CN109492662B - Zero sample image classification method based on confrontation self-encoder model - Google Patents

Zero sample image classification method based on confrontation self-encoder model Download PDF

Info

Publication number
CN109492662B
CN109492662B CN201811134474.XA CN201811134474A CN109492662B CN 109492662 B CN109492662 B CN 109492662B CN 201811134474 A CN201811134474 A CN 201811134474A CN 109492662 B CN109492662 B CN 109492662B
Authority
CN
China
Prior art keywords
visual
features
decoder
encoder
category
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
CN201811134474.XA
Other languages
Chinese (zh)
Other versions
CN109492662A (en
Inventor
冀中
王俊月
于云龙
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.)
Tianjin University
Original Assignee
Tianjin 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 Tianjin University filed Critical Tianjin University
Priority to CN201811134474.XA priority Critical patent/CN109492662B/en
Publication of CN109492662A publication Critical patent/CN109492662A/en
Application granted granted Critical
Publication of CN109492662B publication Critical patent/CN109492662B/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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

A zero sample image classification method based on a confrontation self-encoder model is characterized in that a confrontation self-encoder network trained on visible classes is utilized, network parameters w and v capable of simulating visual feature distribution and enabling visual features to be associated with class semantic features in a best approximation mode are selected, and then class semantic features a of unseen classes are classifiedtThe network is input, the decoder network G is used to generate visual features, and the Euclidean distance between the generated visual features and the real visual features is calculated. And finally, considering the class with the minimum distance as a predicted class, thereby realizing the zero sample classification task. The method provided by the invention better accords with the characteristics of real data, and simultaneously aligns the visual characteristics and the category semantic characteristics, so that a better classification effect can be realized in a zero-sample task.

