field of computer vision image processing, in particular to a domain countermeasure remote sensing image target classification method.
Background
The invention uses two related methods: firstly, a target domain generalization algorithm based on countermeasure and secondly, a data enhancement method based on image generation.
The generalization problem of the target classification domain is to train on one or more source domains, and in the case that the target domain data set cannot be acquired, the target domain is verified to have good performance by a certain algorithm. The invention belongs to the problem of multisource domain generalization.
The challenge-based target domain generalization algorithm is a target domain generalization manner, and can be applied to various tasks such as a target classification task, a target detection task, a target segmentation task, and the like. The main method is to use a feature extractor for learning category features and a domain classifier of which domain the learning sample belongs to conduct countermeasure training, so that the domain classifier cannot judge which domain the features output by the feature extractor belong to, and the feature extractor needs to be trained to output domain independent features as far as possible, thereby realizing domain generalization. For example, ganin et al, in the article "Domain-adversarial training of neural networks," learn a common feature space through the countermeasure learning of source domains, spoof Domain discriminators by generating Domain-invariant representations; li et al in Domain generalization with adversarial feature learning employ Maximum Mean Dispersion (MMD) to align domains, matching potential representations to a priori distributions by resistance learning; in addition, long et al Conditional adversarial domain adaptation also consider the primary task labels when distinguishing samples from different domains.
Data enhancement is an effective way to increase data diversity so it can increase the domain generalization ability of the model and robustness to spurious correlations. Zhang et al uses a generator to enhance data to assist in generating style conversion images in the Exact Feature Distribution Matching for Arbitrary Style Transfer and Domain Generalization article, thereby achieving domain generalization; chalupka et al in Visual causal feature learning designed an active learning scheme to learn causal operations on images, which enriches the data set of observed data and improves generalization of causal and predictive learning tasks; volpi et al in Generalizing to unseen domains via adversarial data augmentation achieve domain generalization by antagonistic data enhancement.
The target domain generalization algorithm based on the countermeasure is effective, however, the effect of the countermeasure training may not be enough to train a complete domain classifier, so that the features extracted by the feature extractor cannot be completely domain independent; moreover, the constraint of just countering the loss is not sufficient, and overfitting may result. Therefore, we propose a way to combine antagonism and data enhancement to improve the domain generalization effect of remote sensing target classification. The method introduces a domain encoder to extract domain features of the image, and combines the domain encoder and the domain classifier for training, so that the domain classifier and the domain encoder with sufficient training are obtained; in addition, the combination of the domain encoder and the category encoder realizes the effect of decoupling the category information and the domain information of the images, and the generator combines the categories and the domains of different images to achieve the effect of data enhancement, thereby increasing the data enhancement constraint and preventing the model from being over fitted.
Disclosure of Invention
Aiming at a remote sensing image, a target classification method of a domain countermeasure remote sensing image is provided, the class encoder is trained through the class encoder and the domain classifier countermeasure, and then the class encoder is fully decoupled from the domain encoder to realize data enhancement so as to restrict the class encoder, so that the class encoder can output domain independent characteristic representation, and the domain generalization effect is realized.
The technical scheme of the invention is as follows: a domain countermeasure remote sensing image target classification method comprises the following steps:
training a domain encoder and a domain classifier through multiple remote sensing image data sets;
the domain encoder and the domain classifier both adopt a ResNet50 structure; the domain encoder is used for outputting domain characteristic information of the remote sensing image, and has the function of decoupling the domain information of the remote sensing image and is used for assisting the class encoder to sufficiently decouple the class information irrelevant to the domain in the second step in a data enhancement mode; the domain classifier is used for outputting the domain to which the remote sensing image sample belongs and is used for antagonizing the class encoder in the second step, and training the class encoder to output the domain independent characteristics so that the domain classifier can not judge the domain information of the remote sensing image;
p remote sensing image data sets are adopted as p domains respectively; inputting the remote sensing image dataset into a domain encoder, inputting the output characteristics of the domain encoder into a domain classifier, and training to obtain the domain encoder only extracting the domain characteristics and the domain classifier outputting the domain information of the remote sensing image;
training loss is as follows:
where N is the total number of samples of the remote sensing image dataset, x i Is an input image, d i Is the domain of the remote sensing image, f d Is a domain encoder, f dc Is a domain classifier;
training a class encoder and a class classifier according to the domain encoder and the domain classifier obtained by training in the step one and the existing generator;
the class encoder and the classifier both adopt a ResNet50 model structure, and two full-connection layers with the length of 512 are newly added in the classifier; the penalty of training the class encoder and classifier includes classification penalty, antagonism penalty, and data enhancement penalty; the classification loss is used for training the correct class of the remote sensing image output by the class encoder and the classifier, and the counterloss and the data enhancement loss are used for improving the generalization capability of the class encoder;
input m-domain samplesGenerating category characteristics by a category encoder; input n-field samples of the same class->Generating domain features into a domain encoder; the method comprises the steps of fixing a domain encoder, a domain classifier and a generator, inputting class features into the classifier, the domain classifier and the generator respectively, inputting the domain features into the generator, and training the class encoder and the classifier;
the classification loss in the classifier is cross entropy loss, and the formula is as follows:
wherein n is k Is the number of samples of each remote sensing image dataset, f s Is a class encoder, f c Is a classifier gt i The true value is corresponding to the remote sensing image sample; x is x i Is a category feature;
fixing the domain classifier obtained in the first step, which is opposed to the class encoder; the domain classifier outputs the domain to which the remote sensing image belongs according to the domain information represented by the features, and trains the class encoder so that the features output by the class encoder do not contain the domain information; the countermeasures loss formula in the domain classifier is as follows:
where N is the total number of samples of the remote sensing image dataset, x i Is a category feature, d i Is the domain of the image, f d Is a domain encoder, f dc Is a domain classifier;
fixing the domain encoder obtained in the first step, and introducing a generator, wherein the generator generates an image according to the category characteristics and the domain characteristics; selecting two images of the same category and different fieldsAnd->Respectively input into a category encoder and a domain encoder to obtain category characteristics x i And domain feature n; category characteristics x i And domain feature n generating an image by a generator +.>The generated image and the image inputted to the domain encoder +.>As a loss; the less domain information the feature generated by the class encoder contains, the less distance the generated image is from the image input to the domain encoder; the data enhancement loss is as follows:
wherein,and->Images of the same category and different domains, G is a generator, f s Is a class encoder, f d Is a domain encoder, l () is a distance function;
in summary, the total loss of the training in the second step is:
L＝L 1 +L 2 +L 3 (5)。
the invention has the beneficial effects that: the invention provides a domain countermeasure remote sensing image target classification method, which realizes the improvement of domain generalization effect through two-step training of a remote sensing image classification network. Training a domain encoder and a domain classifier well through the first step of training, and providing an auxiliary training network for the second step of training; the second step of training is to fight against the class encoder through the domain classifier trained in the first step, so that the domain generalization effect of the class encoder is improved, the domain encoder fully decouples the characteristics of the domain features through the domain encoder trained in the first step, and the class encoder fully decouples the coupled class features through a data enhancement training mode, so that the generalization capability is further improved.
Drawings
Fig. 1 is a network overall structure diagram in a domain countermeasure remote sensing image target classification method.
Fig. 2 is a flowchart of the overall training of the domain countermeasure remote sensing image target classification method.
Detailed Description
Specific embodiments of the present invention will be further described with reference to the accompanying drawings and the technical scheme.
A domain countermeasure remote sensing image target classification method comprises the following steps:
the training step is divided into two steps as shown in fig. 1; the first step is to train the domain encoder and domain classifier with multiple remote sensing image datasets and the second step is to train the class encoder and classifier with the trained domain encoder and domain classifier and the existing generator.
(1) Training a domain encoder and a domain classifier:
the domain encoder and domain classifier employ a ResNet50 structure. The domain encoder is used for outputting domain characteristic information of the image, and the domain encoder is used for decoupling the domain information of the image, so that the class encoder is assisted in a data enhancement mode in a second training stage to fully decouple the class information irrelevant to the domain. The domain classifier is used for outputting which domain the remote sensing image sample belongs to, and has the function of antagonizing with the class encoder in the second training stage, and training the class encoder to output domain independent characteristics so that the domain classifier cannot judge the domain information of the image.
Three remote sensing image datasets DIOR, HRRSD and DOTA were used as three fields, respectively. The domain information is input into a domain encoder and a domain classifier, and the domain encoder which only extracts domain features and the domain classifier which can output domain information of an image are trained.
Training loss is as follows:
where N is the total number of samples of the three remote sensing image datasets, x i Is an input image, d i Is the domain of the remote sensing image, f d Is a domain encoder, f dc Is a domain classifier.
(2) Training a class encoder and a classifier:
the class encoder and the classifier also adopt a ResNet50 model structure, and in addition, two full-connection layers with the length of 512 are added in the classifier for better classification effect. The loss of the training class encoder and the classifier is three, one is the classification loss that the training class encoder and the classifier can output the correct class, and the other two is the countermeasure loss and the data enhancement loss for improving the generalization capability of the class encoder domain.
First, in order for the class encoder and classifier to output the class correctly, training with cross entropy loss is required, as follows:
wherein n is k Is the number of samples of each remote sensing image dataset, f s Is a class encoder, f c Is a classifier gt i Is the true value corresponding to the remote sensing image sample, x i Is a category feature.
Second, in order to generalize the class encoder SE, the class encoder is countered by the domain classifier trained in the first step. The domain classifier is fixed in training, and the domain classifier can output the domain to which the image belongs according to the domain information represented by the features, so that the feature of the domain classifier can be used for training the class encoder, the features output by the class encoder do not contain the domain information, and the domain classifier cannot judge which domain the image belongs to. The loss formula is as follows:
where N is a sample of three remote sensing image datasetsThe sum, x i Is a category feature, d i Is the domain of the image, f d Is a domain encoder, f dc Is a domain classifier.
Next, in order to further enhance the domain generalization capability of the class encoder, the domain encoder is used as an aid to decouple the domain features and class features of the image, so that the image input to the class encoder only outputs the class features, does not include the domain information, and the image input to the domain encoder only outputs the domain features, does not include the class information. Since the domain encoder has been trained in the first step of training, the domain encoder is fixed. To achieve data enhancement, a generator patchGAN is introduced, which is already trained and is capable of generating images from class features and domain features. Selecting two images of the same category and different fieldsAnd->Respectively input into a category encoder and a domain encoder to obtain category characteristics x i And domain feature n, and generating an image via a generator>The distance between the generated image and the image input to the domain encoder is taken as a loss. Since the two image categories are identical, the fewer domain information the characteristics generated by the category encoder contain, the smaller the distance between the generated image and the image input to the domain encoder. The loss function is as follows:
wherein the method comprises the steps ofAnd->Images of different domains of the same categoryG is a generator, f s Is a class encoder, f d Is a domain encoder. l () is a distance function between two samples.
The network training flow chart is shown in fig. 2. Firstly, inputting a DIOR data set, an HRRSD data set and a DOTA data set into a domain encoder and a domain classifier, and training a domain encoder f capable of extracting sample domain information characteristics d Domain classifier f capable of correctly classifying which domain the sample belongs to dc . Second fixed-domain encoder f d Sum domain classifier f dc Training class encoder f s And classifier f c . There are three losses as constraints, the first loss L 1 Is a class cross entropy penalty for training a class encoder f s And classifier f c Make the classifier f c The correct category can be output. The latter two losses are losses for enhancing the generalization capability of the class encoder, one is with the domain classifier f dc And class encoder f s To make countermeasure, fix f dc Training f s Let f dc The generation loss is as large as possible, and it is impossible to identify which domain the sample belongs to. The other is the loss realized by data enhancement, and the generator G is introduced to make two samples of different domains of the same classAnd->Respectively input to the category encoder f s Sum domain encoder f d Decoupling the category characteristics and domain characteristics of the sample, splicing and inputting the sample into a generator to generate a sample +.>The less domain information the feature generated by the class encoder contains, the closer the generated samples are +.>Three losses are used as constraint training to obtain a remote sensing classification network with good generalization.
Claims (1)
1. The domain countermeasure remote sensing image target classification method is characterized by comprising the following steps:
training a domain encoder and a domain classifier through multiple remote sensing image data sets;
the domain encoder and the domain classifier both adopt a ResNet50 structure; the domain encoder is used for outputting domain characteristic information of the remote sensing image, and has the function of decoupling the domain information of the remote sensing image and is used for assisting the class encoder to sufficiently decouple the class information irrelevant to the domain in the second step in a data enhancement mode; the domain classifier is used for outputting the domain to which the remote sensing image sample belongs and is used for antagonizing the class encoder in the second step, and training the class encoder to output the domain independent characteristics so that the domain classifier can not judge the domain information of the remote sensing image;
p remote sensing image data sets are adopted as p domains respectively; inputting the remote sensing image dataset into a domain encoder, inputting the output characteristics of the domain encoder into a domain classifier, and training to obtain the domain encoder only extracting the domain characteristics and the domain classifier outputting the domain information of the remote sensing image;
training loss is as follows:
where N is the total number of samples of the remote sensing image dataset, x i Is an input image, d i Is the domain of the remote sensing image, f d Is a domain encoder, f dc Is a domain classifier;
training a class encoder and a class classifier according to the domain encoder and the domain classifier obtained by training in the step one and the existing generator;
the class encoder and the classifier both adopt a ResNet50 model structure, and two full-connection layers with the length of 512 are newly added in the classifier; the penalty of training the class encoder and classifier includes classification penalty, antagonism penalty, and data enhancement penalty; the classification loss is used for training the correct class of the remote sensing image output by the class encoder and the classifier, and the counterloss and the data enhancement loss are used for improving the generalization capability of the class encoder;
input m-domain samplesGenerating category characteristics by a category encoder; input n-field samples of the same class->Generating domain features into a domain encoder; the method comprises the steps of fixing a domain encoder, a domain classifier and a generator, inputting class features into the classifier, the domain classifier and the generator respectively, inputting the domain features into the generator, and training the class encoder and the classifier;
the classification loss in the classifier is cross entropy loss, and the formula is as follows:
wherein n is k Is the number of samples of each remote sensing image dataset, f s Is a class encoder, f c Is a classifier gt i The true value is corresponding to the remote sensing image sample; x is x i Is a category feature;
fixing the domain classifier obtained in the first step, which is opposed to the class encoder; the domain classifier outputs the domain to which the remote sensing image belongs according to the domain information represented by the features, and trains the class encoder so that the features output by the class encoder do not contain the domain information; the countermeasures loss formula in the domain classifier is as follows:
where N is the total number of samples of the remote sensing image dataset, x i Is a category feature, d i Is the domain of the image, f d Is a domain encoder, f dc Is a domain classifier;
fixing stepThe domain encoder obtained in the step one, a generator is introduced, and the generator generates an image according to the category characteristics and the domain characteristics; selecting two images of the same category and different fieldsAnd->Respectively input into a category encoder and a domain encoder to obtain category characteristics x i And domain feature n; category characteristics x i And domain feature n generating an image by a generator +.>The generated image and the image inputted to the domain encoder +.>As a loss; the less domain information the feature generated by the class encoder contains, the less distance the generated image is from the image input to the domain encoder; the data enhancement loss is as follows:
wherein,and->Images of the same category and different domains, G is a generator, f s Is a class encoder, f d Is a domain encoder, l () is a distance function;
in summary, the total loss of the training in the second step is:
L＝L 1 +L 2 +L 3 (5)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310950762.7A CN116912593B (en) | 2023-07-31 | 2023-07-31 | Domain countermeasure remote sensing image target classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310950762.7A CN116912593B (en) | 2023-07-31 | 2023-07-31 | Domain countermeasure remote sensing image target classification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116912593A CN116912593A (en) | 2023-10-20 |
| CN116912593B true CN116912593B (en) | 2024-01-23 |
Family
ID=88362865
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310950762.7A Active CN116912593B (en) | 2023-07-31 | 2023-07-31 | Domain countermeasure remote sensing image target classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116912593B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102020122345A1 (en) * | 2019-08-27 | 2021-03-04 | Nvidia Corporation | CROSS-DOMAIN IMAGE PROCESSING FOR OBJECT REALIZATION |
| CN113486987A (en) * | 2021-08-04 | 2021-10-08 | 电子科技大学 | Multi-source domain adaptation method based on feature decoupling |
| CN113627443A (en) * | 2021-10-11 | 2021-11-09 | 南京码极客科技有限公司 | Domain self-adaptive semantic segmentation method for enhancing feature space counterstudy |
| CN114078276A (en) * | 2021-11-23 | 2022-02-22 | 华东交通大学 | Face living body detection method with condition-to-immunity domain generalization and network model architecture |
| CN114492640A (en) * | 2022-01-27 | 2022-05-13 | 云南联合视觉科技有限公司 | Domain-adaptive-based model training method, target comparison method and related device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20200075344A (en) * | 2018-12-18 | 2020-06-26 | 삼성전자주식회사 | Detector, method of object detection, learning apparatus, and learning method for domain transformation |
-
2023
- 2023-07-31 CN CN202310950762.7A patent/CN116912593B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102020122345A1 (en) * | 2019-08-27 | 2021-03-04 | Nvidia Corporation | CROSS-DOMAIN IMAGE PROCESSING FOR OBJECT REALIZATION |
| CN113486987A (en) * | 2021-08-04 | 2021-10-08 | 电子科技大学 | Multi-source domain adaptation method based on feature decoupling |
| CN113627443A (en) * | 2021-10-11 | 2021-11-09 | 南京码极客科技有限公司 | Domain self-adaptive semantic segmentation method for enhancing feature space counterstudy |
| CN114078276A (en) * | 2021-11-23 | 2022-02-22 | 华东交通大学 | Face living body detection method with condition-to-immunity domain generalization and network model architecture |
| CN114492640A (en) * | 2022-01-27 | 2022-05-13 | 云南联合视觉科技有限公司 | Domain-adaptive-based model training method, target comparison method and related device |
Non-Patent Citations (3)
| Title |
|---|
| Divergence-agnostic unsupervised domain adaptation by adversarial attacks;Jingjing Li et al.;《 IEEE Transactions on Pattern Analysis and Machine Intelligence》;第44卷(第11期);全文 * |
| Style-Content Metric Learning for Multidomain Remote Sensing Object Recognition;Wenda Zhao et al.;《The Thirty-Seventh AAAI Conference on Artificial Intelligence》;第37卷(第03期);全文 * |
| 面向小样本约束的域适应分类算法;戴宏 等;《计算机学报》;第45卷(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116912593A (en) | 2023-10-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | Learning a recurrent residual fusion network for multimodal matching | |
| CN112446423B (en) | Fast hybrid high-order attention domain confrontation network method based on transfer learning | |
| CN113076994B (en) | Open-set domain self-adaptive image classification method and system | |
| Liu et al. | Reinforced training data selection for domain adaptation | |
| Choudhuri et al. | Distribution alignment using complement entropy objective and adaptive consensus-based label refinement for partial domain adaptation | |
| CN113420775B (en) | Picture classification method based on nonlinearity self-adaptive subdomain field adaptation under extremely small quantity of training samples | |
| Qian et al. | Scene graph refinement network for visual question answering | |
| CN111985538A (en) | Small sample picture classification model and method based on semantic auxiliary attention mechanism | |
| JP2024521118A (en) | Transfer learning in image recognition systems | |
| CN113627543B (en) | Anti-attack detection method | |
| Bhatt et al. | An iterative similarity based adaptation technique for cross-domain text classification | |
| CN113052017B (en) | Unsupervised pedestrian re-identification method based on multi-granularity feature representation and domain self-adaptive learning | |
| WO2024130752A1 (en) | Conditional text-e-commerce picture retrieval method and system based on hierarchical style | |
| Mao et al. | Gap++: Learning to generate target-conditioned adversarial examples | |
| CN114863213B (en) | Domain generalization image recognition method based on causal decoupling generation model | |
| CN117746467A (en) | Modal enhancement and compensation cross-modal pedestrian re-recognition method | |
| Duan et al. | Dynamic unary convolution in transformers | |
| CN116821391A (en) | Cross-modal image-text retrieval method based on multi-level semantic alignment | |
| Zhu et al. | Transferable feature learning on graphs across visual domains | |
| CN116912593B (en) | Domain countermeasure remote sensing image target classification method | |
| CN114399055A (en) | Domain generalization method based on federal learning | |
| Yuan et al. | Unbiased feature enhancement framework for cross-modality person re-identification | |
| Li et al. | Disentangled features with direct sum decomposition for zero shot learning | |
| Wang et al. | Interpolation normalization for contrast domain generalization | |
| US20230196810A1 (en) | Neural ode-based conditional tabular generative adversarial network apparatus and method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |