CN109359623A - High spectrum image based on depth Joint Distribution adaptation network migrates classification method - Google Patents
High spectrum image based on depth Joint Distribution adaptation network migrates classification method Download PDFInfo
- Publication number
- CN109359623A CN109359623A CN201811347067.7A CN201811347067A CN109359623A CN 109359623 A CN109359623 A CN 109359623A CN 201811347067 A CN201811347067 A CN 201811347067A CN 109359623 A CN109359623 A CN 109359623A
- Authority
- CN
- China
- Prior art keywords
- spectrum image
- high spectrum
- aiming field
- probability distribution
- source domain
- 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.)
- Granted
Links
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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
Abstract
A kind of high spectrum image migration classification method based on depth Joint Distribution adaptation network, step are as follows: the high spectrum image of input source domain and aiming field carries out feature normalization and dimension is unified；Combine the feature of source domain and aiming field high spectrum image；Marginal probability distribution adaptation network is constructed, the marginal probability distribution adaptation of source domain and aiming field high spectrum image is carried out；According to one-to-many principle of classification, source domain and the training sample with a small amount of aiming field high spectrum image are chosen；Conditional probability distribution adaptation network is constructed, the conditional probability distribution adaptation of source domain and aiming field high spectrum image is carried out；One-to-many classification is carried out to aiming field high spectrum image.The invention proposes depth Joint Distribution adaptation network is based on, the feature adaptation of source domain and aiming field high spectrum image is realized, reduces the joint probability distribution difference of the two；Meanwhile using one-to-many disaggregated model, the distinction in class between class is improved, and then improves the accuracy of high spectrum image migration classification.
Description
Technical field
The invention belongs to technical field of remote sensing image processing, are based on depth Joint Distribution adaptation network more particularly to one kind
High spectrum image migrate classification.
Background technique
The ground object target spectral details abundant information that high spectrum image spectral resolution is high, band coverage is wide, obtains,
Be conducive to carry out fine terrain analysis.Classification hyperspectral imagery is the important content of high spectrum image interpretation, in mineral exploration, is planted
It is investigated, the fields such as agricultural monitoring are widely used.Since hyperspectral image data amount is big, there are redundancies, same target is not
With in data, there are SPECTRAL DIVERSITYs, influence the effect of classification.
Classification hyperspectral imagery is the process that the features such as the target optical spectrum signal that analysis obtains are classified, and main use has
The method of supervised classification.Supervised classification model is required to a large amount of training sample, could obtain more excellent classification results.
In general, training sample is to need biggish cost of labor by manually marking acquisition.In practical applications, new remote sensing number
According to marker samples be often not easy to obtain.In order to realize automatic classification, new remotely-sensed data should be used by the training of other data
Good classifier is classified.However, during classification hyperspectral imagery, not due to factors such as sensor, wave band coverings
Together, the Data Migration ability of classifier can all be influenced.The classifier nicety of grading on the data of the training on a data set
It is higher, but be difficult to obtain identical effect in other data.
For the problem that Data Migration ability in high spectrum image migration classification is limited, the domain based on transfer learning is adapted to mould
Type is suggested, to reduce the influence of different sensors data difference.For example, Persello C. et al. 2016 in IEEE
" the Kernel- delivered on the 5th phase of volume 54 of Transactions on Geoscience and Remote Sensing
based domain-invariant feature selection in hyperspectral images for transfer
Learning ", the domain invariant features selection method based on core study is proposed, by calculating source in reproducing kernel Hilbert space
The distance in domain and aiming field conditional probability distribution, to measure the stability of Data Migration, the experiment proves that this method feature selecting
Validity.Zhou X. et al. 2018 in IEEE Transactions on Geoscience and Remote
" the Deep feature alignment neural networks for delivered on the 10th phase of volume 56 of Sensing
Domain adaptation of hyperspectral data ", depth convolution loop neural network is proposed, source domain is carried out
It is adapted to the feature learning of aiming field high spectrum image and domain, the experiment proves that this method can be improved using the information of source domain
The precision of aiming field classification hyperspectral imagery.The above method does not analyze the joint probability point of source domain and aiming field high spectrum image
Cloth, the joint probability distribution not yet in effect using data further increase migration classifying quality.
