CN112418257B - Effective zero sample learning method based on potential visual attribute mining - Google Patents
Effective zero sample learning method based on potential visual attribute mining Download PDFInfo
- Publication number
- CN112418257B CN112418257B CN201910778304.3A CN201910778304A CN112418257B CN 112418257 B CN112418257 B CN 112418257B CN 201910778304 A CN201910778304 A CN 201910778304A CN 112418257 B CN112418257 B CN 112418257B
- Authority
- CN
- China
- Prior art keywords
- visual
- domain
- potential
- feature
- attribute
- 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
- 230000000007 visual effect Effects 0.000 title claims abstract description 88
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000005065 mining Methods 0.000 title claims abstract description 22
- 238000013507 mapping Methods 0.000 claims abstract description 26
- 239000011159 matrix material Substances 0.000 claims abstract description 22
- 230000006870 function Effects 0.000 claims description 21
- 238000012549 training Methods 0.000 claims description 18
- 238000010845 search algorithm Methods 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 abstract description 11
- 230000008569 process Effects 0.000 description 7
- 238000000605 extraction Methods 0.000 description 4
- 241000282414 Homo sapiens Species 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 3
- 230000001149 cognitive effect Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 241000283070 Equus zebra Species 0.000 description 1
- 241001465754 Metazoa Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000036548 skin texture Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a zero sample learning method based on potential visual attribute mining. The invention mainly relates to the following steps: firstly, calculating a feature mapping matrix represented by manually defined attributes in an original feature domain, and mining a potential visual attribute set through optimization of a sparse dictionary model. And further constructing a pair-wise visual dictionary model, and guiding the learning of the target feature domain by taking the mapping relation obtained by the original feature domain as constraint information. And finally, extracting visual features of a depth residual error network (ResNet) from the input image, obtaining feature discrimination semantic representation, and realizing the prediction of the image label category. The image recognition method has the advantages of no manual participation, high recognition rate, strong robustness and the like, has certain use value aiming at the specific field of recognition of semantic objects under the condition of a small number of samples and zero samples, and can be applied practically.
Description
Technical Field
The invention provides a zero sample learning method based on potential visual attribute mining, which solves the recognition task of semantic objects in images under the condition of a small number of samples and zero samples and is a novel technology in the field of image recognition.
Background
Nowadays, with the rapid development of digital multimedia technology, image and video data facing people in daily life have been increased explosively, and the human society is promoted to step into the big data era. In many complicated multimedia data, how to effectively analyze and understand the content of visual data becomes increasingly important. In order to solve the problems, many computer vision methods proposed in the past years make certain progress in the aspects of feature extraction, strong supervised learning, object model construction and the like, and are successfully applied to the daily life fields of multimedia data retrieval, scene analysis, visual information text description and the like. However, due to the semantic gap between the bottom-layer features and the middle-layer and high-layer information of the existing visual data, the existing method still progresses slowly on key problems of learning of a small number of samples and zero samples, semantic attribute mining, cross-feature domain object models and the like.
The recognition of image semantic object information plays an important role in multimedia data analysis and understanding. With the construction of large-scale image databases and the wide application of convolutional neural networks, the object recognition method is rapidly improved. However, the current object recognition method still has the following disadvantages: first, the conventional visual recognition algorithm is based on a supervised learning manner, and a large amount of labeled image data is required for training a more robust classification model. However, in practical application, the acquisition of a large amount of labeled data is often difficult to realize, and particularly, the acquisition of samples needing expert knowledge to be subjected to fine labeling and rare type samples is very difficult; secondly, the existing object recognition algorithm is limited by the trained object types, the newly introduced object types cannot be effectively recognized, and the expansibility of the method is limited; in addition, traditional object recognition algorithms are not consistent with human cognitive mechanisms. The recognition of semantic objects by human beings is based on the difference between similar object classes, and the knowledge is migrated to the discrimination of new object classes. For example, a child who never sees a zebra can easily identify the animal by observing a skin texture difference from that of a general horse. In order to solve the above problems to be solved by the current object recognition method, an object recognition method based on zero samples has been proposed in recent years. The method can effectively identify the class of the newly introduced object which does not appear in the training process by means of the visual cognitive mechanism of the discriminator and through the extraction of the middle and high level semantic information.
