CN109582960B - Zero example learning method based on structured association semantic embedding - Google Patents

Zero example learning method based on structured association semantic embedding Download PDF

Info

Publication number
CN109582960B
CN109582960B CN201811423691.0A CN201811423691A CN109582960B CN 109582960 B CN109582960 B CN 109582960B CN 201811423691 A CN201811423691 A CN 201811423691A CN 109582960 B CN109582960 B CN 109582960B
Authority
CN
China
Prior art keywords
pseudo
class
category
space
semantic embedding
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
Application number
CN201811423691.0A
Other languages
Chinese (zh)
Other versions
CN109582960A (en
Inventor
熊红凯
陈宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Jiaotong University
Original Assignee
Shanghai Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Jiaotong University filed Critical Shanghai Jiaotong University
Priority to CN201811423691.0A priority Critical patent/CN109582960B/en
Publication of CN109582960A publication Critical patent/CN109582960A/en
Application granted granted Critical
Publication of CN109582960B publication Critical patent/CN109582960B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • 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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines

Abstract

The invention provides a zero-instance learning method based on structured association semantic embedding, which constructs a semantic embedding space by extracting attribute features of an input image, excavates a semantic embedding association relation in the semantic embedding space and constructs a local linear correlation graph structure, and then utilizes the structured restriction of the semantic embedding space to constrain a classification model of a known class, thereby finally synthesizing a model of an unknown class and realizing the effective prediction of the class of an untrained sample. The invention can keep the global structure of the semantic space while digging the semantic embedded information with similar categories, can obtain effective semantic representation and effectively improve the category prediction accuracy of unseen category image samples.