Summary of the invention
The present invention is intended to provide the model of more preferable migration classifying quality can be obtained, propose a kind of based on depth Joint Distribution
The high spectrum image of adaptation network migrates classification method.
Technical solution of the present invention:
High spectrum image based on depth Joint Distribution adaptation network migrates classification method, and steps are as follows:
(1) high spectrum image for inputting source domain and aiming field carries out feature normalization and dimension is unified:
(1a) is respectively normalized the spectral signature of two panel height spectrum pictures of source domain and aiming field, is distributed in it
Between 0 and 1；
(1b) if the spectral Dimensions of two panel height spectrum pictures are different, the high spectrum image low to dimension carries out zero padding, makes
The image high with dimension realize that dimension is unified；
(2) feature of source domain and aiming field high spectrum image is combined:
(2a) carries out vectorization to the spectral signature of source domain and aiming field high spectrum image respectively；
The spectral signature of source domain and aiming field is combined into a vector set by (2b)；
(3) three layers of marginal probability distribution adaptation network are constructed, the marginal probability of source domain and aiming field high spectrum image is carried out
Distribution adaptation:
(3a) utilizes a linear denoising encoder and a non-linear encoder, and building first layer marginal probability distribution is suitable
Distribution network, solves the weight of linear denoising encoder, and is adapted to source domain and aiming field spectral signature；
(3b) utilizes a linear denoising encoder and a non-linear encoder, and building second layer marginal probability distribution is suitable
Distribution network, solves the weight of linear denoising encoder, and is adapted to source domain and aiming field spectral signature；
(3c) utilizes a linear denoising encoder and a non-linear encoder, and building third layer marginal probability distribution is suitable
Distribution network, solves the weight of linear denoising encoder, and is adapted to source domain and aiming field spectral signature；
(4) according to one-to-many principle of classification, the training sample of source domain and aiming field high spectrum image is chosen:
Classify to C class, the aiming field high spectrum image sample of source domain high spectrum image sample and p% is chosen, according to a pair
More principles of classification form C training sample set, and each training set sample label is to belong to such and be not belonging to such, sample characteristics
For the feature by the output of marginal probability distribution adaptation network；
(5) C three layers of conditional probability distribution adaptation network of building, the condition for carrying out source domain and aiming field high spectrum image are general
Rate distribution adaptation:
(5a) initializes the weight and offset parameter of C three layers of conditional probability distribution adaptation networks respectively, utilizes each instruction
Practice sample set and distinguishes layer-by-layer C network of pre-training；
After (5b) pre-training is complete, the implicit output of third layer conditional probability distribution adaptation network is special as the sample after optimization
Sign；
In (5c) each network, softmax classifier is connected after third layer conditional probability distribution adaptation network, will be trained
Feature and class label after sample optimization are input to classifier, optimize the weight and offset parameter of softmax classifier；
(5d) is finely adjusted C conditional probability distribution adaptation network by top layer to bottom, advanced optimizes each network
Parameter；
(6) one-to-many classification is carried out to aiming field high spectrum image:
The test sample collection of aiming field high spectrum image is separately input to C conditional probability distribution adaptation network by (6a),
It is characterized in the implicit output of third layer conditional probability distribution adaptation network after the network optimization；
(6b) classifies to test sample using trained C softmax classifier, is belonged to and is not belonging to each
The prediction probability of a classification；
(6c) compares the size for belonging to each class prediction probability, takes the corresponding classification of maximum probability as each sample
Prediction label；
(6d) exports the classification knot of aiming field high spectrum image according to the spatial position of prediction label vector sum test sample
Fruit figure.
Compared with prior art, the present invention mainly having the advantage that
First, the invention proposes depth Joint Distribution adaptation network, be utilized respectively marginal probability distribution adaptation network and
Conditional probability distribution adaptation network realizes the feature adaptation of source domain and aiming field high spectrum image, reduces the joint of the two
Probability distribution variances.
Second, the present invention is learnt using the training that source domain sample and a small amount of aiming field sample carry out classifier, is reduced
The demand of aiming field training sample；Meanwhile it being improved using one-to-many disaggregated model by learning the classifier of multiple two classification
Distinction in class class between improves the precision that migration is classified.
