CN111461067B - Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction - Google Patents

Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction Download PDF

Info

Publication number
CN111461067B
CN111461067B CN202010338879.6A CN202010338879A CN111461067B CN 111461067 B CN111461067 B CN 111461067B CN 202010338879 A CN202010338879 A CN 202010338879A CN 111461067 B CN111461067 B CN 111461067B
Authority
CN
China
Prior art keywords
remote sensing
sensing image
mapping
visible
invisible
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
CN202010338879.6A
Other languages
Chinese (zh)
Other versions
CN111461067A (en
Inventor
李彦胜
孔德宇
张永军
季铮
肖锐
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.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
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 Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN202010338879.6A priority Critical patent/CN111461067B/en
Publication of CN111461067A publication Critical patent/CN111461067A/en
Application granted granted Critical
Publication of CN111461067B publication Critical patent/CN111461067B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]

Landscapes

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

Abstract

The invention provides a zero sample remote sensing image scene recognition method based on prior knowledge mapping and correction. Based on a visible remote sensing image scene sample with a class label and a priori knowledge representation vector set of a visible class, a depth feature extractor and a mapping model from robust visual features to priori knowledge representation features are obtained through remote sensing scene class learning and cross-mode learning between the visual feature vectors and the priori knowledge representation vectors. Based on the category prior knowledge representation vector of the whole category and the invisible remote sensing image scene sample, the prior knowledge representation vector of the invisible category is progressively corrected through unsupervised collaborative representation learning and unsupervised k-nearest neighbor algorithm respectively, so that the classification precision of the zero-sample remote sensing image scene is effectively improved.

Description

Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction
Technical Field
The invention belongs to the technical field of remote sensing and photogrammetry, relates to a zero sample remote sensing image scene classification method, and particularly relates to a zero sample remote sensing image scene identification method based on priori knowledge mapping and correction.
Background
After the 21 st century, the development of remote sensing technology is more rapid, and the remote sensing technology plays an important role in land resource investigation, ecological environment monitoring, disaster analysis and prediction and the like. With the improvement of the resolution of the remote sensing image, the classification method based on the pixels and the objects is widely influenced by the phenomena of 'same-object different spectrum and same-spectrum foreign matter' of the high-resolution remote sensing image, and the requirement for efficient and stable remote sensing image interpretation cannot be met. Based on this consideration, remote sensing image scene classification is widely concerned by researchers at home and abroad. The remote sensing image scene classification aims at predicting the semantic category of an image block by excavating visual primitives and the spatial relationship among the visual primitives in the remote sensing image scene (image block), and can greatly reduce the confusion degree of pixel-level or object-level ground object interpretation, thereby improving the stability and accuracy of high-resolution remote sensing image interpretation, and having important application in the aspects of content-based remote sensing image retrieval, remote sensing image target detection and the like.
With the continuous opening of remote sensing image scene data sets, a large number of remote sensing image scene classification methods based on artificial features or deep learning are provided by multi-field researchers. However, most of the existing remote sensing image scene classification methods rely on all types of remote sensing image samples to learn classification models. With the coming of the remote sensing big data era, the remote sensing ground object categories show an explosive growth trend, so that it is unrealistic to collect sufficient remote sensing image samples for all the categories. How to introduce the priori knowledge in the field of remote sensing into the scene understanding process of the remote sensing image can identify the remote sensing image scene with the class never appearing in the training stage only by learning partial classes containing the remote sensing image, and the method has important practical significance in the era of remote sensing big data. Therefore, the development of Zero-sample learning (Zero-shot learning) in recent years provides a new idea for remote sensing image scene classification. Zero-sample learning aims to simulate the process of human learning, and samples in an invisible class (unseen) are inferentially identified with the aid of class prior knowledge (e.g. attribute vectors of classes, natural language semantic vectors of classes) through visible class (seen) sample learning. Currently, zero sample learning is mainly focused on the field of computer vision, the research on the classification of remote sensing image scenes is few, and a great deal of research work is needed to promote the development of the zero sample remote sensing image scene classification technology.
Disclosure of Invention
The invention provides a zero-sample remote sensing image scene recognition method based on priori knowledge mapping and correction, which is based on the problems of large modal span between a bottom remote sensing image scene sample and high-level priori knowledge representation, drift of a visible class priori knowledge space and an invisible class priori knowledge space, and offset of an invisible class priori knowledge representation space generated by remote sensing image scene mapping and an invisible class semantic space corrected based on the visible class priori knowledge space. Based on a visible remote sensing image scene sample with a class label and a priori knowledge representation vector set of a visible class, a depth feature extractor and a mapping model from robust visual features to priori knowledge representation features are obtained through remote sensing scene class learning and cross-mode learning between the visual feature vectors and the priori knowledge representation vectors. Based on the category prior knowledge representation vector of the whole category and the invisible remote sensing image scene sample, the prior knowledge representation vector of the invisible category is progressively corrected through unsupervised collaborative representation learning and unsupervised k-nearest neighbor algorithm respectively, so that the classification precision of the zero-sample remote sensing image scene is effectively improved.
The technical scheme adopted by the invention is as follows: a zero sample remote sensing image scene recognition method based on priori knowledge mapping and correction comprises the following steps:
a training stage:
step 1: creating a priori knowledge representation vector corresponding to each category of visible classes based on open natural language corpus or domain expert knowledge
Figure BDA0002467631140000021
Vector of prior knowledge representation corresponding to each category of invisible classes
Figure BDA0002467631140000022
Where p and q represent the number of classes, visible and invisible, respectively, dsRepresenting the dimensionality of the vector for a priori knowledge;
and 2, step: input original remote sensing image scene data set D { (x)i，yi)：i＝1，...，M}，
Figure BDA0002467631140000023
Where D is a visible class data set, xiRepresenting in visible classesIth remote sensing image scene, yiRepresenting a category label of the ith image in the visible category, wherein M is the total number of samples of the visible remote sensing data; dUIn the case of a data set of the invisible class,
Figure BDA0002467631140000024
representing the kth remote sensing image scene in the invisible class,
Figure BDA0002467631140000025
a category label of the kth image in the invisible category is represented, and N is the total number of samples of the invisible category data;
Figure BDA0002467631140000026
extracting image characteristics F of visible class data set and image characteristics F of invisible class data set by utilizing deep convolutional networkU；
And step 3: solving a mapping matrix W from F to S based on a robust cross-modal mapping target function of visual feature self-coding constraint, and thus finishing the learning of depth cross-modal mapping;
and 4, step 4: correction based on unsupervised collaborative representation learningUTo obtain
Figure BDA0002467631140000027
And 5: using mapping matrix W in step 3 to convert FUMapping to
Figure BDA0002467631140000028
Step 6: solving by using k nearest neighbor algorithm
Figure BDA0002467631140000031
Semantic vector obtained through mapping
Figure BDA0002467631140000032
The neighbor vectors in (1) are averaged to obtain
Figure BDA0002467631140000033
And (3) a testing stage:
and 7: giving an invisible test remote sensing image scene, extracting visual features and mapping to obtain semantic vectors according to the steps 2-5
Figure BDA0002467631140000034
And step 8: computing
Figure BDA0002467631140000035
And
Figure BDA0002467631140000036
cosine similarity between the two images is obtained, and a label of a test remote sensing image scene is obtained.
Furthermore, in the step 2, T is used for representing the convolutional layer hyper-parameter of the deep convolutional network, and V is the mapping hyper-parameter of the last fully-connected layer feature and the classification layer; learning the hyper-parameter T of the convolution layer and the hyper-parameter V of the mapping of the full connection layer by fine tuning the deep convolution network, and extracting the image characteristics of the visible data set by utilizing the hyper-parameter T of the convolution layer
Figure BDA0002467631140000037
The fine tuning deep network process only uses visible data; wherein f isi＝Q(xi(ii) a T), Q (; a) represents a non-linear mapping of a deep convolutional network, the deep convolutional network based on the remote sensing image scene data set optimizes a loss function as in equation one, wherein ci＝σ(fiV), σ () denotes the Softmax map,
Figure BDA0002467631140000038
wherein M is the total number of samples of the visible remote sensing data, and p represents the number of categories of the visible remote sensing data.
Further, the mapping matrix W in step 3 is obtained by the self-encoder, and the objective function is as follows:
Figure BDA0002467631140000039
where alpha is a self-encoding regularization coefficient,
Figure BDA00024676311400000310
denotes the F norm, s denotes the sum of FiAnd (3) simplifying a corresponding priori knowledge semantic vector into a Sylvester equation, and solving W by using a Bartels-Stewart algorithm.
Further, the objective function of the collaborative representation coefficient ρ in unsupervised collaborative representation learning in step 4 is:
Figure BDA00024676311400000311
where β is the regularization constant, the closed form solution of the above equation:
Figure BDA00024676311400000312
wherein, I is a discrimination matrix, and the optimal co-expression coefficient is obtained by formula
Figure BDA00024676311400000313
Performing matrix operation with S to obtain reconstructed invisible semantic vector
Figure BDA00024676311400000314
Figure BDA00024676311400000315
Further, in step 5
Figure BDA00024676311400000316
Calculated as follows:
Figure BDA0002467631140000041
further, in step 6
Figure BDA0002467631140000042
Calculated as follows:
Figure BDA0002467631140000043
wherein the content of the first and second substances,
Figure BDA0002467631140000044
to represent
Figure BDA0002467631140000045
The k-th invisible class prior knowledge of the medium represents that the vector is in
Figure BDA0002467631140000046
The m neighbor prior knowledge searched in (1) represents a vector, k is 1 … q, and o is 1 … m.
Further, the label of the invisible type test remote sensing image scene in step 8 is calculated according to the following formula:
Figure BDA0002467631140000047
specifically, a set of test remote sensing image scenes is given
Figure BDA0002467631140000048
Visual features of remote sensing scene images
Figure BDA0002467631140000049
Further mapping it into semantic vector with matrix W
Figure BDA00024676311400000410
Calculating out
Figure BDA00024676311400000411
And
Figure BDA00024676311400000412
cosine similarity between them, wherein,
Figure BDA00024676311400000413
is an image of a scene
Figure BDA00024676311400000414
D (-) is the cosine distance equation.
The invention has the following advantages: the invention aims at the problems of mapping learning and reference correction of prior knowledge in a remote sensing scene zero sample classification task. Based on the category prior knowledge representation vector of the visible category and the remote sensing image scene sample, the depth cross-modal mapping from the visual space of the remote sensing image scene to the category prior knowledge representation space is realized by combining the scene category classification and the multitask learning of self-coding cross-modal mapping. Aiming at the offset problem of a visible prior knowledge representation space and an invisible prior knowledge representation space and the offset problem of an invisible prior knowledge representation space after self-coding cross-modal mapping model mapping and the offset problem of an invisible prior knowledge representation space after collaborative representation, the invention corrects the prior knowledge representation vector of the invisible category through unsupervised collaborative representation learning and unsupervised k-nearest neighbor algorithm respectively based on category prior knowledge representation vectors of all categories and invisible remote sensing image samples, and realizes a stable invisible remote sensing image scene recognition task.
Drawings
FIG. 1: is a general flow diagram of an embodiment of the invention;
FIG. 2: is a sample diagram of a data set according to an embodiment of the present invention.
Detailed Description
In order to facilitate the understanding and implementation of the present invention for those of ordinary skill in the art, the present invention is further described in detail with reference to the accompanying drawings and examples, it is to be understood that the embodiments described herein are merely illustrative and explanatory of the present invention and are not restrictive thereof.
Referring to fig. 1, the method for identifying the zero-sample remote sensing image scene based on prior knowledge mapping and correction provided by the invention comprises the following steps:
step 1: creating a priori knowledge representation vector corresponding to each category of visible classes based on open natural language corpus or domain expert knowledge
Figure BDA0002467631140000051
Vector of prior knowledge representation corresponding to each category of invisible classes
Figure BDA0002467631140000052
Where p and q represent the number of classes, visible and invisible, respectively, dsIs a semantic vector dimension.
Step 2: input original remote sensing image scene data set D { (x)i，yi)：i＝1，...，M}，
Figure BDA0002467631140000053
Extracting image characteristics F of visible class data set and image characteristics F of invisible class data set by utilizing deep convolutional networkU；
Let T denote the convolutional layer hyper-parameter of the deep convolutional network Resnet-50, and V is the mapping hyper-parameter of the last fully-connected layer feature f and the classification layer y. And learning the convolutional layer hyper-parameter T and the full connection layer mapping hyper-parameter V through a fine tuning deep convolutional network. The network optimization loss function based on the remote sensing image scene data set is as the formula I, wherein ci＝σ(fiV), σ () denotes the Softmax mapping, fi＝Q(xi(ii) a T), Q (; a.) represents a non-linear mapping of the deep convolutional network.
Figure BDA0002467631140000054
And learning the convolutional layer hyper-parameter T and the full-connection layer mapping hyper-parameter V through a fine tuning deep convolutional network. Extracting image features of visible class data set by using parameter T
Figure BDA0002467631140000055
Extracting image features of visible class datasets
Figure BDA0002467631140000056
D is a visible class data set, xiRepresenting the ith remote-sensing image scene in the visible class, yiRepresenting a category label of the ith image in the visible category, wherein M is the total number of samples of the visible remote sensing data; dUIn the case of a data set of the invisible class,
Figure BDA0002467631140000057
representing the ith remote sensing image scene in the invisible class,
Figure BDA0002467631140000058
a category label representing the ith image in the invisible category, wherein N is the total number of samples of the invisible category data;
Figure BDA0002467631140000059
the fine tuning depth network process only uses visible remote sensing image scene samples.
And step 3: and solving a mapping matrix W from F to S. The mapping matrix W is obtained by the self-encoder, and the objective function is as follows:
Figure BDA00024676311400000510
wherein, alpha is a regularization coefficient of self-coding, and the optimal value is 0.001 through experimental analysis.
Figure BDA0002467631140000061
Denotes the F norm, s denotes the sum of FiThe corresponding priori knowledge semantic vector, formula one, can be simplified into Sylvester equation, and the W is solved by using Bartels-Stewart algorithm.
And 4, step 4: correcting S with collaborative representationUTo obtain
Figure BDA0002467631140000062
The objective function of the co-expression coefficient ρ is:
Figure BDA0002467631140000063
where β is the regularization constant. The closed-form solution of the above formula is:
Figure BDA0002467631140000064
wherein, I is a discrimination matrix. Optimal co-expression coefficient obtained by formula III
Figure BDA0002467631140000065
Performing matrix operation with S to obtain reconstructed invisible semantic vector
Figure BDA0002467631140000066
Figure BDA0002467631140000067
And 5: using mapping matrix W in step 3 to convert FUMapping to
Figure BDA0002467631140000068
Figure BDA0002467631140000069
Calculated as follows:
Figure BDA00024676311400000610
step 6: solving by using k nearest neighbor algorithm
Figure BDA00024676311400000611
In the process of passingMapped a priori knowledge representative vector
Figure BDA00024676311400000612
The neighbor vectors in (1) are averaged to obtain
Figure BDA00024676311400000613
Wherein
Figure BDA00024676311400000614
Calculated as follows:
Figure BDA00024676311400000615
Figure BDA00024676311400000616
to represent
Figure BDA00024676311400000617
The j-th invisible class prior knowledge represents the vector is in
Figure BDA00024676311400000618
The m neighbor prior knowledge sought in (1) represents a vector.
And 7: giving an invisible image, extracting visual features and mapping to obtain a priori knowledge expression vector
Figure BDA00024676311400000619
And 8: computing
Figure BDA00024676311400000620
And
Figure BDA00024676311400000621
the cosine similarity between the two images, and the label of the test image is predicted. The label of the invisible type test image can be calculated as follows:
Figure BDA00024676311400000622
specifically, a set of test remote sensing scene images is given
Figure BDA00024676311400000623
Visual features of remote sensing scene images
Figure BDA00024676311400000624
Further mapping it to a priori knowledge representation vector by a matrix W
Figure BDA00024676311400000625
Computing
Figure BDA00024676311400000626
And
Figure BDA00024676311400000627
cosine similarity between them, wherein,
Figure BDA00024676311400000628
is an image of a scene
Figure BDA00024676311400000629
D (-) is the cosine distance equation.
In order to verify the effectiveness of the disclosed technology, a plurality of existing disclosed remote sensing image scene data sets are integrated, and a remote sensing image scene data set with more scene types is established. Based on the natural language models Word2vec and Bert, two types of class prior knowledge expression vectors are created for each class of the newly constructed remote sensing scene data set. Based on two different classification prior knowledge representation methods, experimental results show that the algorithm disclosed by the invention can obtain ideal classification precision under the condition of dividing multiple different visible classes and invisible classes.
The described method has been an evaluation test on a new data set obtained by integrating public data sets, which may reflect the effectiveness of the method. Specifically, a public evaluation dataset is shown in fig. 2, which comprises 70 classes of scene categories, each class comprising 800 images. Table 1 shows the results of vector testing using two prior knowledge of Word2vec and Bert under different partitioning modes for the visible class and the invisible class.
Table 1. visible class and invisible class are divided into two kinds of prior knowledge expression vectors of Word2vec and Bert according to different proportions, and the overall accuracy of the method on the test data set is measured
Figure BDA0002467631140000071
It should be understood that parts of the specification not set forth in detail are well within the prior art.
It should be understood that the above description of the preferred embodiments is given for clarity and not for any purpose of limitation, and that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (7)

1. A zero sample remote sensing image scene recognition method based on prior knowledge mapping and correction is characterized by comprising the following steps:
a training stage:
step 1: creating a priori knowledge representation vector corresponding to each category of visible classes based on open natural language corpus or domain expert knowledge
Figure FDA0003617033620000011
Vector of prior knowledge representation corresponding to each category of invisible classes
Figure FDA0003617033620000012
Where p and q represent the number of classes, visible and invisible, respectively, dsRepresenting the dimensionality of the vector for a priori knowledge;
step 2: input original remote sensing image scene data set D { (x)i，yi)：i＝1，…，M}，
Figure FDA0003617033620000013
Where D is a visible class data set, xiRepresenting the ith remote-sensing image scene in the visible class, yiRepresenting a category label of the ith image in the visible category, wherein M is the total number of samples of the visible remote sensing data; dUIn the case of a data set of the invisible class,
Figure FDA0003617033620000014
representing the kth remote sensing image scene in the invisible class,
Figure FDA0003617033620000015
a category label representing the kth image in the invisible category, wherein N is the total number of samples of the invisible category data;
Figure FDA0003617033620000016
extracting image characteristics F of visible class data set and image characteristics F of invisible class data set by utilizing deep convolutional networkU；
And step 3: solving a mapping matrix W from F to S based on a robust cross-modal mapping target function of visual feature self-coding constraint, and thus finishing the learning of depth cross-modal mapping;
and 4, step 4: correction based on unsupervised collaborative representation learningUTo obtain
Figure FDA0003617033620000017
And 5: using mapping matrix W in step 3 to convert FUMapping to
Figure FDA0003617033620000018
Step 6:solving by using k nearest neighbor algorithm
Figure FDA0003617033620000019
Semantic vector obtained through mapping
Figure FDA00036170336200000110
The neighbor vectors in (1) are averaged to obtain
Figure FDA00036170336200000111
And (3) a testing stage:
and 7: giving an invisible test remote sensing image scene, extracting visual features and mapping to obtain semantic vectors according to the steps 2-5
Figure FDA00036170336200000112
And 8: computing
Figure FDA00036170336200000113
And
Figure FDA00036170336200000114
cosine similarity between the two images is obtained, and a label of a test remote sensing image scene is obtained.
2. The method for identifying the scene of the zero-sample remote sensing image mapped and corrected based on the prior knowledge as claimed in claim 1, wherein: in the step 2, T is used for representing the convolutional layer hyper-parameter of the deep convolutional network, and V is the mapping hyper-parameter of the last fully-connected layer feature and the classification layer; learning the hyper-parameter T of the convolution layer and the hyper-parameter V of the mapping of the full connection layer by fine tuning the deep convolution network, and extracting the image characteristics of the visible data set by utilizing the hyper-parameter T of the convolution layer
Figure FDA0003617033620000021
The fine tuning deep network process only uses visible data; wherein f isi＝Q(xi；T)，Q represents the nonlinear mapping of the depth convolution network, the optimization loss function of the depth convolution network based on the remote sensing image scene data set is shown as the formula I, wherein ci＝σ(fiV), σ () denotes the Softmax map,
Figure FDA0003617033620000022
wherein x isiRepresenting the ith remote-sensing image scene in the visible class, yiClass label indicating the ith image in the visible class, dfAnd representing the dimension of the characteristic, wherein M is the total number of samples of the visible remote sensing data, and p represents the number of categories of the visible category.
3. The zero-sample remote sensing image scene recognition method based on priori knowledge mapping and correction according to claim 2, wherein: the mapping matrix W in the step 3 is obtained through a self-encoder, and the objective function is as follows:
Figure FDA0003617033620000023
where alpha is a self-encoding regularization coefficient,
Figure FDA0003617033620000024
denotes the F norm, s denotes the sum of FiAnd (3) simplifying a corresponding priori knowledge semantic vector into a Sylvester equation, and solving W by using a Bartels-Stewart algorithm.
4. The method for identifying the scene of the zero-sample remote sensing image mapped and corrected based on the prior knowledge as claimed in claim 1, wherein: the objective function of the collaborative representation coefficient ρ in unsupervised collaborative representation learning in step 4 is:
Figure FDA0003617033620000025
where β is the regularization constant, the closed form solution of the above equation:
Figure FDA0003617033620000026
wherein, I is a discrimination matrix, and the optimal co-expression coefficient is obtained by formula
Figure FDA0003617033620000027
Performing matrix operation with S to obtain reconstructed invisible semantic vector
Figure FDA0003617033620000028
Figure FDA0003617033620000029
5. The method for identifying the scene of the zero-sample remote sensing image mapped and corrected based on the prior knowledge as claimed in claim 1, wherein: in step 5
Figure FDA00036170336200000210
Calculated as follows:
Figure FDA0003617033620000031
6. the method for zero-sample remote sensing image scene recognition based on priori knowledge mapping and correction of claim 3, wherein: in step 6
Figure FDA0003617033620000032
Calculated as follows:
Figure FDA0003617033620000033
wherein the content of the first and second substances,
Figure FDA0003617033620000034
to represent
Figure FDA0003617033620000035
The k-th invisible class prior knowledge of the medium represents that the vector is in
Figure FDA0003617033620000036
The m neighbor prior knowledge searched in (1) represents a vector, k is 1 … q, and o is 1 … m.
7. The method for zero-sample remote sensing image scene recognition based on priori knowledge mapping and correction of claim 6, wherein: the label of the invisible test remote sensing image scene in the step 8 is calculated according to the following formula:
Figure FDA0003617033620000037
specifically, a set of test remote sensing image scenes is given
Figure FDA0003617033620000038
Visual features of remote sensing scene images
Figure FDA0003617033620000039
Further mapping it into semantic vector with matrix W
Figure FDA00036170336200000310
Calculating out
Figure FDA00036170336200000311
And
Figure FDA00036170336200000315
cosine similarity between them, wherein,
Figure FDA00036170336200000313
is an image of a scene
Figure FDA00036170336200000314
D (-) is the cosine distance equation.
CN202010338879.6A 2020-04-26 2020-04-26 Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction Active CN111461067B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010338879.6A CN111461067B (en) 2020-04-26 2020-04-26 Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010338879.6A CN111461067B (en) 2020-04-26 2020-04-26 Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction

Publications (2)

Publication Number Publication Date
CN111461067A CN111461067A (en) 2020-07-28
CN111461067B true CN111461067B (en) 2022-06-14

Family

ID=71686040

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010338879.6A Active CN111461067B (en) 2020-04-26 2020-04-26 Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction

Country Status (1)

Country Link
CN (1) CN111461067B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112070023B (en) * 2020-09-09 2022-08-16 郑州轻工业大学 Neighborhood prior embedded type collaborative representation mode identification method
CN115100532B (en) * 2022-08-02 2023-04-07 北京卫星信息工程研究所 Small sample remote sensing image target detection method and system
CN115018472B (en) * 2022-08-03 2022-11-11 中国电子科技集团公司第五十四研究所 Interactive incremental information analysis system based on interpretable mechanism

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102054274A (en) * 2010-12-01 2011-05-11 南京大学 Method for full automatic extraction of water remote sensing information in coastal zone
CN109558890A (en) * 2018-09-30 2019-04-02 天津大学 Zero sample image classification method of confrontation network is recycled based on adaptive weighting Hash
CN110334781A (en) * 2019-06-10 2019-10-15 大连理工大学 A kind of zero sample learning algorithm based on Res-Gan
CN110728187A (en) * 2019-09-09 2020-01-24 武汉大学 Remote sensing image scene classification method based on fault tolerance deep learning

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150310862A1 (en) * 2014-04-24 2015-10-29 Microsoft Corporation Deep learning for semantic parsing including semantic utterance classification

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102054274A (en) * 2010-12-01 2011-05-11 南京大学 Method for full automatic extraction of water remote sensing information in coastal zone
CN109558890A (en) * 2018-09-30 2019-04-02 天津大学 Zero sample image classification method of confrontation network is recycled based on adaptive weighting Hash
CN110334781A (en) * 2019-06-10 2019-10-15 大连理工大学 A kind of zero sample learning algorithm based on Res-Gan
CN110728187A (en) * 2019-09-09 2020-01-24 武汉大学 Remote sensing image scene classification method based on fault tolerance deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Zero-Shot Scene Classification for High Spatial Resolution Remote Sensing Images;Aoxue Li 等;《IEEE Transactions on Geoscience and Remote Sensing》;20170731;全文 *
结合深度学习和半监督学习的遥感影像分类进展;谭琨 等;《中国图象图形学报》;20191130;第24卷(第11期);全文 *

Also Published As

Publication number Publication date
CN111461067A (en) 2020-07-28

Similar Documents

Publication Publication Date Title
CN111461067B (en) Zero sample remote sensing image scene identification method based on priori knowledge mapping and correction
Zhu et al. Intelligent logging lithological interpretation with convolution neural networks
CN110516095B (en) Semantic migration-based weak supervision deep hash social image retrieval method and system
CN109871875B (en) Building change detection method based on deep learning
CN112232371B (en) American license plate recognition method based on YOLOv3 and text recognition
CN113223042B (en) Intelligent acquisition method and equipment for remote sensing image deep learning sample
CN115934990B (en) Remote sensing image recommendation method based on content understanding
CN110929746A (en) Electronic file title positioning, extracting and classifying method based on deep neural network
CN108805102A (en) A kind of video caption detection and recognition methods and system based on deep learning
CN112149758A (en) Hyperspectral open set classification method based on Euclidean distance and deep learning
CN113988147A (en) Multi-label classification method and device for remote sensing image scene based on graph network, and multi-label retrieval method and device
CN115049841A (en) Depth unsupervised multistep anti-domain self-adaptive high-resolution SAR image surface feature extraction method
CN115393666A (en) Small sample expansion method and system based on prototype completion in image classification
CN115546553A (en) Zero sample classification method based on dynamic feature extraction and attribute correction
CN117572457B (en) Cross-scene multispectral point cloud classification method based on pseudo tag learning
CN109002771A (en) A kind of Classifying Method in Remote Sensing Image based on recurrent neural network
CN114579794A (en) Multi-scale fusion landmark image retrieval method and system based on feature consistency suggestion
CN113269274A (en) Zero sample identification method and system based on cycle consistency
CN114511787A (en) Neural network-based remote sensing image ground feature information generation method and system
CN117315556A (en) Improved Vision Transformer insect fine grain identification method
CN115497006B (en) Urban remote sensing image change depth monitoring method and system based on dynamic mixing strategy
CN111652265A (en) Robust semi-supervised sparse feature selection method based on self-adjusting graph
CN115496950A (en) Neighborhood information embedded semi-supervised discrimination dictionary pair learning image classification method
CN115511214A (en) Multi-scale sample unevenness-based mineral product prediction method and system
CN114708501A (en) Remote sensing image building change detection method based on condition countermeasure network

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