Description

Zero example learning method based on structured association semantic embedding
Technical Field
The invention relates to the technical field of computer vision, in particular to a zero-example learning method based on structured association semantic embedding.
Background
Zero example learning is a very important and challenging problem in the field of computer vision. State of the art supervised learning approaches have achieved significant achievements in a number of areas, however the performance of supervised learning is highly dependent on large scale labeled data. Conversely, zero-instance learning may take knowledge from samples of trained classes, migrate to samples of untrained classes, and enable differentiation of untrained class samples. To achieve this goal, it is important that the knowledge between the trained class samples and the untrained class samples be migrated and that the untrained class samples be efficiently described.
Through the literature search of the prior art, l.c.h, n.hannes and h.stafan, in the article "Attribute-based classification for zero-shot visual object classification" published in IEEE Transactions on Pattern Analysis and Machine Analysis conference 2014, a method for learning an Attribute classifier is proposed, i.e., a classifier is learned for each Attribute, and samples with no category are distinguished through the Attribute classifier output of discriminant test data. However, a bias is introduced in the process of establishing the classifier, so that the discrimination of the unseen class sample is influenced. To solve this problem, researchers have proposed to directly measure the correlation between images and class embedding, by learning a linear or non-linear compatibility function, to directly correlate images with class embedding. However, such methods can be biased when the training data is not consistent with the test data distribution, so researchers have attempted to solve the zero-instance learning problem using fusion model-based methods, i.e., considering class embedding as a combination of already seen class components. The 'Zero-shot Learning by y context communication of semantic embedding of unseen classes' published by Mohammad et al in the International Conference on Learning retrieval meeting of 2014 proposes that semantic embedding of unseen classes is combined and synthesized by the embedded weights of the seen classes, and the combined weights are corresponding probability outputs of pre-trained classifiers thereof. However, this method requires training the classifier in advance, which is time consuming and inefficient. Soravit et al, in the article "Synthesized classifiers for zero-shot learning" published at the IEEE Conference on Computer Vision and Pattern Recognition Conference of 2016, proposed a method for implementing semantic embedding and knowledge migration between classifiers using pseudo-classes, and synthesizing new classifiers by convex combination of the pseudo-classes. This approach ignores the semantically embedded local structure, which may provide more information for the exact synthesis classifier.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a zero-example learning method based on structured association semantic embedding.
The invention provides a zero example learning method based on structured association semantic embedding, which comprises the following steps:
step 1: extracting attribute features of the known category images;
step 2: establishing a semantic embedding space according to the attribute characteristics, and establishing a model space formed by classification models of known class samples;
and step 3: introducing R pseudo categories into the semantic embedding space and the model space to serve as relationship migration carriers of the semantic embedding space and the model space;
and 4, step 4: acquiring a scaled squared Euclidean distance between a real category and a pseudo category of a known category image, and taking the scaled squared Euclidean distance as a connection weight;
and 5: reconstructing a graph structure of a semantic embedding space in a model space according to the connection weight;
step 6: obtaining R of each real category of the known category image with the connection weight within a preset rangenA pseudo category;
and 7: constructing a set of semantic embedding rules for each semantic embedding in the semantic embedding spacenAn objective function synthesized by pseudo classes;
and 8: constructing a classification model, wherein the classification model consists of contiguous RnSynthesizing the pseudo classes;
and step 9: learning R pseudo classes, R, of the semantic embedding space and model space according to the objective functionnPseudo classes to minimize loss of the classification model;
step 10: extracting attribute features of unknown image, and using R as said attribute featuresnCharacterizing each pseudo category to obtain a graph structure of the unknown category image;
step 11: and projecting the graph structure of the unknown class image from the semantic embedding space to a model space to obtain projected pseudo classes, and synthesizing the classification model of the unknown class image according to the projected pseudo classes.
Optionally, the step 4 includes:
in a training sample, the number of known classes is S, the number of unknown classes is U, and the number of pseudo classes is R; defining the coordinates of each class i as si(i ═ 1,2, …, S + U), all classes of the pre-set training samples are on the same manifold of the semantic embedding space, with the help of the introduced pseudo-class pm(m-1, 2, …, R) preserving the correlation between training sample attributes, presetting each semantic embedding and adjacent RnPseudo classes are embedded in local linear blocks of the space manifold semantically; defining the scaled squared Euclidean distance as a similarity measure; each siPseudo class p with neighborsmThe calculation formula of the connection weight between the two is as follows:
Figure BDA0001881092950000031
wherein s isiSemantic embedding representing class i, with neighboring pseudo-classes sharing RnA plurality of; p is a radical ofmDenotes siNearest neighbor RnOne of the pseudo categories; w is amiEach representsS isiPseudo class p of its neighborsmWeight of connection between wmiThe larger the value of (A), the more the pseudo class p is representedmEmbedding with semantics siThe greater the similarity between;
embedding R corresponding to neighbor according to each semanticnPseudo categories, namely constructing a category association diagram of a semantic embedding space; the vertex set of the class association graph comprises semantically embedded attributes and introduced pseudo classes, and the edge set comprises edges connecting the attributes and the pseudo classes; the semantic embedding representation of each pseudo category neighborhood is as follows:
Figure BDA0001881092950000032
wherein, mumiRepresenting the weight between semantic embeddings and pseudo classes.
Optionally, the step 8 includes:
defining the classification model corresponding to each class i as ui(i ═ 1,2, …, S + U), all classification models are preset on the same manifold, and similarly, R pseudo classes are introduced on the manifold, then each classification model UiWeights w that can be embedded in space by semanticsmiAnd RnA contiguous pseudo class tm(m ═ 1,2, …, R) represents:
Figure BDA0001881092950000033
wherein: u. ofiRepresenting a classification model; t is tmRepresenting a pseudo category; w is amiRepresents each siPseudo class p of its neighborsmThe weight of the connection between.
Optionally, the formula of the objective function is as follows:
Figure BDA0001881092950000034
wherein:
Figure BDA0001881092950000035
pseudo class t representing a learning model spacemAnd weight μ of semantic embedding spacemiMinimizing multi-class classification loss, R representing a total number of false classes, S representing a total number of seen classes, RnNumber of pseudo classes, x, representing the neighbourhood of each class ikRepresents input data, λ represents a weighting coefficient of the classification model, and γ represents a weighting coefficient of the weight;
the first term of the objective function is the squared chained penalty with a one-to-many classification strategy
Figure BDA0001881092950000036
Figure BDA0001881092950000037
Wherein u isiClassifier, x, representing class ikRepresenting input data for ensuring that the classifier is more concerned with gross classification errors, such that gross classification errors tend to be minimal;
indicating function
Figure BDA0001881092950000038
Represents a category ykWhether the same as the category i is detected, if the same as the category i, 1 is selected, and if not, 1 is selected; the second term of the objective function is the classification model uiTo prevent overfitting;
the third target item of the target function is the weight mu of semantic embedding spacemiConstraint is carried out to lead the film to tend to be sparse; | g | represents a modulo operation;
Figure BDA0001881092950000041
representing a norm operation.
Optionally, when the category correlation is considered, the method is further experimented by using a Crammer-Singer multi-category support vector machine strategy, that is, the first term of the objective function is
Figure BDA0001881092950000042
Is replaced by
Figure BDA0001881092950000043
Wherein: i belongs to S- { ykIndicates that category i belongs to the known category set S except for category ykOne of the other categories;
Figure BDA0001881092950000044
representing an optimized class i such that it is associated with class ykThe difference between the semantic embedding space and the model space is maximum; sgSemantic embedding of the representation class g in the semantic embedding space, ugDenotes the classifier, x, corresponding to the class gkRepresenting input data; u. ofiA classifier representing the class i is provided,
Figure BDA0001881092950000045
represents a category ykA classifier ofstructRepresenting Cramer-Singer multi-class losses, s, taking into account class correlationsiThe semantic embedding representing the category i is performed,
Figure BDA0001881092950000046
represents a category ykThe semantic embedding of (2) is carried out,
Figure BDA0001881092950000047
denotes siAnd
Figure BDA0001881092950000048
performing norm operation of the difference value;
optionally, the objective function is solved by using an alternative optimization method.
Compared with the prior art, the invention has the following beneficial effects:
the zero example learning method based on the structural association semantic embedding provided by the invention assumes and utilizes more structural associations in the semantic embedding space through the zero example learning frame, utilizes the semantic representation of similar categories to represent the assumed conditions which can be projected to the adjacent positions in the embedding space, and can be helpful for predicting the classifier of the unseen categories. The invention can keep the global structure of the semantic space while digging the semantic embedded information with similar categories, can obtain effective semantic representation and effectively improve the category prediction accuracy of unseen category image samples.
In the alternative, the present invention takes into account the proximity embedding associations of similar classes, exploits the structured associations and proposes to predict a new classifier by building a locally linearly connected graph structure for unseen classes. The invention reserves the global structure while utilizing the local incidence relation, and obtains more effective semantic information representation by strengthening the influence of adjacent embedding.
Drawings
Other features, objects and advantages of the invention will become more apparent upon reading of the detailed description of non-limiting embodiments with reference to the following drawings:
FIG. 1 is a schematic diagram illustrating a schematic structure of a zero-example learning method based on structured association semantic embedding according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a process of migration of a structural relationship in an embodiment of the present invention.
Detailed Description
The present invention will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the invention, but are not intended to limit the invention in any way. It should be noted that it would be obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit of the invention. All falling within the scope of the present invention.
Fig. 1 is a schematic structural diagram of a principle of a zero-instance learning method based on structured association semantic embedding according to an embodiment of the present invention, as shown in fig. 1, the embodiment includes: constructing a class relation of a manifold space, constructing a structural constraint classification model, and synthesizing an unseen class model, wherein: the category relation construction is that a graph structure retaining a local linear relation is constructed for unseen category samples by mining a structural relation of similar category semantic embedding and considering the local adjacency relation of similar categories; the structured constraint classification model introduces a pseudo category into a semantic space, and projects a graph structure of the semantic space to a model space formed by a known category classifier by means of the pseudo category; and synthesizing the unseen type model by utilizing the projected pseudo type and the weight of the semantic space to reconstruct the image structure in the model space and synthesizing the unseen type model. The process of structured relationship migration is shown in FIG. 2.
The method firstly calculates R embedded in each semantic in the training processnNearest neighbor and form a graph structure with local linear correlation. And then projecting the pseudo categories in the semantic space into a model space. In the prediction phase, the semantics of each unseen category are embedded by its RnAnd (3) expressing the pseudo classes of the nearest neighbors, projecting the pseudo classes to a model space from a semantic embedding space, reserving the weight in the semantic embedding space, and reconstructing the model in the model space to synthesize a new classifier and distinguish unseen classes. Since overlapping local structures can provide global aggregate information, the method preserves global structures in the semantic space while embedding local distributions with semantics. When the local semantic representation is projected into the model space, the global non-linear structure is also recovered from the local connection relations. The optimization objectives of the training process are as follows:
wherein:
Figure BDA0001881092950000051
pseudo class t representing a learning model spacemAnd weight μ of semantic spacemiMinimizing multi-class classification loss, R representing a total number of false classes, S representing a total number of seen classes, RnNumber of pseudo classes, x, representing the nearest neighbor of each class ikRepresents the input data, λ represents the weighting coefficient of the classification model, and γ represents the weighting coefficient of the weight. The first term of the objective function is the square chain loss l (x, y; u) ═ max (0, 1-yu) using a one-to-many classification strategyTx)2Wherein y represents a class label, u represents a classifier, and x represents input data, which is used to ensure that the classifier focuses more on overall classification errors, so that the overall classification errors tend to be minimal; indicating function
Figure BDA0001881092950000052
Represents a category ykWhether the same as the category i is detected, if the same as the category i, 1 is selected, and if not, 1 is selected; the second term of the objective function is the classification model uiTo prevent overfitting; the third target item is the weight mu to semantic spacemiConstraint is carried out to lead the film to tend to be sparse; | g | represents a modulo operation;
Figure BDA0001881092950000053
representing a norm operation. In addition, in consideration of the class correlation, the first term of the objective function is also tested by using a Crammer-Singer multi-class support vector machine, namely
Figure BDA0001881092950000054
Wherein: i belongs to S- { ykIndicates that category i belongs to the known category set S except for category ykOne of the other categories;
Figure BDA0001881092950000061
representing an optimized class i such that it is associated with class ykThe difference is greatest in both the semantic space and the model space. sgSemantic embedding of the representation class g in the semantic space, ugDenotes the classifier, x, corresponding to the class gkRepresenting the input data.
The invention obtains attribute representations at a class level by averaging the attribute representations of images of the same class. All variables are initialized randomly, and the hyper-parameters in the objective function are obtained by performing cross validation on the training data of each data set. The cross-validation strategy in this embodiment employs class-based segmentation of the training data rather than the traditional direct segmentation of the training data. Further, the unconnected edge distance in the constructed graph model is set to ∞. Further, the number of neighboring pseudo classes RnInitially set to a decimal multiple of the number R of pseudo classes and then gradually narrowed to obtain more accurate results depending on the change in classification accuracy.
In this embodiment, training and testing are performed on the AWA data set and the SUN data set, respectively. The AWA dataset is a dataset of animal attributes containing 85 binary attributes and 30475 pictures, for a total of 50 categories. The SUN dataset is a dataset of scene attributes, which contains 102 attributes and 14340 pictures, for a total of 717 scene categories. In this embodiment, the semantic space attribute features adopted by the AWA data set are extracted and obtained by AlexNet, and the semantic space attribute features adopted by the SUN data set are extracted and obtained by GoogleNet. The effectiveness and the generalization of semantic representation extraction are evaluated through the multi-class classification accuracy of untrained class images. The classification accuracy of the zero-example learning method based on structured relevance semantic embedding and the zero-example learning method of the synthetic classifier proposed by Soravit et al in "Synthesized classifiers for zero-shot learning" under two loss strategies of a pair of rest and structured SVM are compared.
The result is: by applying a zero-example learning method of structured associated semantic embedding, the classification accuracy of a test sample under an AWA data set reaches 72.3% and 74.6% respectively under a pair of other loss strategies and structured SVM loss strategies, and is respectively improved by 2.6% and 2.3% compared with a method for synthesizing a classifier model; the classification accuracy of the test sample under the SUN data set respectively reaches 62.4% and 62.7% under a pair of loss strategies of the rest SVM and the structured SVM, and is respectively improved by 0.3% and 0.3% compared with the method for synthesizing the classifier model. Due to the fact that structural association of semantic embedding is considered, the method obtains higher classification accuracy under different loss conditions, and the effectiveness of the method is verified through experimental results.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims (4)

1. A zero example learning method based on structured association semantic embedding is characterized by comprising the following steps:
step 1: extracting attribute features of the known category images;
step 2: establishing a semantic embedding space according to the attribute characteristics, and establishing a model space formed by classification models of known class samples;
and step 3: introducing R pseudo categories into the semantic embedding space and the model space to serve as relationship migration carriers of the semantic embedding space and the model space;
and 4, step 4: acquiring a scaled squared Euclidean distance between a real category and a pseudo category of a known category image, and taking the scaled squared Euclidean distance as a connection weight;
and 5: reconstructing a graph structure of a semantic embedding space in a model space according to the connection weight;
step 6: obtaining R of each real category of the known category image with the connection weight within a preset rangenA pseudo category;
and 7: constructing a set of semantic embedding rules for each semantic embedding in the semantic embedding spacenAn objective function synthesized by pseudo classes;
and 8: constructing a classification model, wherein the classification model consists of contiguous RnSynthesizing the pseudo classes;
and step 9: learning R pseudo classes, adjacent R of the semantic embedding space and the model space according to the objective functionnPseudo classes to minimize loss of the classification model;
step 10: extracting attribute features of an unknown type image, and using adjacent R for the attribute featuresnCharacterizing each pseudo category to obtain a graph structure of the unknown category image;
step 11: projecting the graph structure of the unknown class image from the semantic embedding space to a model space to obtain projected pseudo classes, and synthesizing a classification model of the unknown class image according to the projected pseudo classes;
the step 8 comprises the following steps:
defining the classification model corresponding to each class i as ui(i ═ 1,2, …, S + U), preset all classification modesThe types are on the same manifold, and similarly, R pseudo classes are introduced on the manifold, so that each classification model uiWeights w that can be embedded in space by semanticsmiAnd RnA contiguous pseudo class tmRepresents:
Figure FDA0002729919150000011
wherein: u. ofiRepresenting a classification model; t is tmRepresenting a pseudo category; w is amiRepresents each siPseudo class p of its neighborsmA connection weight between; siSemantic embedding representing category i;
the formula of the objective function is as follows:
Figure FDA0002729919150000021
wherein:
Figure FDA0002729919150000022
pseudo class t representing a learning model spacemAnd weight μ of semantic embedding spacemiMinimizing multi-class classification loss, m 1,2, …, R, R representing the total number of false classes, S representing the total number of seen classes, R representing the total number of false classesnNumber of pseudo classes, x, representing the neighbourhood of each class ikRepresents input data, λ represents a weighting coefficient of the classification model, and γ represents a weighting coefficient of the weight;
the first term of the objective function is the squared chaining penalty with a one-to-many, multi-classification strategy
Figure FDA0002729919150000023
Figure FDA0002729919150000024
Wherein u isiClass model, x, representing class ikRepresenting input data for ensuring that the classification model is more concerned with gross classification errorsThe total classification error tends to be minimal;
indicating function
Figure FDA0002729919150000025
Represents a category ykWhether the same as the category i is detected, if the same as the category i, 1 is selected, and if not, 1 is selected; the second term of the objective function is the classification model uiTo prevent overfitting;
the third target item of the target function is the weight mu of semantic embedding spacemiConstraint is carried out to lead the film to tend to be sparse; | g | represents a modulo operation;
Figure FDA0002729919150000026
representing a norm operation.
2. The zero-example learning method based on structured association semantic embedding according to claim 1, wherein the step 4 comprises:
in a training sample, the number of known classes is S, the number of unknown classes is U, and the number of pseudo classes is R; defining the coordinates of each class i as si(i ═ 1,2, …, S + U), all classes of the pre-set training samples are on the same manifold of the semantic embedding space, with the help of the introduced pseudo-class pm(m-1, 2, …, R) preserving the correlation between training sample attributes, presetting each semantic embedding and adjacent RnPseudo classes are embedded in local linear blocks of the space manifold semantically; defining the scaled squared Euclidean distance as a similarity measure; each siPseudo class p with neighborsmThe calculation formula of the connection weight between the two is as follows:
Figure FDA0002729919150000027
wherein s isiSemantic embedding representing class i, with neighboring pseudo-classes sharing RnA plurality of; p is a radical ofmDenotes siNearest neighbor RnOne of the pseudo categories; w is amiRepresents each siIs close to itNeighbor pseudo class pmWeight of connection between wmiThe larger the value of (A), the more the pseudo class p is representedmEmbedding with semantics siThe greater the similarity between;
embedding R corresponding to neighbor according to each semanticnPseudo categories, namely constructing a category association diagram of a semantic embedding space; the vertex set of the category association graph comprises attributes embedded by semantics and introduced pseudo categories, and the edge set comprises edges connected between the attributes and the pseudo categories; the semantic embedding representation of each pseudo category neighborhood is as follows:
Figure FDA0002729919150000028
wherein, mumiRepresenting the weight between semantic embeddings and pseudo classes.
3. The zero-example learning method based on structured relevance semantic embedding according to claim 1, wherein the method is further experimented with Cramer-Singer multi-class support vector machine strategy, namely, the first term of the objective function, when considering class correlation
Figure FDA0002729919150000031
Is replaced by
Figure FDA0002729919150000032
Wherein: i belongs to S- { ykIndicates that category i belongs to the known category set S except for category ykOne of the other categories;
Figure FDA0002729919150000033
representing an optimized class i such that it is associated with class ykThe difference between the semantic embedding space and the model space is maximum; sgSemantic embedding of the representation class g in the semantic embedding space, ugIndicates the score corresponding to the category gClass device, xkRepresenting input data; u. ofiA classification model representing the class i,
Figure FDA0002729919150000034
represents a category ykA classifier ofstructRepresenting Cramer-Singer multi-class losses, s, taking into account class correlationsiThe semantic embedding representing the category i is performed,
Figure FDA0002729919150000035
represents a category ykThe semantic embedding of (2) is carried out,
Figure FDA0002729919150000036
denotes siAnd
Figure FDA0002729919150000037
and (5) performing norm operation of the difference value.
4. The zero-example learning method based on structured relevance semantic embedding according to claim 1, wherein the objective function is solved by adopting an alternative optimization method.
CN201811423691.0A 2018-11-27 2018-11-27 Zero example learning method based on structured association semantic embedding Active CN109582960B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811423691.0A CN109582960B (en) 2018-11-27 2018-11-27 Zero example learning method based on structured association semantic embedding

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811423691.0A CN109582960B (en) 2018-11-27 2018-11-27 Zero example learning method based on structured association semantic embedding

Publications (2)

Publication Number Publication Date
CN109582960A CN109582960A (en) 2019-04-05
CN109582960B true CN109582960B (en) 2020-11-24

Family

ID=65924782

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811423691.0A Active CN109582960B (en) 2018-11-27 2018-11-27 Zero example learning method based on structured association semantic embedding

Country Status (1)

Country Link
CN (1) CN109582960B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110083778A (en) * 2019-04-08 2019-08-02 清华大学 The figure convolutional neural networks construction method and device of study separation characterization
CN110309875A (en) * 2019-06-28 2019-10-08 哈尔滨工程大学 A kind of zero sample object classification method based on the synthesis of pseudo- sample characteristics
CN112418351B (en) * 2020-12-11 2023-04-07 天津大学 Zero sample learning image classification method based on global and local context sensing
CN112529772A (en) * 2020-12-18 2021-03-19 深圳龙岗智能视听研究院 Unsupervised image conversion method under zero sample setting

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9278255B2 (en) * 2012-12-09 2016-03-08 Arris Enterprises, Inc. System and method for activity recognition
CN105701504B (en) * 2016-01-08 2019-09-13 天津大学 Multi-modal manifold embedding grammar for zero sample learning
US10592788B2 (en) * 2016-12-30 2020-03-17 Hrl Laboratories, Llc Zero-shot learning using multi-scale manifold alignment
EP3593284A4 (en) * 2017-03-06 2021-03-10 Nokia Technologies Oy A transductive and/or adaptive max margin zero-shot learning method and system
CN107563444A (en) * 2017-09-05 2018-01-09 浙江大学 A kind of zero sample image sorting technique and system
CN108399421B (en) * 2018-01-31 2022-04-01 南京邮电大学 Deep zero sample classification method based on word embedding

Also Published As

Publication number Publication date
CN109582960A (en) 2019-04-05

Similar Documents

Publication Publication Date Title
US10229346B1 (en) Learning method, learning device for detecting object using edge image and testing method, testing device using the same
CN109582960B (en) Zero example learning method based on structured association semantic embedding
US10262214B1 (en) Learning method, learning device for detecting lane by using CNN and testing method, testing device using the same
US7983490B1 (en) Adaptive Bayes pattern recognition
US20230316137A1 (en) Data analytics methods for spatial data, and related systems and devices
US10275667B1 (en) Learning method, learning device for detecting lane through lane model and testing method, testing device using the same
CN111369572A (en) Weak supervision semantic segmentation method and device based on image restoration technology
Li et al. Robust deep neural networks for road extraction from remote sensing images
CN111444816A (en) Multi-scale dense pedestrian detection method based on fast RCNN
CN115063664A (en) Model learning method, training method and system for industrial vision detection
Gupta et al. Automated Valve Detection in Piping and Instrumentation (P&ID) Diagrams
Kajabad et al. YOLOv4 for urban object detection: Case of electronic inventory in St. Petersburg
Wang et al. Subspace prototype learning for few-Shot remote sensing scene classification
CN113066537B (en) Compound classification method based on graph neural network
CN113408546B (en) Single-sample target detection method based on mutual global context attention mechanism
Cao et al. A comprehensive understanding of machine learning and deep learning methods for 3d architectural cultural heritage point cloud semantic segmentation
CN112132169B (en) Information processing apparatus and information processing method
CN113627522A (en) Image classification method, device and equipment based on relational network and storage medium
CN112651996A (en) Target detection tracking method and device, electronic equipment and storage medium
Kumari et al. A Deep Learning-Based Segregation of Housing Image Data for Real Estate Application
Kumar et al. Image classification in python using Keras
Mohammad et al. Updating of GIS maps with Change Detection of Buildings using Deep Learning techniques
Bhanbhro et al. Symbol Detection in a Multi-class Dataset Based on Single Line Diagrams using Deep Learning Models
Umapathy et al. Segmentation of Floorplans and Heritage Sites: An Approach to Unbalanced Dataset
Claes Deep Learning for the Classification and Detection of Animals in Artworks

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