However, the current research aiming at the zero sample object identification method still faces the following problems to be solved: first, there is a lack of research into potential visual attribute mining methods. The existing zero sample object recognition algorithm is usually only based on manually defined attribute representation, and the exploration of discriminant visual attributes is omitted, so that the semanteme of attribute space feature representation is difficult to improve. In addition, current zero-sample learning methods lack the study of object recognition models with cross-feature domain adaptation. In the process of identifying the category of a newly introduced object, the existing method always directly acts the original feature domain mapping relation on a target domain, and the defect that the feature transformation does not have domain adaptability is overcome.
Disclosure of Invention
The technical problem to be solved is as follows:
aiming at the defects of the current zero sample learning-based object identification method in the aspects of image feature representation discrimination capability and cross-feature domain adaptability, the invention tries to analyze and solve the problem from the perspective of mining the potential visual attributes of image data with more practical value. Compared with the existing zero sample learning-based object identification algorithm, the method has the main advantages that: firstly, the problem that the existing zero sample learning-based object recognition algorithm only depends on manually defined attributes, so that the feature representation semantic information is deficient is solved. The method provided by the invention excavates the potential visual attribute of the image data, establishes the mapping relation from the visual feature space to the semantic attribute space through the joint optimization of the objective function by combining the manually defined attribute, and enhances the discriminability of image feature representation. Secondly, the method takes the mapping relation learned in the original characteristic domain as a regularization constraint term, guides the learning of the mapping function from the visual characteristic to the semantic attribute representation in the target characteristic domain, improves the adaptability of the object recognition model in the characteristic domain migration process, and improves the generalization performance of the zero sample learning algorithm.
The technical scheme is as follows:
in order to establish the correlation between the image visual characteristics and the semantic attribute representation, the method comprises two parts in the training stage, namely an original characteristic domain (known object type) training process and a target characteristic domain (unknown object type) training process, wherein the original characteristic domain and the target characteristic domain have no intersection. Firstly, by constructing an optimization model of the following objective function, a mapping function from visual features to manually defined attribute representation is learned in the original feature domain training process. The mathematical form of the objective function is as follows:
wherein, X s All the graphs in the invented method are visual feature matrixes of all training samples in the original feature domainThe visual features of the image are all image features extracted by using a depth residual error network (ResNet); a. The s A manually defined attribute representation matrix, D, representing all training samples s For the mapping of visual features to manually defined attribute representations requiring an optimization solution, d i Representation matrix D s The ith column vector. The least mean square optimization problem with the above objective function as the standard can be obtained by the following mathematical form with respect to the mapping moment D s Closed solution of the array.
The manually defined attribute representation is mainly based on the sharing characteristics among all the classes of objects, and in order to obtain feature semantic representation with higher discriminability, the method further excavates the potential visual attributes of all the classes of objects besides the attributes based on manual definition so as to further improve the discriminability after the visual features of the images are mapped to the semantic attribute space. Specifically, in the raw feature domain, the underlying visual properties of the image data are mined by optimizing the following objective function.
Wherein,
set of potential visual attributes learned for the original feature domain, d j Is a matrix>
Corresponds to the jth column vector, i.e. the jth potential visual attribute, in>
Matrix X representing visual characteristics s Is at>
Is represented by a value of>
The correlation between visual features and potential visual attributes is described, parameter λ 1 For balancing regularization constraints in the objective function. Sparse representation model with the above objective function as a criterion can be paired with ∑ er>
And carrying out iterative solution by adopting an alternating optimization mode.
Mapping matrix D extracted from original characteristic domain s And a set of visual attributes
The method further establishes the association between the visual features and the semantic attribute representation aiming at the target feature domain, and simultaneously mines the potential discriminant visual attributes. Specifically, the method constructs a cross-feature-domain paired visual dictionary learning model, and jointly realizes the solution of a characteristic mapping matrix and the extraction of potential visual attributes. The specific mathematical form of the optimization model is as follows:
in the above formula, X t Training the feature matrix of the sample for the target domain, Y t Artificially defined attribute representation of training samples for a target domain, D t A mapping of visual features in the target domain to manually defined attribute representations is described,
for a mined set of target domain potential visual attributes, ->
Represents X t A representation coefficient on a visual property.
In the target domain training model, the first two cost items
Reconstructing residual terms for data, ensuring the use of a learned mapping matrix D t And the potential visual attribute->
Can effectively reconstruct input characteristic data X t (ii) a Cost item
Enabling the manually defined attribute representations corresponding to the same visual features and the potential visual attribute representations to have similarity; the method aims to solve the problem of inconsistency of data feature distribution between an original feature domain and a target feature domain. Cost item is introduced into dictionary model>
And &>
Learning in the domain by means of primitive features D s 、
And using the feature mapping matrix D as constraint information to guide the feature mapping matrix D in the target feature domain t And the potential visual attribute->
The extraction of (2) to solve the problem of domain adaptability among different feature domains; in addition, a regularization term in the model>
The effectiveness of the final solution of the objective function is ensured; parameters alpha, beta, rho and mu in the formula 1 、μ 2 、μ 3 For balancing dictionary modelsDifferent cost terms in (1). The dictionary learning objective function is a multivariable optimization problem, and the iterative computation is specifically carried out by adopting an alternative optimization strategy in the method. Specifically, when one variable in the objective function is optimized, the other variables are fixed, so that the original optimization problem can be converted into a plurality of optimization sub-problems to be solved.
In the testing stage, an input image is given, the image depth residual error network (ResNet) features are firstly extracted, the semantic attribute representation of the input image is further calculated by optimizing the following objective function, and the specific solved mathematical form is as follows:
in the formula, x is the depth characteristic of the input image, y is the characteristic x in a mapping matrix D t The above expression coefficient, i.e. the semantic attribute representation of the input visual feature x, the parameter λ 2 The method is used for balancing the residual error item of data reconstruction and coefficient sparsity constraint. In the method, for solving the objective function, a Feature-signature search (Feature-signature) algorithm is specifically adopted to calculate the variable y to be optimized. After the semantic attribute representation y of the input image is obtained, the semantic label category of the input image is further obtained through a Nearest Neighbor (NN) search algorithm, and finally the problem of identification of the zero sample object category is solved.
[1] Sparse representation models, see literature (M.Aharon, M.Elad, A.Bruckstein, K-svd: an algorithm for designing-complete dictionary for sparse representation, IEEE trans.Signal Process.Vol.54, 4311-4322, 2006.)
[2] Deep residual error network (ResNet), see literature (K.He, X.Zhang, S.Ren, and J.Sun.deep residual error for image Recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.)
[3] Feature assignment search (Feature-signature) algorithm, see literature (Honglak Lee, alexis Battle, rajat Raina, and Andrew Y. Ng. Effectionspace coding algorithms of the Conference on Neural Information Processing Systems, pages 801-808.2007.)
[4] Zero sample method DAP, see literature (C.H.Lambert, H.Nickisch, and S.Harmeling.Attribute-based classification for zero-shot visual object localization. IEEE Transactions on Pattern Analysis and Machine Analysis, vol.36,453-465,2014.)
[5] Zero sample method SAE, see literature (E.Kodirov, T.Xiaong, and S.Gong.Semantic Autoencoder for zero-shot
[6] Zero sample method ESZSL, see literature (B.Romera-seeds and P.H.Torr.an embrassin simple approach zero-shot Learning. In International Conference on Machine Learning, pages2152-2161, 2015.)
Has the advantages that:
the effective zero sample learning method based on potential visual attribute mining provided by the invention has the characteristics of no manual participation, high classification accuracy and the like. The main innovation points of the invention are the proposed potential visual attribute mining method and the constructed cross-feature-domain pairwise visual dictionary learning model. The method provided by the invention is different from the current zero sample learning method which only depends on manually defined attributes by mining the potential visual attributes of the image data, and overcomes the problem of lack of semantic expression of the current image data features. The inventive method is practical and effective.
Drawings
FIG. 1 is a flow chart of a zero sample learning method based on potential visual attribute mining
FIG. 2 evaluation of accuracy of the method of the present invention and existing zero sample learning methods on the AWA2 object recognition database
FIG. 3 recognition accuracy of the method of the present invention on typical object class images
Detailed Description
The method is realized on a Matlab R2016b experimental platform, and mainly comprises four steps as shown in fig. 1, namely original feature domain artificial definition attribute representation mapping matrix learning, original feature domain potential visual attribute mining, target feature domain paired visual dictionary learning and input image semantic label prediction. The method comprises the following specific steps:
step one, learning a mapping matrix represented by an original feature domain manually defined attribute:
training sample visual characteristic data X according to original characteristic domain s And a manually defined attribute representation A corresponding to the sample data s Calculating a mapping matrix D between the visual features of the image in the original feature domain and the manually defined attribute representation s . Specifically, the method is obtained by optimizing the following objective function:
step two, mining the potential visual attributes of the original characteristic domain:
and (4) excavating potential visual properties of image data in the original features by optimizing the following sparse dictionary learning model.
In the formula
The set of potential visual attributes for the acquired original feature domain.
Step three, learning a target feature domain pair visual dictionary:
constructing a cross-feature-domain paired visual dictionary learning model in a target feature domain and utilizing a mapping matrix D learned in an original feature domain s And a set of potential visual attributes
Guiding the solving of the mapping function between the image characteristics and the semantic attribute representation, and mining the potential visual attributes. The objective function of the pairwise visual dictionary model is as follows:
step four, input image semantic label prediction:
inputting an image, extracting a depth residual network (ResNet) feature x of the image, and calculating semantic attribute representation of the input image by optimizing the following objective functions:
and further utilizing the solved semantic attribute to express y, and obtaining the semantic label category of the input image through a Nearest domain search algorithm (NN) to realize the identification of the zero sample object category.
The invention is characterized in step two and step three. Any use of step two and step three is within the protection field of the present invention.
Claims (5)
1. An effective zero sample learning method based on potential visual attribute mining is characterized by comprising the following steps:
(1) Firstly, extracting depth residual error network (ResNet) visual characteristics of images for all training sample images of an image data set in an original characteristic domain, then inputting the visual characteristics into an objective function of a least mean square problem, and solving a mapping matrix represented by manually defined attributes of the original characteristic domain;
(2) In order to realize the mining of the potential visual attributes of the image data, inputting an original characteristic domain sparse dictionary model by using the training sample image characteristics extracted in the step (1) to obtain an original characteristic domain potential visual attribute set; the mining formula of the potential visual attributes of the image data is as follows:
X s for training the visual feature matrix of the image, < >>
Set of potential visual attributes learned for a raw feature domain，Y s v Is X s Is at>
Is a coefficient of expression, parameter λ 1 For balancing regularization constraints in the objective function;
(3) Inputting the mapping matrix and the potential visual attribute set learned in the steps (1) and (2) into a pair visual dictionary model constructed, and solving the mapping matrix and the potential visual attribute set in the target feature domain; wherein, the objective function of the paired visual dictionary model is as follows:
wherein, X t Training the feature matrix of the sample for the target domain, Y t Artificially defined attribute representation of training samples for a target domain, D t A mapping of visual features in a target domain to manually defined attribute representations is described,
set of potential visual properties for mined target domain, Y t v Represents X t Representation coefficients on a visual attribute>
Parameters α, β, ρ, μ for the regularization term 1 、μ 2 、μ 3 For balancing different cost terms in the dictionary model;
(4) Extracting depth residual error network (ResNet) features from the test image, inputting the features into a target feature domain sparse dictionary model, solving semantic feature representation of the test image, and predicting input image class labels through a nearest domain search algorithm (NN).
2. The effective zero-sample learning method based on potential visual attribute mining as claimed in claim 1, wherein the obtaining of the original feature domain potential visual attribute set comprises extracting deep residual error network (ResNet) visual features from a training image set, optimizing a sparse dictionary learning model, and realizing mining of the original feature domain potential visual attribute set.
3. The effective zero sample learning method based on potential visual attribute mining as claimed in claim 1, wherein the number of visual dictionary words in the original feature domain sparse dictionary model is 200.
4. The efficient zero-sample learning method based on potential visual attribute mining as claimed in claim 1, wherein the paired visual dictionary model introduces consistency constraints between the manually defined attribute representations and the potential visual attribute representations, and guides the learning of the target feature domain by using the mapping matrix and the set of potential visual attributes in the original target domain.
5. The method of claim 4, wherein the consistency constraint provides similarity between the manually defined attribute representation of the input image features and the potential visual attribute representation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910778304.3A CN112418257B (en) | 2019-08-22 | 2019-08-22 | Effective zero sample learning method based on potential visual attribute mining |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910778304.3A CN112418257B (en) | 2019-08-22 | 2019-08-22 | Effective zero sample learning method based on potential visual attribute mining |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112418257A CN112418257A (en) | 2021-02-26 |
| CN112418257B true CN112418257B (en) | 2023-04-18 |
Family
ID=74779771
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910778304.3A Active CN112418257B (en) | 2019-08-22 | 2019-08-22 | Effective zero sample learning method based on potential visual attribute mining |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112418257B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113221814A (en) * | 2021-05-26 | 2021-08-06 | 华瑞新智科技（北京）有限公司 | Road traffic sign identification method, equipment and storage medium |
| CN114037879A (en) * | 2021-10-22 | 2022-02-11 | 北京工业大学 | Dictionary learning method and device for zero sample recognition |
| CN116363762B (en) * | 2022-12-23 | 2024-09-03 | 南京羽丰视讯科技有限公司 | Living body detection method, training method and device of deep learning model |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019055114A1 (en) * | 2017-09-12 | 2019-03-21 | Hrl Laboratories, Llc | Attribute aware zero shot machine vision system via joint sparse representations |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10074041B2 (en) * | 2015-04-17 | 2018-09-11 | Nec Corporation | Fine-grained image classification by exploring bipartite-graph labels |
| CN109643384A (en) * | 2016-08-16 | 2019-04-16 | 诺基亚技术有限公司 | Method and apparatus for zero sample learning |
| CN106485271B (en) * | 2016-09-30 | 2019-11-15 | 天津大学 | A kind of zero sample classification method based on multi-modal dictionary learning |
| US10908616B2 (en) * | 2017-05-05 | 2021-02-02 | Hrl Laboratories, Llc | Attribute aware zero shot machine vision system via joint sparse representations |
| EP3619644A4 (en) * | 2017-05-05 | 2021-01-20 | HRL Laboratories, LLC | Zero shot machine vision system via joint sparse representations |
| CN107491788A (en) * | 2017-08-21 | 2017-12-19 | 天津大学 | A kind of zero sample classification method based on dictionary learning |
| CN108376267B (en) * | 2018-03-26 | 2021-07-13 | 天津大学 | Zero sample classification method based on class transfer |
-
2019
- 2019-08-22 CN CN201910778304.3A patent/CN112418257B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019055114A1 (en) * | 2017-09-12 | 2019-03-21 | Hrl Laboratories, Llc | Attribute aware zero shot machine vision system via joint sparse representations |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112418257A (en) | 2021-02-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107766933B (en) | Visualization method for explaining convolutional neural network | |
| CN112418257B (en) | Effective zero sample learning method based on potential visual attribute mining | |
| CN108595636A (en) | The image search method of cartographical sketching based on depth cross-module state correlation study | |
| CN115934990B (en) | Remote sensing image recommendation method based on content understanding | |
| CN111680176A (en) | Remote sensing image retrieval method and system based on attention and bidirectional feature fusion | |
| CN105389326B (en) | Image labeling method based on weak matching probability typical relevancy models | |
| CN110443293A (en) | Based on double zero sample image classification methods for differentiating and generating confrontation network text and reconstructing | |
| CN106022254A (en) | Image recognition technology | |
| CN113761259A (en) | Image processing method and device and computer equipment | |
| CN113269224A (en) | Scene image classification method, system and storage medium | |
| CN114926725A (en) | Online financial group partner fraud identification method based on image analysis | |
| Adnan et al. | An improved automatic image annotation approach using convolutional neural network-Slantlet transform | |
| CN115131313A (en) | Hyperspectral image change detection method and device based on Transformer | |
| CN113988147A (en) | Multi-label classification method and device for remote sensing image scene based on graph network, and multi-label retrieval method and device | |
| CN110705384B (en) | Vehicle re-identification method based on cross-domain migration enhanced representation | |
| Kaur et al. | A methodology for the performance analysis of cluster based image segmentation | |
| López-Cifuentes et al. | Attention-based knowledge distillation in scene recognition: the impact of a dct-driven loss | |
| CN114187183A (en) | Fine-grained insect image classification method | |
| CN112990340B (en) | Self-learning migration method based on feature sharing | |
| CN114187546A (en) | Combined action recognition method and system | |
| CN113779520A (en) | Cross-space target virtual identity correlation method based on multilayer attribute analysis | |
| CN111291787B (en) | Image labeling method based on forward-multi-reverse collaborative sparse representation classifier | |
| CN112364193A (en) | Image retrieval-oriented method for fusing multilayer characteristic deep neural network model | |
| CN116012840B (en) | Three-dimensional point cloud semantic segmentation labeling method based on active learning and semi-supervision | |
| Jafrasteh et al. | Generative adversarial networks as a novel approach for tectonic fault and fracture extraction in high resolution satellite and airborne optical images |
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 |