Description

Zero sample image classification method based on confrontation self-encoder model
Technical Field
The invention relates to a zero sample classification method. In particular to a zero sample classification method based on a confrontation self-encoder model.
Background
Deep learning has greatly facilitated the development of computer vision, such as object classification, image retrieval, and motion recognition. The performance of these tasks is typically evaluated after training with a large amount of annotation data. However, some tasks have only a small portion of training data or even no training data, making traditional classification models less performing. To improve the classification performance of the conventional classification model for classes with little or no data, zero sample learning has attracted extensive attention. The task of Zero sample Learning (Zero Shot Learning) is to classify classes without training data. The human beings have the ability of reasoning, that is to say that the human beings can successfully reason out the class of the unseen object according to the description and the priori knowledge of the object. For example, when a description is given of: the shape of the unicorn is similar to that of a horse, except that the unicorn is added with a long angle, people can immediately recognize the unicorn. Zero sample learning identifies new classes by simulating the reasoning ability of humans. In zero-sample learning, the data is divided into two parts, training data (visible class) and test data (invisible class), and the classes of the two are different. The identification of unseen classes is usually realized by knowledge migration from visible classes to unseen classes, in the process, in order to characterize semantic association between classes, semantic features common to the visible classes and the unseen classes are used, and the commonly used class semantic features include attribute features and text vector features. The attribute features are manually labeled, and the text vector features are obtained by processing a large text corpus by using a natural language technology.
The image is usually represented by visual features, semantic gaps exist between the image and the semantic features, and the image cannot be directly connected with a semantic space. Most of the existing zero sample learning methods comprise two steps, firstly, mapping functions of a visual space and a semantic space are learned, then, the learned mapping functions are used for calculating the similarity between the visual features of the test data and the semantic features of unseen classes, and the classes with larger similarities are taken as labels of the test data.
Compared with human reasoning processes, the methods take semantic features of visible classes as prior knowledge and semantic features of unseen classes as description of objects, but essentially human does not learn the mapping function, but imagines rough outlines of unseen objects in brain for classification. Therefore, we consider that zero-sample learning can simulate human behavior to generate unseen classes of visual features.
The generative countermeasure network (GAN) is a generative model that can learn a particular data distribution. GAN mainly solves the problem of generation, and can generate images by using an arbitrary random number. GAN includes two network models, a generative model g (generator) and a discriminant model d (discriminator). G takes random noise as input to generate an image G (z), then G (z) and a real image x are input into D, and G (z) and x are subjected to two classifications to detect who is the real image and who is the generated false image. G and D will improve themselves continuously according to the output condition of D, G will improve the similarity of G (z) and x to deceive D as much as possible, and D will not be deceived by G as much as possible through learning. G obtains the ability to generate an image when the generated image is not different from the real image, i.e., the output of D is 0.5. When the class information and the noise are input together into G, an image satisfying a certain distribution can be generated to be used in the zero sample method.
In the zero-sample approach, it is generally assumed that N triplets are given in the training phase
Figure GDA0003034668550000021
Data of a defined visible category, where xi∈RpIs a representation of the ith visual feature of the visible category, ai∈RqIs a category semantic feature of the ith visual feature,
Figure GDA0003034668550000022
is the category label for the ith visual feature, and p and q are the dimensions of the visual and semantic spaces, respectively. In the testing stage, according to the category semantic features and category labels of the unseen categories { at,ytFor its visual feature xtIs classified, wherein
Figure GDA0003034668550000023
And is provided with
Figure GDA0003034668550000024
The task of the zero sample is to train a model by using data of visible classes and then predict labels y of unseen classes by using the trained modelt。
Existing methods based on generating classes mainly comprise the following steps:
1) realizing the mapping relation from the category semantic space A to the visual space X through a linear model or a depth model by utilizing a training sample
Figure GDA0003034668550000025
2) And mapping the real category semantic features of the unseen categories to a visual space by using the mapping relation f learned by the training samples to obtain the predicted visual features corresponding to the unseen categories.
3) And determining the category to which the unseen category belongs by using the similarity relation between the visual features obtained by prediction and the real visual features of the unseen category. The discrimination criteria used to determine the class is typically the nearest neighbor method.
However, the method based on generation of classes has the following problems:
when a linear model is used to obtain the mapping relationship from the category semantic space to the visual space, the linear model is likely to cause the loss of some discrimination information of the visible categories in the training stage, but the discrimination information may be included in the unseen categories. When the mapping relation is obtained by using a depth model, a generative countermeasure network is generally used. The confrontation network trains a generator G which can fit the real visual feature distribution by using the confrontation learning between the generator G and the discriminator D. However, most of the countermeasure networks only focus on generating the distribution approximating the real visual features, but ignore the corresponding relationship between the visual features and the category semantic features, so that the generated visual features lack discriminative information to a certain extent.
Disclosure of Invention
The invention aims to solve the technical problem of providing a zero sample image classification method based on a confrontation self-encoder model, which can be more conveniently and accurately applied to image recognition and information retrieval.
The technical scheme adopted by the invention is as follows: a zero sample image classification method based on a confrontation self-encoder model comprises the following steps:
1) initializing parameters r, w and v of discriminator D, encoder E and decoder G;
2) respectively and randomly selecting a group of data in set batches from the visual characteristic x and the category semantic characteristic a of the training sample, and respectively and correspondingly taking the data as the input of an encoder E and a decoder G;
3) training an encoder E and a decoder G according to a self-encoder model, optimizing the model parameters by using an Adam optimizer, and reserving parameters w and v of the encoder E and the decoder G which enable the model calculation result to be minimum:
Figure GDA0003034668550000026
wherein, when the first item represents the input category semantic feature a, the process of obtaining the visual feature through a decoder G; when the second item represents the input category semantic feature a, reconstructing the category semantic feature by a decoder G and an encoder E in sequence;
Figure GDA0003034668550000027
is the corresponding antagonistic self-encoder model parameter regularization term; λ is a parameter corresponding to the regularization term;
Figure GDA0003034668550000031
is expressed by a 2 norm;
4) according to the selected data of the set batch, three inputs x, x' and of the discriminator D are obtained by using the trained encoder E and decoder G
Figure GDA0003034668550000032
Wherein x corresponds to a true visual characteristic; the x' corresponds to the reconstructed visual features, namely the features obtained by the encoder E and the decoder network G in sequence of x belong to real visual features;
Figure GDA0003034668550000033
the correspondingly generated visual features, namely the features obtained by the category semantic features a through a decoder network G, belong to false visual features;
5) the discriminator D is trained on the following model of the discriminator D, the parameters of which are optimized by means of an Adam optimizer, preserving the parameter r that makes the discriminator D perform best:
Figure GDA0003034668550000034
wherein ExAnd eaRespectively representing the distribution of visual features x and category semantic features a, wherein log is logarithm operation, and sigma is a softmax function;
6) training a decoder G according to a model of the discriminator D, optimizing the model parameters by using an Adam optimizer, and reserving a parameter v which enables the decoder G to have the best performance;
7) repeating the steps 2) to 6) according to the set times to obtain final parameters r, w and v;
8) semantic feature a of category of unseen categorytInputting the visual characteristics into a decoder G to obtain the visual characteristics generated by the unseen category
Figure GDA0003034668550000035
9) Comparing visual features generated by unseen categories according to the principle of minimum Euclidean distance
Figure GDA0003034668550000036
And the visual characteristics x of the test specimentThe predicted class label is obtained.
The invention relates to a zero sample image classification method based on a confrontation self-encoder model, which simulates the generation process of visual features and the association between the visual features and category semantic features by utilizing a self-encoder method, better explores the distribution of the visual features, and has the advantages that:
(1) the invention introduces the self-encoder into the counterstudy for the first time, constructs a network structure for generating features bidirectionally, completes the alignment relation between vision and semantics, and designs the zero sample classification technology suitable for the image data features.
(2) The invention can synthesize visual characteristics which are closer to real distribution. The model comprises a countermeasure network, real visual features, reconstructed visual features and generated pseudo visual features are used as input of the discriminator, the reconstructed visual features and the real visual features can be similar as much as possible, therefore, the association of the visual features and the category semantic features can be completed, most semantic information can be reserved, and more real visual features can be synthesized.
Drawings
FIG. 1 is a flow chart of a zero-sample image classification method based on a confrontation self-encoder model according to the present invention.
Detailed Description
The following describes a zero-sample image classification method based on a robust auto-encoder model according to the present invention in detail with reference to the following embodiments and the accompanying drawings.
The invention discloses a zero sample image classification method based on a countermeasure self-encoder model, which assumes that the category semantic features are generated by using the category semantic features, and simultaneously considers the reverse process of generating the category semantic features by the visual features. Therefore, on the basis of using the countermeasure network, the self-encoder is introduced, and the bidirectional generation process is completed through the encoding and decoding processes of the self-encoder, so that the purposes of generating the visual features and associating the visual features with the category semantic features are achieved.
An autoencoder is a type of neural network that is trained to copy an input to an output. The self-encoder consists of two parts, namely an encoder h ═ E (x) and a decoder x ═ G (h), wherein h is used as a middle hidden layer, and x' correspond to input and output. When the dimensionality of x and x' is the same as that of the visual feature and the dimensionality of h is the same as that of the category semantic feature, the purposes of generating the visual feature and associating the visual feature with the category semantic feature can be achieved.
The zero-sample image classification method based on the confrontation self-encoder model is to link visual features and category semantic features through a bidirectional generation process. Specifically, when the input x and the output x' are visual features, the encoder E compresses the visual features x into a hidden space h, and the hidden space h is supervised by the real category semantic features so as to associate the visual features with the category semantic features; the decoder G reconstructs the features of the hidden space to obtain a visual feature x', to obtain:
Figure GDA0003034668550000041
Figure GDA0003034668550000042
where w and v are the parameters of the encoder E and decoder G respectively,
Figure GDA0003034668550000043
is a feature of the hidden space h.
When the input x and the output x' are the category semantic features, the category semantic features directly obtain the generated pseudo-visual features through an encoder E of the category semantic features, and the encoder E is a decoder G used when the input is the visual features; the generated pseudo-visual features are further used for reconstructing the input category semantic features through a decoder G of the pseudo-visual features, and the decoder G corresponds to an encoder E when the input is the visual features.
As shown in FIG. 1, the zero-sample image classification method based on the confrontation self-encoder model of the present invention assumes that x is the visual feature of the training sample, a is the category semantic feature of the training sample, and xtVisual features of the unseen category, atThe category semantic features of the unseen category. The method comprises the following steps:
1) initializing parameters r, w and v of discriminator D, encoder E and decoder G;
2) respectively and randomly selecting a group of data in set batches from the visual characteristic x and the category semantic characteristic a of the training sample, and respectively and correspondingly taking the data as the input of an encoder E and a decoder G;
3) training an encoder E and a decoder G according to a self-encoder model, optimizing the model parameters by using an Adam optimizer, and reserving parameters w and v of the encoder E and the decoder G which enable the model calculation result to be minimum:
Figure GDA0003034668550000044
wherein, when the first item represents the input category semantic feature a, the process of obtaining the visual feature through a decoder G; when the second item represents the semantic feature a of the input category, the second item passes through a decoder G and codes in sequenceA process of reconstructing category semantic features by the device E;
Figure GDA0003034668550000045
is the corresponding antagonistic self-encoder model parameter regularization term; λ is a parameter corresponding to the regularization term;
Figure GDA0003034668550000046
is expressed by a 2 norm;
4) in order to make decoder G obtain better capability of generating visual features, discriminator D is added, and according to the selected set batch of data, the trained encoder E and decoder G are used to obtain three inputs x, x' and
Figure GDA0003034668550000047
wherein x corresponds to a true visual characteristic; the x' corresponds to the reconstructed visual features, namely the features obtained by the encoder E and the decoder network G in sequence of x belong to real visual features;
Figure GDA0003034668550000048
the correspondingly generated visual features, namely the features obtained by the category semantic features a through a decoder network G, belong to false visual features;
5) the discriminator D is trained on the following model of the discriminator D, the parameters of which are optimized by means of an Adam optimizer, preserving the parameter r that makes the discriminator D perform best:
Figure GDA0003034668550000051
wherein ExAnd eaRespectively representing the distribution of visual features x and category semantic features a, wherein log is logarithm operation, and sigma is a softmax function;
6) training a decoder G according to a model of the discriminator D, optimizing the model parameters by using an Adam optimizer, and reserving a parameter v which enables the decoder G to have the best performance;
7) repeating the steps 2) to 6) according to the set times to obtain final parameters r, w and v;
8) semantic feature a of category of unseen categorytInputting the visual characteristics into a decoder G to obtain the visual characteristics generated by the unseen category
Figure GDA0003034668550000052
9) Comparing visual features generated by unseen categories according to the principle of minimum Euclidean distance
Figure GDA0003034668550000053
Visual feature x from unseen categorytThe predicted class label is obtained.
For the zero sample image classification task, for the visual feature x of the unseen classtThe invention utilizes a confrontation self-encoder model which is well trained on visible classes, selects parameters w and v of an encoder E and a decoder G which can best approximately simulate visual feature distribution and enable the visual features to be associated with the class semantic features, and then selects the class semantic features a of unseen classestInputting the visual characteristics into a decoder G, generating the visual characteristics by using the decoder G, and calculating and outputting Euclidean distance between the generated visual characteristics and the real visual characteristics. And finally, considering the class with the minimum distance as a predicted class, thereby realizing the zero sample classification task. The method provided by the invention better accords with the characteristics of real data, simultaneously aligns the visual characteristics and the category semantic characteristics, and can realize a better classification effect in a zero-sample task.

Claims (1)

1. A zero sample image classification method based on a confrontation self-encoder model is characterized by comprising the following steps:
1) initializing parameters r, w and v of discriminator D, encoder E and decoder G;
2) respectively and randomly selecting a group of data in set batches from the visual characteristic x and the category semantic characteristic a of the training sample, and respectively and correspondingly taking the data as the input of an encoder E and a decoder G;
3) training an encoder E and a decoder G according to a self-encoder model, optimizing the model parameters by using an Adam optimizer, and reserving parameters w and v of the encoder E and the decoder G which enable the model calculation result to be minimum:
Figure FDA0003034668540000011
wherein, when the first item represents the input category semantic feature a, the process of obtaining the visual feature through a decoder G; when the second item represents the input category semantic feature a, reconstructing the category semantic feature by a decoder G and an encoder E in sequence;
Figure FDA0003034668540000012
is the corresponding antagonistic self-encoder model parameter regularization term; λ is a parameter corresponding to the regularization term;
Figure FDA0003034668540000013
is expressed by a 2 norm;
4) according to the selected data of the set batch, three inputs x, x' and of the discriminator D are obtained by using the trained encoder E and decoder G
Figure FDA0003034668540000014
Wherein x corresponds to a true visual characteristic; the x' corresponds to the reconstructed visual features, namely the features obtained by the encoder E and the decoder network G in sequence of x belong to real visual features;
Figure FDA0003034668540000015
the correspondingly generated visual features, namely the features obtained by the category semantic features a through a decoder network G, belong to false visual features;
5) the discriminator D is trained on the following model of the discriminator D, the parameters of which are optimized by means of an Adam optimizer, preserving the parameter r that makes the discriminator D perform best:
Figure FDA0003034668540000016
wherein ExAnd eaRespectively representing the distribution of visual features x and category semantic features a, wherein log is logarithm operation, and sigma is a softmax function;
6) training a decoder G according to a model of the discriminator D, optimizing the model parameters by using an Adam optimizer, and reserving a parameter v which enables the decoder G to have the best performance;
7) repeating the steps 2) to 6) according to the set times to obtain final parameters r, w and v;
8) semantic feature a of category of unseen categorytInputting the visual characteristics into a decoder G to obtain the visual characteristics generated by the unseen category
Figure FDA0003034668540000017
9) Comparing visual features generated by unseen categories according to the principle of minimum Euclidean distance
Figure FDA0003034668540000018
Visual feature x from unseen categorytThe predicted class label is obtained.
CN201811134474.XA 2018-09-27 2018-09-27 Zero sample image classification method based on confrontation self-encoder model Active CN109492662B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811134474.XA CN109492662B (en) 2018-09-27 2018-09-27 Zero sample image classification method based on confrontation self-encoder model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811134474.XA CN109492662B (en) 2018-09-27 2018-09-27 Zero sample image classification method based on confrontation self-encoder model

Publications (2)

Publication Number Publication Date
CN109492662A CN109492662A (en) 2019-03-19
CN109492662B true CN109492662B (en) 2021-09-14

Family

ID=65690082

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811134474.XA Active CN109492662B (en) 2018-09-27 2018-09-27 Zero sample image classification method based on confrontation self-encoder model

Country Status (1)

Country Link
CN (1) CN109492662B (en)

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110097095B (en) * 2019-04-15 2022-12-06 天津大学 Zero sample classification method based on multi-view generation countermeasure network
CN110135459B (en) * 2019-04-15 2023-04-07 天津大学 Zero sample classification method based on double-triple depth measurement learning network
CN110427967A (en) * 2019-06-27 2019-11-08 中国矿业大学 The zero sample image classification method based on embedded feature selecting semanteme self-encoding encoder
CN110443293B (en) * 2019-07-25 2023-04-07 天津大学 Zero sample image classification method for generating confrontation network text reconstruction based on double discrimination
CN110580501B (en) * 2019-08-20 2023-04-25 天津大学 Zero sample image classification method based on variational self-coding countermeasure network
CN110598759A (en) * 2019-08-23 2019-12-20 天津大学 Zero sample classification method for generating countermeasure network based on multi-mode fusion
CN110795585B (en) * 2019-11-12 2022-08-09 福州大学 Zero sample image classification system and method based on generation countermeasure network
CN110826638B (en) * 2019-11-12 2023-04-18 福州大学 Zero sample image classification model based on repeated attention network and method thereof
CN111914929B (en) * 2020-07-30 2022-08-23 南京邮电大学 Zero sample learning method
CN112364894B (en) * 2020-10-23 2022-07-08 天津大学 Zero sample image classification method of countermeasure network based on meta-learning
US20230410487A1 (en) * 2020-11-30 2023-12-21 Intel Corporation Online learning method and system for action recognition
CN113191381B (en) * 2020-12-04 2022-10-11 云南大学 Image zero-order classification model based on cross knowledge and classification method thereof
CN112487193B (en) * 2020-12-18 2022-11-22 贵州大学 Zero sample picture classification method based on self-encoder
CN112364851B (en) * 2021-01-13 2021-11-02 北京邮电大学 Automatic modulation recognition method and device, electronic equipment and storage medium
CN112733954A (en) * 2021-01-20 2021-04-30 湖南大学 Abnormal traffic detection method based on generation countermeasure network
CN113111917B (en) * 2021-03-16 2022-07-01 重庆邮电大学 Zero sample image classification method and device based on dual self-encoders
CN113361611B (en) * 2021-06-11 2023-12-12 南京大学 Robust classifier training method under crowdsourcing task
CN113269274B (en) * 2021-06-18 2022-04-19 南昌航空大学 Zero sample identification method and system based on cycle consistency
CN113657172B (en) * 2021-07-20 2023-08-01 西安理工大学 Cross-domain human body action recognition method based on constant characteristics of semantic level field

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778804A (en) * 2016-11-18 2017-05-31 天津大学 The zero sample image sorting technique based on category attribute transfer learning
CN107679556A (en) * 2017-09-18 2018-02-09 天津大学 The zero sample image sorting technique based on variation autocoder
CN107977629A (en) * 2017-12-04 2018-05-01 电子科技大学 A kind of facial image aging synthetic method of feature based separation confrontation network
CN108491874A (en) * 2018-03-19 2018-09-04 天津大学 A kind of image list sorting technique for fighting network based on production
CN108537257A (en) * 2018-03-26 2018-09-14 天津大学 The zero sample classification method based on identification dictionary matrix pair

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778804A (en) * 2016-11-18 2017-05-31 天津大学 The zero sample image sorting technique based on category attribute transfer learning
CN107679556A (en) * 2017-09-18 2018-02-09 天津大学 The zero sample image sorting technique based on variation autocoder
CN107977629A (en) * 2017-12-04 2018-05-01 电子科技大学 A kind of facial image aging synthetic method of feature based separation confrontation network
CN108491874A (en) * 2018-03-19 2018-09-04 天津大学 A kind of image list sorting technique for fighting network based on production
CN108537257A (en) * 2018-03-26 2018-09-14 天津大学 The zero sample classification method based on identification dictionary matrix pair

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Autoencoding beyond pixels using a learned similarity metric;Anders Boesen Lindbo Larsen;《arXiv:1512.09300v2》;20160210;全文 *
Zero-Shot Learning via Class-Conditioned Deep Generative Models;Wenlin Wang,Yunchen Pu,Vinay Kumar Verma;《arXiv:1711.05820v2》;20171119;全文 *
Zero-Shot Learning via Latent Space Encoding;Yunlong Yu,Zhong Ji;《arXiv:1712.09300v2》;20180420;全文 *
基于深度示例差异化的零样本多标签图像分类;冀中，李慧慧，何宇清;《计算机科学与探索》;20180326;第97-105页 *
基于语义属性的零样本图像分类;潘兴会;《中国优秀硕士学位论文全文数据库 信息科技辑》;20180315;第2-4章 *

Also Published As

Publication number Publication date
CN109492662A (en) 2019-03-19

Similar Documents

Publication Publication Date Title
CN109492662B (en) Zero sample image classification method based on confrontation self-encoder model
CN110717431B (en) Fine-grained visual question and answer method combined with multi-view attention mechanism
CN110580501B (en) Zero sample image classification method based on variational self-coding countermeasure network
Mathieu et al. Disentangling factors of variation in deep representation using adversarial training
CN110097095B (en) Zero sample classification method based on multi-view generation countermeasure network
CN110826638A (en) Zero sample image classification model based on repeated attention network and method thereof
CN111652049A (en) Face image processing model training method and device, electronic equipment and storage medium
CN113642621A (en) Zero sample image classification method based on generation countermeasure network
Akhtar et al. Attack to fool and explain deep networks
CN110598759A (en) Zero sample classification method for generating countermeasure network based on multi-mode fusion
CN111242948A (en) Image processing method, image processing device, model training method, model training device, image processing equipment and storage medium
CN112017255A (en) Method for generating food image according to recipe
CN114821196A (en) Zero sample image identification method and identification device, medium and computer terminal thereof
CN113988180A (en) Model fingerprint-based generated image tracing method
Noroozi et al. Seven: deep semi-supervised verification networks
Baraheem et al. Image synthesis: a review of methods, datasets, evaluation metrics, and future outlook
CN111445545B (en) Text transfer mapping method and device, storage medium and electronic equipment
Vijayaraju Image retrieval using image captioning
CN115965836A (en) Human behavior posture video data amplification system and method with controllable semantics
CN112215282B (en) Meta-generalization network system based on small sample image classification
Pallavi Suggestive GAN for supporting Dysgraphic drawing skills
Wu et al. Efficient multi-domain dictionary learning with gans
Santha Deepfakes generation using LSTM based generative adversarial networks
Polyak et al. Unsupervised generation of free-form and parameterized avatars
Kurfess Neural networks and structured knowledge: Rule extraction and applications

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