Detailed description of the invention
Fig. 1 is the implementation flow chart of high spectrum image migration classification；
Fig. 2 is source domain Pavia University high-spectral data；
Fig. 3 is aiming field Pavia Center high-spectral data；
Fig. 4 is not do the classification results figure migrated using one-to-many softmax classifier；
Fig. 5 is the classification results figure using the method for the present invention.
Specific embodiment
Below with reference to specific example and attached drawing, detailed elaboration is made to the present invention.
According to Fig. 1, the high spectrum image based on depth Joint Distribution adaptation network migrates classification method, including walks as follows
It is rapid:
(1) high spectrum image for reading source domain and aiming field carries out feature normalization and dimension is unified:
(1a) carries out linear normalization to the spectral signature of the high spectrum image of source domain and aiming field respectively, is distributed in it
Between 0 and 1；
(1b) if the spectral Dimensions of source domain and the high spectrum image of aiming field are different, the high spectrum image low to dimension into
Row zero padding is allowed to realize that dimension is unified with the high image of dimension；
(2) feature of source domain and aiming field high spectrum image is combined:
The spectral signature vector of source domain high spectrum image is turned to X by (2a)S, by the spectral signature of aiming field high spectrum image
Vector turns to XT；
The spectral signature of source domain and aiming field is combined into a vector set X=[X by (2b)S XT]；
(3) three layers of conditional probability distribution adaptation network are constructed, the marginal probability of source domain and aiming field high spectrum image is carried out
Distribution adaptation:
(3a) utilizes a linear denoising encoder and a non-linear encoder, and building first layer marginal probability distribution is suitable
Distribution network:
To sample xi, by each dimensional characteristics zero setting, M disturbed versions being obtained, wherein m-th is disturbed at random
Dynamic version is x 'i,m；By linear denoising encoder to x 'i,mLinearly insinuate to restore original sample,Optimize weight
The target equation of W is as follows:
Wherein, first item is average reconstruction error, and Section 2 is edge penalty term, and λ indicates balance factor, NSAnd NTRespectively
Indicate that the sample number of source domain EO-1 hyperion and aiming field EO-1 hyperion, N are total sample number, i.e. N=NS+NT；Enable disturbed sample moment
Battle array X 'm=[x '1,m,x′2,m,…,x′N,m], M times of original sample matrixM disturbed sample matrix
Combine X '=[X '1,X′2,…,X′M], such above formula target equation is converted into
Wherein, D indicates differential index (di) matrix, is defined as
In this way, linear denoising encoder weight W has following closed solution:
Later, it is insinuated using a non-linear encoder is non-linear to feature progress, formula is as follows
Here,Indicate the output of first layer marginal probability distribution adaptation network；
(3b) using a linear denoising encoder and a non-linear encoder, constructs the according to (3a) step as above
Two layers of marginal probability distribution adaptation network, solve the weight of linear denoising encoder, and to source domain and aiming field spectral signature into
Row adaptation, obtains the output of the second layer network
(3c) using a linear denoising encoder and a non-linear encoder, constructs the according to (3a) step as above
Three layers of marginal probability distribution adaptation network, solve the weight of linear denoising encoder, and to source domain and aiming field spectral signature into
Row adaptation, obtains the output of third layer network
(4) according to one-to-many principle of classification, the training sample of source domain and aiming field high spectrum image is chosen:
Classify to 9 classes, chooses each 500 samples of classification of source domain high spectrum image and each class of aiming field high spectrum image
Other 1% sample forms 9 training sample sets according to one-to-many principle of classification, and each training set sample label is to belong to such
It (being expressed as 1) and is not belonging to such (being expressed as 0), sample characteristics are the output of third layer network
(5) 9 three layers of conditional probability distribution adaptation networks are constructed, the condition for carrying out source domain and aiming field high spectrum image is general
Rate distribution adaptation:
(5a) first initializes the weight and offset parameter of three layers of conditional probability distribution adaptation network, recycles c-th of training
Each layer network of the layer-by-layer pre-training of sample set；The cataloged procedure and decoding process of kth layer conditional probability distribution adaptation network are as follows:
Wherein,WithRespectively indicate the input, implicit expression and output, f () and g of kth layer network
() is respectively the nonlinear function coded and decoded,WithThe respectively weight and biasing of cataloged procedure,WithPoint
Not Wei decoding process weight and biasing, * indicates S (source domain) or T (aiming field)；
The target equation of pre-training weight and offset parameter are as follows:
Wherein, first item indicates the average reconstruction error on training sample, and Section 2 is weight penalty term, prevents weight
It is excessive,NtrainIndicate training samples number, λ ' expression weight penalty factor；It is passed using reversed
It broadcasts algorithm and solves above formula；
After (5b) pre-training is complete, the implicit output of third layer conditional probability distribution adaptation networkAs the sample after optimization
Eigen；
(5c) connects softmax classifier after third layer conditional probability distribution adaptation network, after training sample is optimized
Feature and class label be input to classifier, optimize the weight and offset parameter of softmax classifier；
(5d) is finely adjusted conditional probability distribution adaptation network by top layer to bottom, advanced optimizes the ginseng of each network
Number；Reversely finely tune the target equation of entire network weight and offset parameter are as follows:
Wherein, first item is the total reconstruction error of coding layer, and Section 2 is condition penalty term, for reducing source domain and target
The difference of domain sample conditions probability distribution；WithThe respectively sample number of c class source domain EO-1 hyperion and aiming field EO-1 hyperion；
Equally, above formula is solved using back-propagation algorithm, each layer conditional probability distribution adaptation network and softmax after being finely tuned point
The parameter of class device；
The step of (5e) is according to above-mentioned (5a)-(5d) is utilized respectively 9 training sample set training and obtains 9 three layers of conditions
Probability distribution adaptation network realizes the conditional probability distribution adaptation of source domain and aiming field high spectrum image；
(6) one-to-many classification is carried out to aiming field high spectrum image:
The test sample collection of aiming field high spectrum image is input to c-th of conditional probability distribution adaptation network, net by (6a)
It is characterized in the implicit output of third layer conditional probability distribution adaptation network after network optimization
(6b) classifies to test sample using trained softmax classifier, and the prediction that acquisition belongs to c class is general
Rate:
Wherein,WithRespectively correspond fractional weight and the biasing of softmax classifier；
The step of (6c) is according to above-mentioned (6a)-(6b), is separately input to 9 for the test sample collection of aiming field high spectrum image
A conditional probability distribution adaptation network and softmax classifier obtain the probability that sample belongs to each classification；
(6d) compares the size for belonging to each class prediction probability, takes the corresponding classification of maximum probability as each sample
Prediction label, it is specific as follows
(6e) exports the classification knot of aiming field high spectrum image according to the spatial position of prediction label vector sum test sample
Fruit figure.
Below by way of emulation experiment, technical effect of the invention is illustrated:
1, simulated conditions and content
Experimental data of the invention is Pavia University high-spectral data and Pavia Center high-spectral data,
The data obtained by ROSIS sensor.It is source domain Pavia University high-spectral data, size 610 as shown in 2
× 340 pixels, 103 wave bands, Fig. 2 (a) be its 57th, 34,3 wave bands synthesis image, Fig. 2 (b) is corresponding true atural object
Label figure, shares 9 classes, including asphalt road, meadow, rubble, tree, metal plate, soil, pitch, brick and shade.It is as shown in 3
Aiming field Pavia Center high-spectral data, size be 1096 × 715 pixels, 102 wave bands, Fig. 3 (a) be its 57th, 34,
The image of 3 wave bands synthesis, Fig. 3 (b) is corresponding truly substance markers figure, shares 9 classes, including water, tree, meadow, brick, soil,
Asphalt road, pitch, tile and shade.Fig. 4 is not do the Pavia Center bloom migrated using one-to-many softmax classifier
For spectrogram as classification results figure, Fig. 5 is the knot for carrying out migration classification to Pavia Center high spectrum image using the method for the present invention
Fruit figure, table one is source domain and aiming field training sample corresponds to situation and its nicety of grading comparison.In emulation experiment, the present invention and right
Ratio method is all to program to realize in Matlab R2017a.
2, analysis of simulation result
The comparison of one nicety of grading of table
As shown in Table 1, migration classification is carried out using depth Joint Distribution adaptation network of the present invention to obtain than non-migratory classification
Obtained higher nicety of grading, it was demonstrated that validity of the method for the present invention in migration classification.Comparison diagram 4 and Fig. 5 are it is found that the present invention
The misclassified gene of the classification results figure of method is less, closer to truly substance markers figure.In short, method of the invention can have
Effect improves high spectrum image and migrates classifying quality.
Claims (1)
1. a kind of high spectrum image based on depth Joint Distribution adaptation network migrates classification method, which is characterized in that step is such as
Under:
(1) high spectrum image for reading source domain and aiming field carries out feature normalization and dimension is unified:
(1a) carries out linear normalization to the spectral signature of the high spectrum image of source domain and aiming field respectively, it is made to be distributed in 0 and 1
Between；
(1b) if source domain is different with the spectral Dimensions of the high spectrum image of aiming field, the high spectrum image low to dimension is mended
Zero, it is allowed to realize that dimension is unified with the high image of dimension；
(2) feature of source domain and aiming field high spectrum image is combined:
The spectral signature vector of source domain high spectrum image is turned to X by (2a)S, by the spectral signature vector of aiming field high spectrum image
Turn to XT；
The spectral signature of source domain and aiming field is combined into a vector set X=[X by (2b)S XT]；
(3) three layers of conditional probability distribution adaptation network are constructed, the marginal probability distribution of source domain and aiming field high spectrum image is carried out
Adaptation:
(3a) utilizes a linear denoising encoder and a non-linear encoder, and building first layer marginal probability distribution is adapted to net
Network:
To sample xi, by each dimensional characteristics zero setting, M disturbed versions being obtained, wherein m-th disturbed at random
Version is x 'i,m；By linear denoising encoder to x 'i,mLinearly insinuate to restore original sample,Optimize weight W's
Target equation is as follows:
Wherein, first item is average reconstruction error, and Section 2 is edge penalty term, and λ indicates balance factor, NSAnd NTIt respectively indicates
The sample number of source domain EO-1 hyperion and aiming field EO-1 hyperion, N are total sample number, i.e. N=NS+NT；Enable disturbed sample matrix
X′m=[x '1,m,x′2,m,…,x′N,m], M times of original sample matrixM disturbed sample matrix groups
Close X '=[X '1,X′2,…,X′M], such above formula target equation is converted into
Wherein, D indicates differential index (di) matrix, is defined as
In this way, linear denoising encoder weight W has following closed solution:
Later, it is insinuated using a non-linear encoder is non-linear to feature progress, formula is as follows
Here,Indicate the output of first layer marginal probability distribution adaptation network；
(3b) constructs the second layer using a linear denoising encoder and a non-linear encoder according to (3a) step as above
Marginal probability distribution adaptation network, solves the weight of linear denoising encoder, and fits to source domain and aiming field spectral signature
Match, obtains the output of the second layer network
(3c) constructs third layer using a linear denoising encoder and a non-linear encoder according to (3a) step as above
Marginal probability distribution adaptation network, solves the weight of linear denoising encoder, and fits to source domain and aiming field spectral signature
Match, obtains the output of third layer network
(4) according to one-to-many principle of classification, the training sample of source domain and aiming field high spectrum image is chosen:
Classify to 9 classes, chooses each 500 samples of classification of source domain high spectrum image and each classification of aiming field high spectrum image
1% sample forms 9 training sample sets according to one-to-many principle of classification, and each training set sample label is to belong to such (table
It is shown as 1) and is not belonging to such (being expressed as 0), sample characteristics are the output of third layer network
(5) 9 three layers of conditional probability distribution adaptation networks are constructed, the conditional probability point of source domain and aiming field high spectrum image is carried out
Cloth adaptation:
(5a) first initializes the weight and offset parameter of three layers of conditional probability distribution adaptation network, recycles c-th of training sample
Collect each layer network of layer-by-layer pre-training；The cataloged procedure and decoding process of kth layer conditional probability distribution adaptation network are as follows:
Wherein,WithThe input, implicit expression and output of kth layer network are respectively indicated, f () and g () divide
The nonlinear function that Wei do not code and decode, W1 kWithThe respectively weight and biasing of cataloged procedure,WithRespectively solve
The weight and biasing of code process, * indicate S (source domain) or T (aiming field)；
The target equation of pre-training weight and offset parameter are as follows:
Wherein, first item indicates the average reconstruction error on training sample, and Section 2 is weight penalty term, prevents weight mistake
Greatly,NtrainIndicate training samples number, λ ' expression weight penalty factor；Utilize backpropagation
Algorithm solves above formula；
After (5b) pre-training is complete, the implicit output of third layer conditional probability distribution adaptation networkIt is special as the sample after optimization
Sign；
(5c) connects softmax classifier, the spy after training sample is optimized after third layer conditional probability distribution adaptation network
Class label of seeking peace is input to classifier, optimizes the weight and offset parameter of softmax classifier；
(5d) is finely adjusted conditional probability distribution adaptation network by top layer to bottom, advanced optimizes the parameter of each network；
Reversely finely tune the target equation of entire network weight and offset parameter are as follows:
Wherein, first item is the total reconstruction error of coding layer, and Section 2 is condition penalty term, for reducing source domain and aiming field sample
The difference of this conditional probability distribution；WithThe respectively sample number of c class source domain EO-1 hyperion and aiming field EO-1 hyperion；Equally,
Above formula, each layer conditional probability distribution adaptation network and softmax classifier after being finely tuned are solved using back-propagation algorithm
Parameter；
The step of (5e) is according to above-mentioned (5a)-(5d) is utilized respectively 9 training sample set training and obtains 9 three layers of conditional probabilities
It is distributed adaptation network, realizes the conditional probability distribution adaptation of source domain and aiming field high spectrum image；
(6) one-to-many classification is carried out to aiming field high spectrum image:
The test sample collection of aiming field high spectrum image is input to c-th of conditional probability distribution adaptation network by (6a), and network is excellent
It is characterized in the implicit output of third layer conditional probability distribution adaptation network after change
(6b) classifies to test sample using trained softmax classifier, obtains the prediction probability for belonging to c class:
Wherein, Wj KWithRespectively correspond fractional weight and the biasing of softmax classifier；
The test sample collection of aiming field high spectrum image is separately input to 9 items by the step of (6c) is according to above-mentioned (6a)-(6b)
Part probability distribution adaptation network and softmax classifier obtain the probability that sample belongs to each classification；
(6d) compares the size for belonging to each class prediction probability, takes prediction of the corresponding classification of maximum probability as each sample
Label, it is specific as follows
(6e) exports the classification results of aiming field high spectrum image according to the spatial position of prediction label vector sum test sample
Figure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811347067.7A CN109359623B (en) | 2018-11-13 | 2018-11-13 | Hyperspectral image migration classification method based on depth joint distribution adaptive network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811347067.7A CN109359623B (en) | 2018-11-13 | 2018-11-13 | Hyperspectral image migration classification method based on depth joint distribution adaptive network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109359623A true CN109359623A (en) | 2019-02-19 |
| CN109359623B CN109359623B (en) | 2021-05-11 |
Family
ID=65344913
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811347067.7A Expired - Fee Related CN109359623B (en) | 2018-11-13 | 2018-11-13 | Hyperspectral image migration classification method based on depth joint distribution adaptive network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109359623B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961096A (en) * | 2019-03-19 | 2019-07-02 | 大连理工大学 | A kind of multimode high spectrum image migration classification method |
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110321941A (en) * | 2019-06-24 | 2019-10-11 | 西北工业大学 | The Compression of hyperspectral images and classification method of identifiable feature learning |
| CN110598636A (en) * | 2019-09-09 | 2019-12-20 | 哈尔滨工业大学 | Ship target identification method based on feature migration |
| CN111898635A (en) * | 2020-06-24 | 2020-11-06 | 华为技术有限公司 | Neural network training method, data acquisition method and device |
| CN112016392A (en) * | 2020-07-17 | 2020-12-01 | 浙江理工大学 | Hyperspectral image-based small sample detection method for soybean pest damage degree |
| CN112331313A (en) * | 2020-11-25 | 2021-02-05 | 电子科技大学 | Automatic grading method for sugar net image lesions based on label coding |
| CN113030197A (en) * | 2021-03-26 | 2021-06-25 | 哈尔滨工业大学 | Gas sensor drift compensation method |
| CN113505856A (en) * | 2021-08-05 | 2021-10-15 | 大连海事大学 | Hyperspectral image unsupervised self-adaptive classification method |
| CN115410088A (en) * | 2022-10-10 | 2022-11-29 | 中国矿业大学 | Hyperspectral image field self-adaption method based on virtual classifier |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011026167A1 (en) * | 2009-09-03 | 2011-03-10 | National Ict Australia Limited | Illumination spectrum recovery |
| CN106530246A (en) * | 2016-10-28 | 2017-03-22 | 大连理工大学 | Image dehazing method and system based on dark channel and non-local prior |
| CN107239759A (en) * | 2017-05-27 | 2017-10-10 | 中国科学院遥感与数字地球研究所 | A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic |
| JP2018081404A (en) * | 2016-11-15 | 2018-05-24 | パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカＰａｎａｓｏｎｉｃ Ｉｎｔｅｌｌｅｃｔｕａｌ Ｐｒｏｐｅｒｔｙ Ｃｏｒｐｏｒａｔｉｏｎ ｏｆ Ａｍｅｒｉｃａ | Discrimination method, discrimination device, discriminator generation method and discriminator generation device |
| CN108171232A (en) * | 2017-11-15 | 2018-06-15 | 中山大学 | The sorting technique of bacillary and viral children Streptococcus based on deep learning algorithm |
| CN108280396A (en) * | 2017-12-25 | 2018-07-13 | 西安电子科技大学 | Hyperspectral image classification method based on depth multiple features active migration network |
-
2018
- 2018-11-13 CN CN201811347067.7A patent/CN109359623B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2011026167A1 (en) * | 2009-09-03 | 2011-03-10 | National Ict Australia Limited | Illumination spectrum recovery |
| CN106530246A (en) * | 2016-10-28 | 2017-03-22 | 大连理工大学 | Image dehazing method and system based on dark channel and non-local prior |
| JP2018081404A (en) * | 2016-11-15 | 2018-05-24 | パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカＰａｎａｓｏｎｉｃ Ｉｎｔｅｌｌｅｃｔｕａｌ Ｐｒｏｐｅｒｔｙ Ｃｏｒｐｏｒａｔｉｏｎ ｏｆ Ａｍｅｒｉｃａ | Discrimination method, discrimination device, discriminator generation method and discriminator generation device |
| CN107239759A (en) * | 2017-05-27 | 2017-10-10 | 中国科学院遥感与数字地球研究所 | A kind of Hi-spatial resolution remote sensing image transfer learning method based on depth characteristic |
| CN108171232A (en) * | 2017-11-15 | 2018-06-15 | 中山大学 | The sorting technique of bacillary and viral children Streptococcus based on deep learning algorithm |
| CN108280396A (en) * | 2017-12-25 | 2018-07-13 | 西安电子科技大学 | Hyperspectral image classification method based on depth multiple features active migration network |
Non-Patent Citations (2)
| Title |
|---|
| CHUN-HUNG YEH: "Reduced Encoding Diffusion Spectrum Imaging Implemented With a Bi-Gaussian Model", 《IEEE TRANSACTIONS ON MEDICAL IMAGING 》 * |
| 樊利恒: "基于分类器集成的高光谱遥感图像分类方法", 《光学学报》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109961096B (en) * | 2019-03-19 | 2021-01-05 | 大连理工大学 | Multimode hyperspectral image migration classification method |
| CN109961096A (en) * | 2019-03-19 | 2019-07-02 | 大连理工大学 | A kind of multimode high spectrum image migration classification method |
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110321941A (en) * | 2019-06-24 | 2019-10-11 | 西北工业大学 | The Compression of hyperspectral images and classification method of identifiable feature learning |
| CN110598636A (en) * | 2019-09-09 | 2019-12-20 | 哈尔滨工业大学 | Ship target identification method based on feature migration |
| CN110598636B (en) * | 2019-09-09 | 2023-01-17 | 哈尔滨工业大学 | Ship target identification method based on feature migration |
| CN111898635A (en) * | 2020-06-24 | 2020-11-06 | 华为技术有限公司 | Neural network training method, data acquisition method and device |
| CN112016392A (en) * | 2020-07-17 | 2020-12-01 | 浙江理工大学 | Hyperspectral image-based small sample detection method for soybean pest damage degree |
| CN112331313A (en) * | 2020-11-25 | 2021-02-05 | 电子科技大学 | Automatic grading method for sugar net image lesions based on label coding |
| CN112331313B (en) * | 2020-11-25 | 2022-07-01 | 电子科技大学 | Automatic grading method for sugar net image lesions based on label coding |
| CN113030197A (en) * | 2021-03-26 | 2021-06-25 | 哈尔滨工业大学 | Gas sensor drift compensation method |
| CN113505856A (en) * | 2021-08-05 | 2021-10-15 | 大连海事大学 | Hyperspectral image unsupervised self-adaptive classification method |
| CN113505856B (en) * | 2021-08-05 | 2024-04-09 | 大连海事大学 | Non-supervision self-adaptive classification method for hyperspectral images |
| CN115410088A (en) * | 2022-10-10 | 2022-11-29 | 中国矿业大学 | Hyperspectral image field self-adaption method based on virtual classifier |
| CN115410088B (en) * | 2022-10-10 | 2023-10-31 | 中国矿业大学 | Hyperspectral image field self-adaption method based on virtual classifier |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109359623B (en) | 2021-05-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109359623A (en) | High spectrum image based on depth Joint Distribution adaptation network migrates classification method | |
| CN110321963B (en) | Hyperspectral image classification method based on fusion of multi-scale and multi-dimensional space spectrum features | |
| Ming et al. | Land cover classification using random forest with genetic algorithm-based parameter optimization | |
| CN103971123B (en) | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) | |
| CN107122733B (en) | Hyperspectral image classification method based on NSCT and SAE | |
| CN106991382A (en) | A kind of remote sensing scene classification method | |
| CN104408478A (en) | Hyperspectral image classification method based on hierarchical sparse discriminant feature learning | |
| CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
| CN109145832A (en) | Polarimetric SAR image semisupervised classification method based on DSFNN Yu non local decision | |
| CN104809471B (en) | A kind of high spectrum image residual error integrated classification method based on spatial spectral information | |
| CN108446582A (en) | Hyperspectral image classification method based on textural characteristics and affine propagation clustering algorithm | |
| CN104484681A (en) | Hyperspectral remote sensing image classification method based on space information and ensemble learning | |
| CN104182767B (en) | The hyperspectral image classification method that Active Learning and neighborhood information are combined | |
| CN107145836A (en) | Hyperspectral image classification method based on stack boundary discrimination self-encoding encoder | |
| Yousefi et al. | Comparison of different algorithms for land use mapping in dry climate using satellite images: a case study of the Central regions of Iran. | |
| CN105787517B (en) | Classification of Polarimetric SAR Image method based on the sparse self-encoding encoder of small echo | |
| CN105160351B (en) | Semi-supervised hyperspectral classification method based on anchor point sparse graph | |
| CN105760900A (en) | Hyperspectral image classification method based on affinity propagation clustering and sparse multiple kernel learning | |
| CN104778482A (en) | Hyperspectral image classifying method based on tensor semi-supervised scale cutting dimension reduction | |
| Ghasrodashti et al. | Hyperspectral image classification using an extended Auto-Encoder method | |
| CN110414616A (en) | A kind of remote sensing images dictionary learning classification method using spatial relationship | |
| CN113095409A (en) | Hyperspectral image classification method based on attention mechanism and weight sharing | |
| Savelonas et al. | Computer Vision and Pattern Recognition for the Analysis of 2D/3D Remote Sensing Data in Geoscience: A Survey | |
| CN108460326A (en) | A kind of high spectrum image semisupervised classification method based on sparse expression figure | |
| Wang et al. | TBC-YOLOv7: a refined YOLOv7-based algorithm for tea bud grading detection |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210511 Termination date: 20211113 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |