CN114220003B - Multi-target unsupervised domain self-adaption method for large-range ground object segmentation - Google Patents
Multi-target unsupervised domain self-adaption method for large-range ground object segmentation Download PDFInfo
- Publication number
- CN114220003B CN114220003B CN202111423886.7A CN202111423886A CN114220003B CN 114220003 B CN114220003 B CN 114220003B CN 202111423886 A CN202111423886 A CN 202111423886A CN 114220003 B CN114220003 B CN 114220003B
- Authority
- CN
- China
- Prior art keywords
- domain
- target
- branch
- image
- model
- 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
Images
Classifications
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
A multi-target unsupervised domain self-adaptive method for large-range ground object segmentation comprises the following steps: step 1: selecting data of one of a plurality of fields from the obtained large-range remote sensing image, cutting the data to manufacture a source field data set after marking, and directly cutting images of other fields to manufacture a plurality of target field data sets; step 2: putting the source domain data set and the target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a segmentation model; and step 3: putting a plurality of target domain data sets into a segmentation model, and making a target domain image with a high-confidence pseudo label into a pseudo source domain data set through entropy minimum sorting; and 4, step 4: putting the pseudo source domain data set and the plurality of target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a final segmentation model; and 5: and putting all the remote sensing images into a model to obtain a segmentation result.
Description
Technical Field
The invention relates to the technical field of semantic segmentation in remote sensing images, in particular to a multi-target unsupervised domain self-adaptive method for large-range ground object segmentation.
Background
Land resource utilization and human activities are closely related, and accurate acquisition of land utilization information has great significance in better environmental protection, development of dominant economic industries, disaster prevention and the like. With the development of remote sensing technology and deep learning, the method for segmenting the ground features on the remote sensing image through the semantic segmentation algorithm is regarded as an efficient and convenient scheme.
The semantic segmentation algorithm can learn better feature extraction capability under the supervision of a large amount of data, but the trained model is not always effective. The acquisition of the remote sensing image is interfered by a plurality of aspects, so that the difference exists in the distribution of the characteristics, and the difference of the characteristics makes the semantic segmentation model unable to obtain the expected result on the untrained image. In such a scenario, the image with the label for training may be referred to as a source domain image, and the image without the label may be referred to as a target domain image. The segmentation of the ground objects in remote sensing images over a large area becomes a difficult task, since the drastic changes of the objects and scenes captured in remote sensing images are difficult to segment by a single model trained on data sets collected at several specific locations.
In order to solve the problem, the traditional unsupervised domain self-adaptive method enables the model to obtain better segmentation results on data in different fields in a mode of aligning the characteristics of a source domain and a target domain. However, these domain adaptive methods are all established under the condition of a single source domain and a single target domain, which still has a great problem in practical use, a large-range remote sensing image may have multiple domains, all target domains are divided into one target domain for domain adaptive learning, and the performance of the obtained segmentation model is far inferior to that of the segmentation model obtained by single source-single target domain adaptive learning. And aligning multiple target domain fields with the source domain respectively is costly and long-term. Therefore, designing a method which can align a plurality of target domains and does not reduce the model performance is an important premise for realizing the ground feature segmentation of the large-range remote sensing image.
According to the method, the source domain and the target domains are directly aligned by constructing the multi-branch unsupervised domain self-adaptive model, and the characteristics are further aligned by constructing the pseudo source domain data set through entropy minimization. The method has the advantages that the segmentation model learns the characteristics of multiple fields with low cost, the problem of field deviation between the source field and multiple target fields is well relieved, and therefore the effect of the segmentation model on the large-range remote sensing image is further improved.
Disclosure of Invention
The invention aims to realize the ground feature segmentation of a large-range remote sensing image, and provides a model which can directly align a source and a plurality of target domains, so that the segmentation model can obtain a better segmentation result on data of the plurality of domains with lower cost.
A multi-target unsupervised domain self-adaptive method for large-range ground object segmentation comprises the following steps:
step 1: selecting data of any one of a plurality of fields, cutting the data after marking to manufacture a source field data set, and directly cutting images of other fields except the selected field to manufacture a plurality of target field data sets;
step 2: putting the source domain data set and the target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a segmentation model;
and step 3: inputting a plurality of target domain data sets into a segmentation model, and making a target domain image with a high-confidence pseudo label into a pseudo source domain data set through entropy minimum sorting;
and 4, step 4: putting the pseudo source domain data set and the plurality of target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a final segmentation model;
and 5: putting all remote sensing images into a model to obtain a result;
in step 1, when the data of the selected region is marked, the feature type recognized by the model is the same as the number of marked types.
In step 2, when training the multi-branch unsupervised domain adaptive model to obtain a segmentation model, the method includes the following steps:
1) Constructing a multi-branch unsupervised domain self-adaptive model;
2) Taking the source domain data set and the plurality of target domain data sets as input, and training until the model performance reaches the optimum;
in step 1), the multi-branch unsupervised domain adaptive model is composed of a feature extractor, an invariant feature classifier, K target classifiers, K branch modules and K fine-grained discriminators (the value of K is equal to the number of target domains). The system comprises a feature extractor, a K target classifiers and K branch modules, wherein the K target classifiers and the K branch modules are connected behind the feature extractor, the K branch modules are respectively connected with K corresponding fine-grained discriminators, and meanwhile, the outputs of the K branch modules are also connected with the invariant feature classifiers.
In step 2), training each module in the multi-branch unsupervised domain self-adaptive model is supervised by different data, wherein the feature extractor and the invariant feature classifier are trained through all data; the kth target classifier, the kth branch module and the K fine-grain discriminators are trained on the source domain dataset and the kth target domain dataset (K =1,2, · ·, K).
The loss function used to train the feature extractor is defined as follows:
wherein L is seg Is the loss of segmentation from the source domain data,andare the countermeasure loss and classification loss generated by the target domain data.Is a domain label of the target domain image.Is the predicted probability that the target domain image will be generated after passing through the discriminator.Is the resulting prediction probability of the target domain image on the kth target classifier. n is s Andis the number of samples of the source domain and the kth target domain.Andare weighting factors that control the impact on the immunity to loss and classification loss.
The dimensionality of the domain label is the same as the number of the surface feature types, and the value of the domain label represents the probability that the model prediction pixel point belongs to a certain type.
The data flow of the source domain data set in the multi-branch unsupervised domain adaptive model comprises the following steps:
s 1) generating a feature map by a feature extractor from the source domain image;
s 2) extracting invariant features from the source domain feature map through all branch modules, inputting the generated invariant features into an invariant feature classifier to generate a prediction result, and calculating segmentation loss;
s 3) inputting the source domain invariant features into the discriminator and training the discriminator;
the data flow of the multiple target domain data sets in the multi-branch unsupervised domain adaptive model comprises the following steps:
s-1) generating a feature map from the target domain image through a feature extractor;
s-2) extracting invariant features from a feature map generated by the image of the kth target domain through a kth branch module, inputting the generated invariant features into an invariant feature classifier to generate a pseudo label, and inputting the invariant features into a kth discriminator to calculate the countermeasure loss;
s-3) inputting a feature map generated by the image of the kth target domain into a kth target classifier, and calculating classification loss;
s-4) inputting the invariant features of the kth target domain into the kth discriminator to train the kth discriminator;
in step 3, the target domain image is input into the model to obtain a prediction result, the image segmentation conditions are sequenced by calculating the entropy of the prediction result, and the image with high confidence coefficient is extracted.
In step 4, the target domain label graph with high confidence coefficient and the corresponding target domain image are used as a pseudo source domain, and the pseudo source domain and the target domain image are input into a multi-branch unsupervised domain self-adaptive model for training.
In step 5, all the remote sensing images are cut into small pictures which can be identified by the model, and the small pictures are input into the model to obtain a segmentation result.
Compared with the prior art, the invention has the following technical effects:
(1) The invention is a multi-target unsupervised domain self-adaptive model, and only needs to mark the data of one domain, so that the knowledge of a source domain can be migrated to other domains at one time. The segmentation model is better popularized, and meanwhile, the labor cost for realizing ground object segmentation in a large-range remote sensing image is reduced;
(2) According to the invention, through designing a multi-branch structure, firstly, a source domain and different target domains are aligned on different branches, so that the target domains are not directly aligned, and the condition that the result is poor due to overlarge gaps among the target domains is avoided; secondly, the branch module in the multi-branch structure separates the domain characteristics from the domain invariant characteristics, so that the domain characteristics are reserved, more effective segmentation can be realized, and meanwhile, the domain invariant characteristics enable the model to align the characteristics of different domains;
(3) According to the method, the pseudo labels with high confidence coefficient are obtained through entropy minimization, and the target domain images with the pseudo labels are aligned with other target domain images, so that the interior of the field is also aligned, and the performance of the model after intra-domain alignment is further improved;
drawings
The invention is further illustrated by the following examples in conjunction with the accompanying drawings:
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of source domain data flow of the present invention;
FIG. 3 is a flow chart of a target domain data flow in the present invention;
FIG. 4 is a diagram of a network architecture according to the present invention;
Detailed Description
The invention aims to realize the segmentation of the ground features in a wide-range remote sensing image, and provides a model which can directly align a source and a plurality of target domains, so that the segmentation model can obtain a better segmentation result on data of a plurality of domains with lower cost.
A multi-target unsupervised domain self-adaptive method for dividing ground features in a large-range remote sensing image comprises the following steps:
step 1: selecting data of one of a plurality of fields from the obtained large-range remote sensing image, cutting the data to manufacture a source field data set after marking, and directly cutting images of other fields to manufacture a plurality of target field data sets;
step 2: putting the source domain data set and the target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a segmentation model;
and step 3: inputting a plurality of target domain data sets into a segmentation model, and making a target domain image with a high-confidence pseudo label into a pseudo source domain data set through entropy minimum sorting;
and 4, step 4: putting the pseudo source domain data set and the plurality of target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a final segmentation model;
and 5: putting all remote sensing images into a model to obtain a result;
in step 2, when training the multi-branch unsupervised domain adaptive model to obtain a segmentation model, the method includes the following steps:
1) Constructing a multi-branch unsupervised domain self-adaptive model;
2) Taking the source domain data set and the plurality of target domain data sets as input, and training until the model performance reaches the optimum;
in step 1), the multi-branch unsupervised domain adaptive model is composed of a feature extractor, an invariant feature classifier, K target classifiers, K branch modules and K fine-grained discriminators (the value of K is equal to the number of target domains). The system comprises a feature extractor, a K target classifiers and K branch modules, wherein the K target classifiers and the K branch modules are connected behind the feature extractor, the K branch modules are respectively connected with K corresponding fine-grained discriminators, and meanwhile, the outputs of the K branch modules are also connected with the invariant feature classifiers.
In step 2), training of each module in the multi-branch unsupervised domain adaptive model is influenced by different data. Wherein the training of the feature extractor and invariant feature classifier is affected by all data; the kth target classifier, the kth branch module and the kth fine-grain discriminator training are influenced by the source domain dataset and the kth target domain dataset (K =1,2, ·, K).
The penalty function for training the feature extractor is defined as follows:
wherein L is seg Is the loss of segmentation from the source domain data,andare the countermeasure loss and classification loss generated by the target domain data.Is the domain label of the target domain image.Is the predicted probability that the target domain image will be generated after passing through the discriminator.Is the resulting prediction probability of the target domain image on the kth target classifier. n is s Andis the number of samples of the source domain and the kth target domain.Andare the weighting factors that control the impact on the immunity to loss and classification loss.
The data flow of the source domain data set in the multi-branch unsupervised domain adaptive model comprises the following steps:
s 1) generating a feature map by a feature extractor from the source domain image;
s 2) extracting invariant features from the source domain feature map through all the branch modules, inputting the generated invariant features into an invariant feature classifier to generate a prediction result, and calculating the segmentation loss;
s 3) inputting the source domain invariant features into the discriminator and training the discriminator;
the data flow of the multiple target domain data sets in the multi-branch unsupervised domain adaptive model comprises the following steps:
s-1) generating a feature map by the target domain image through a feature extractor;
s-2) extracting invariant features from a feature map generated by the image of the kth target domain through a kth branch module, inputting the generated invariant features into an invariant feature classifier to generate a pseudo label, and inputting the invariant features into a kth discriminator to calculate the countermeasure loss;
s-3) inputting a feature map generated by the image of the kth target domain into the kth target classifier, and calculating classification loss;
s-4) inputting the invariant features of the kth target domain into the kth discriminator to train the kth discriminator;
in step 3, the target domain image is input into the model to obtain a prediction result, the image segmentation conditions are sequenced by calculating the entropy of the prediction result, and the image with high confidence coefficient is extracted.
In step 4, the target domain label graph with high confidence coefficient and the corresponding target domain image are used as a pseudo source domain, and the pseudo source domain and the target domain image are input into a multi-branch unsupervised domain self-adaptive model for training.
In step 5, all remote sensing images are cut into small pictures which can be identified by the model, and the model is input to obtain a segmentation result.
Example (b):
the invention provides a method for segmenting ground objects in a large-range remote sensing image on the remote sensing image, which is carried out according to the following modes:
step 1: acquiring data of one of a plurality of fields from a large-range remote sensing image, and marking different ground object categories through ENVI; cutting the obtained image and the corresponding label to prepare a source domain data set, and directly cutting the images in other fields to prepare a plurality of target domain data sets; the cropped image size is 512 × 512 pixels.
Step 2: and constructing a multi-branch unsupervised domain self-adaptive model, wherein the multi-branch unsupervised domain self-adaptive model consists of a feature extractor, an invariant feature classifier, K target classifiers, K branch modules and K fine-grained discriminators (the value of K is equal to the number of target domains).
And 3, step 3: a loss function in the model is defined. The loss function used to train the feature extractor is defined as follows:
wherein L is seg Is the loss of segmentation from the source domain data,andare the competing and categorical penalties incurred by the target domain data.Is the domain label of the target domain image.Is the predicted probability that the target domain image will be generated after passing through the discriminator.Is the resulting prediction probability of the target domain image on the kth target classifier. n is s Andis the number of samples of the source domain and the kth target domain.Andis a weight controlling the impact on the immunity to loss and classification lossAnd (4) a heavy factor.
And 4, step 4: the model is trained using the source domain dataset and the plurality of target domain datasets as inputs. The data flow of the source domain data set in the multi-branch unsupervised domain adaptive model is as follows: generating a feature map from the source domain image through a feature extractor; extracting invariant features from the source domain feature map through all branch modules, inputting the generated invariant features into an invariant feature classifier to generate a prediction result, and calculating segmentation loss; and inputting the source domain invariant features into a discriminator and training the discriminator.
The data flow of the multiple target domain data sets in the multi-branch unsupervised domain adaptive model is as follows: generating a feature map from the target domain image through a feature extractor; extracting invariant features from a feature map generated by the image of the kth target domain through a kth branch module, inputting the generated invariant features into an invariant feature classifier to generate a pseudo label, and inputting the invariant features into a kth discriminator to calculate the antagonistic loss; inputting a feature map generated by the image of the kth target domain into a kth target classifier, and calculating classification loss; inputting the invariant features of the kth target domain into a kth discriminator, and training and inputting the kth discriminator;
and 5: the image of the target domain is input into the model to obtain pseudo labels, the confidence degrees of the image pseudo labels of the target domain are sequenced through entropy minimization, the target domain images with high confidence degrees are selected to be made into a pseudo source domain data set, the pseudo source domain data set and the images of the target domain are input into the multi-branch unsupervised domain self-adaptive model to be trained again, and a final segmentation model is obtained.
And 6: and cutting all the remote sensing images into small pictures which can be identified by the model, and inputting the small pictures into the model to obtain a segmentation result.
Claims (3)
1. A multi-target unsupervised domain self-adaptive method for large-range ground object segmentation comprises the following steps:
step 1: selecting data of any one of a plurality of fields from the obtained large-range remote sensing image, cutting the data to manufacture a source field data set after marking, and directly cutting images of other fields except the selected field to manufacture a plurality of target field data sets;
and 2, step: putting the source domain data set and the plurality of target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a segmentation model;
and step 3: inputting a plurality of target domain data sets into a segmentation model, and making a target domain image with a pseudo label with high confidence degree into a pseudo source domain data set through entropy minimum sorting;
and 4, step 4: putting the pseudo source domain data set and the plurality of target domain data sets into a multi-branch unsupervised domain self-adaptive model for training to obtain a final segmentation model;
and 5: putting all remote sensing images into a model to obtain a segmentation result;
in step 2, when the multi-branch unsupervised domain adaptive model is trained to obtain a segmentation model, the method comprises the following steps:
1) Constructing a multi-branch unsupervised domain self-adaptive model;
2) Taking the source domain data set and the plurality of target domain data sets as input, and training until the model performance reaches the optimum;
in the step 1), the multi-branch unsupervised domain self-adaptive model consists of a feature extractor, an invariant feature classifier, K target classifiers, K branch modules and K fine-grained discriminators; the system comprises a feature extractor, a K target classifiers and K branch modules, wherein the K target classifiers and the K branch modules are connected behind the feature extractor, the K branch modules are respectively connected with K corresponding fine-grained discriminators, and meanwhile, the outputs of the K branch modules are also connected with an invariant feature classifier;
in step 2), training each module in the multi-branch unsupervised domain self-adaptive model is supervised by different data, wherein the feature extractor and the invariant feature classifier are trained by all data; the kth target classifier, the kth branch module and the kth fine-grained discriminator are trained through a source domain data set and a kth target domain data set;
the loss function used to train the feature extractor is defined as follows:
wherein L is seg Is the loss of segmentation from the source domain data,andis the countermeasure loss and classification loss generated by the target domain data,is a domain label of the target domain image,is the prediction probability generated after the target domain image passes through the discriminator,is the resulting prediction probability of the target domain image on the kth target classifier,is the number of samples of the kth target domain,andis a weighting factor that controls the impact on loss resistance and classification loss;
the data flow of the source domain data set in the multi-branch unsupervised domain adaptive model comprises the following steps:
s 1) generating a feature map by a feature extractor from the source domain image;
s 2) extracting invariant features from the source domain feature map through all the branch modules, inputting the generated invariant features into an invariant feature classifier to generate a prediction result, and calculating the segmentation loss;
s 3) inputting the source domain invariant features into a discriminator and training the discriminator;
the data flow of the multiple target domain data sets in the multi-branch unsupervised domain adaptive model comprises the following steps:
s-1) generating a feature map by the target domain image through a feature extractor;
s-2) extracting invariant features from a feature map generated by the image of the kth target domain through a kth branch module, inputting the generated invariant features into an invariant feature classifier to generate a pseudo label, and inputting the invariant features into a kth discriminator to calculate the countermeasure loss;
s-3) inputting a feature map generated by the image of the kth target domain into the kth target classifier, and calculating classification loss;
s-4) inputting the invariant feature of the kth target domain into the kth discriminator, and training the kth discriminator.
2. The method according to claim 1, wherein in the step 3, the target domain image is input into the model to obtain a prediction result, and the image segmentation conditions are sorted by calculating entropy of the prediction result to extract the image with high confidence.
3. The method according to claim 1, wherein in step 4, the target domain label map with high confidence and the corresponding target domain image are used as a pseudo source domain, and the pseudo source domain and the target domain image are input into a multi-branch unsupervised domain adaptive model for training.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111423886.7A CN114220003B (en) | 2021-11-26 | 2021-11-26 | Multi-target unsupervised domain self-adaption method for large-range ground object segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111423886.7A CN114220003B (en) | 2021-11-26 | 2021-11-26 | Multi-target unsupervised domain self-adaption method for large-range ground object segmentation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114220003A CN114220003A (en) | 2022-03-22 |
CN114220003B true CN114220003B (en) | 2022-10-21 |
Family
ID=80698523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111423886.7A Active CN114220003B (en) | 2021-11-26 | 2021-11-26 | Multi-target unsupervised domain self-adaption method for large-range ground object segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114220003B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115082725B (en) * | 2022-05-17 | 2024-02-23 | 西北工业大学 | Multi-source domain self-adaption method based on reliable sample selection and double-branch dynamic network |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109948648A (en) * | 2019-01-31 | 2019-06-28 | 中山大学 | A kind of multiple target domain adaptive migration method and system based on member confrontation study |
CN111291705A (en) * | 2020-02-24 | 2020-06-16 | 北京交通大学 | Cross-multi-target-domain pedestrian re-identification method |
CN111382871A (en) * | 2020-03-11 | 2020-07-07 | 中国人民解放军军事科学院国防科技创新研究院 | Domain generalization and domain self-adaptive learning method based on data expansion consistency |
CN113096137A (en) * | 2021-04-08 | 2021-07-09 | 济南大学 | Adaptive segmentation method and system for OCT (optical coherence tomography) retinal image field |
CN113128411A (en) * | 2021-04-22 | 2021-07-16 | 深圳市格灵精睿视觉有限公司 | Cross-domain capture identification method and device, electronic equipment and storage medium |
CN113255823A (en) * | 2021-06-15 | 2021-08-13 | 中国人民解放军国防科技大学 | Unsupervised domain adaptation method and unsupervised domain adaptation device |
CN113486827A (en) * | 2021-07-13 | 2021-10-08 | 上海中科辰新卫星技术有限公司 | Multi-source remote sensing image transfer learning method based on domain confrontation and self-supervision |
CN113536972A (en) * | 2021-06-28 | 2021-10-22 | 华东师范大学 | Self-supervision cross-domain crowd counting method based on target domain pseudo label |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108062753B (en) * | 2017-12-29 | 2020-04-17 | 重庆理工大学 | Unsupervised domain self-adaptive brain tumor semantic segmentation method based on deep counterstudy |
US10885383B2 (en) * | 2018-05-16 | 2021-01-05 | Nec Corporation | Unsupervised cross-domain distance metric adaptation with feature transfer network |
US10915792B2 (en) * | 2018-09-06 | 2021-02-09 | Nec Corporation | Domain adaptation for instance detection and segmentation |
US11087142B2 (en) * | 2018-09-13 | 2021-08-10 | Nec Corporation | Recognizing fine-grained objects in surveillance camera images |
CN112991353B (en) * | 2021-03-12 | 2022-10-18 | 北京航空航天大学 | Unsupervised semantic segmentation method for cross-domain remote sensing image |
-
2021
- 2021-11-26 CN CN202111423886.7A patent/CN114220003B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109948648A (en) * | 2019-01-31 | 2019-06-28 | 中山大学 | A kind of multiple target domain adaptive migration method and system based on member confrontation study |
CN111291705A (en) * | 2020-02-24 | 2020-06-16 | 北京交通大学 | Cross-multi-target-domain pedestrian re-identification method |
CN111382871A (en) * | 2020-03-11 | 2020-07-07 | 中国人民解放军军事科学院国防科技创新研究院 | Domain generalization and domain self-adaptive learning method based on data expansion consistency |
CN113096137A (en) * | 2021-04-08 | 2021-07-09 | 济南大学 | Adaptive segmentation method and system for OCT (optical coherence tomography) retinal image field |
CN113128411A (en) * | 2021-04-22 | 2021-07-16 | 深圳市格灵精睿视觉有限公司 | Cross-domain capture identification method and device, electronic equipment and storage medium |
CN113255823A (en) * | 2021-06-15 | 2021-08-13 | 中国人民解放军国防科技大学 | Unsupervised domain adaptation method and unsupervised domain adaptation device |
CN113536972A (en) * | 2021-06-28 | 2021-10-22 | 华东师范大学 | Self-supervision cross-domain crowd counting method based on target domain pseudo label |
CN113486827A (en) * | 2021-07-13 | 2021-10-08 | 上海中科辰新卫星技术有限公司 | Multi-source remote sensing image transfer learning method based on domain confrontation and self-supervision |
Non-Patent Citations (4)
Title |
---|
Multi-Target Adversarial Frameworks for Domain Adaptation in Semantic Segmentation;Antoine Saporta et al;《arXiv》;20210916;第3.2-3.3节,图2-3 * |
Multi-Target Domain Adaptation with Collaborative Consistency Learning;Takashi Isobe et al;《arXiv》;20210608;1-10 * |
遥感图像语义分割中的弱监督域自适应算法;丁一鹏 等;《计算机工程与应用》;20210825;1-10 * |
领域自适应研究综述;李晶晶 等;《计算机工程》;20210630;第47卷(第6期);1-13 * |
Also Published As
Publication number | Publication date |
---|---|
CN114220003A (en) | 2022-03-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
Saha et al. | Unsupervised deep change vector analysis for multiple-change detection in VHR images | |
Wu et al. | Rapid target detection in high resolution remote sensing images using YOLO model | |
WO2020163455A1 (en) | Automatic optimization of machine learning algorithms in the presence of target datasets | |
CN108875816A (en) | Merge the Active Learning samples selection strategy of Reliability Code and diversity criterion | |
CN105574505A (en) | Human body target re-identification method and system among multiple cameras | |
CN112906606B (en) | Domain self-adaptive pedestrian re-identification method based on mutual divergence learning | |
CN112819065B (en) | Unsupervised pedestrian sample mining method and unsupervised pedestrian sample mining system based on multi-clustering information | |
Tung et al. | Collageparsing: Nonparametric scene parsing by adaptive overlapping windows | |
CN108537816A (en) | A kind of obvious object dividing method connecting priori with background based on super-pixel | |
CN111783521A (en) | Pedestrian re-identification method based on low-rank prior guidance and based on domain invariant information separation | |
CN112364721A (en) | Road surface foreign matter detection method | |
CN102750712A (en) | Moving object segmenting method based on local space-time manifold learning | |
CN108805102A (en) | A kind of video caption detection and recognition methods and system based on deep learning | |
CN114220003B (en) | Multi-target unsupervised domain self-adaption method for large-range ground object segmentation | |
CN104063701B (en) | Fast electric television stations TV station symbol recognition system and its implementation based on SURF words trees and template matches | |
CN107609509A (en) | A kind of action identification method based on motion salient region detection | |
CN109214430A (en) | A kind of recognition methods again of the pedestrian based on feature space topology distribution | |
Zhang et al. | Few-shot object detection with self-adaptive global similarity and two-way foreground stimulator in remote sensing images | |
Xu et al. | Hierarchical online domain adaptation of deformable part-based models | |
CN114782752A (en) | Small sample image grouping classification method and device based on self-training | |
CN114329031A (en) | Fine-grained bird image retrieval method based on graph neural network and deep hash | |
Casagrande et al. | Abnormal motion analysis for tracking-based approaches using region-based method with mobile grid | |
CN104517127A (en) | Self-learning pedestrian counting method and apparatus based on Bag-of-features model | |
CN113627443A (en) | Domain self-adaptive semantic segmentation method for enhancing feature space counterstudy |
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 |