CN104866869B - Timing SAR image classification method based on distributional difference and incremental learning - Google Patents
Timing SAR image classification method based on distributional difference and incremental learning Download PDFInfo
- Publication number
- CN104866869B CN104866869B CN201510292485.0A CN201510292485A CN104866869B CN 104866869 B CN104866869 B CN 104866869B CN 201510292485 A CN201510292485 A CN 201510292485A CN 104866869 B CN104866869 B CN 104866869B
- Authority
- CN
- China
- Prior art keywords
- image
- sample
- sar image
- incremental learning
- distributional difference
- 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.)
- Expired - Fee Related
Links
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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A kind of timing SAR image classification method based on distributional difference and incremental learning, the incremental learning of selection and increment sample including increment sample.Detection is changed to time series SAR image using the method based on distributional difference first, utilize the statistical distribution of Egdeworth approximation theory estimation source domain and aiming field SAR image, calculate the cross entropy differential index (di) of image distribution, and variation detection figure is obtained according to the detection threshold value that CFAR method obtains, using unchanged region as increment sample；Then SVM classifier is initialized using the known training set of source domain image, and the parameter for the incremental learning and svm classifier model for completing entire increment sample set by iterative process updates.Distributional difference change detecting method and support vector machines incremental learning combine, and the domain that can be completed from source domain to aiming field is adaptive, to realize the high-precision classification of target area image.
Description
Technical field
The invention belongs to technical field of image processing, in particular to a kind of timing based on distributional difference and incremental learning
SAR image classification method.
Background technique
In recent years, with the appearance of plenty of time sequence image resource, a series of new theoretical methods are used for time sequence
The interpretation of column image, it is all in earth's surface drawing, vegetation dynamic changes monitoring, phenology information extraction, land use pattern monitoring etc.
It is multi-field to be widely used, it is that people's research and tracking natural history evolution, monitoring environment and resource dynamic become
The important and effective means changed.In the classification of time-series image, from starting point difference, it is broadly divided into the side of two mainstreams
To:
1) using the image of training set known to certain width in time-series image come the figure to other certain unknown training set
As classifying, classification method such as adaptive based on domain.This method using source domain and aiming field when correlation come to mesh
The image in mark domain is classified, and semisupervised classification method and Active Learning classification method are generally divided into.Semisupervised classification method is first
First with the effective training set training classifier of source domain image, then in conjunction with unlabelled sample in target area image to classifier
Parameter be updated, finally classified using the classifier after update to target area image, i.e., for target area image
In sample prosthetic label cost.Active Learning classification method is using the effective training set of source domain image as aiming field figure
As initial training set, then according to certain criterion function, by the method for iteration, select in aiming field most rich in information
The unmarked sample of amount, and handmarking is carried out, it is added to initial training concentration, the training set of target area image is carried out excellent
Change, is finally classified according to the training set after optimization to target area image.
2) time variability and spatial variability for combining multiple image in time series, can be realized to increased surface covering
Preferably divide.Such time-series image sorting algorithm mainly includes three parts: image preprocessing, and the extraction of feature divides
The selection of class device.Different feature extracting methods and different classifiers derive different sorting algorithms.The main packet of pretreatment
Include the filtering of radiation calibration, image calibration and image；The feature extracting method of mainstream includes feature based on ecological index, base
In feature, the feature based on image coherence and the textural characteristics etc. of backscattering coefficient variation；Classifier mainly includes maximum
Likelihood classifier, SVM, neural network, decision tree, Ensemble classifier etc..
These two kinds of methods inherently focus on the excavation of the substantive characteristics of time-series image and the selection two of classifier
Aspect.In these existing algorithms, there are a variety of trials in terms of the selection of classifier, but in the training process of classifier
In and the distinctive advantage of sequence image between underusing then.
Summary of the invention
It is an object of the invention to be directed to time series SAR image classification problem, propose a kind of new based on distributional difference
The time series SAR image classification method of detection and support vector machines incremental learning.
The technical scheme is that a kind of time series based on distributional difference detection and support vector machines incremental learning
SAR image classification method, comprising the following steps:
Step 1, detection being changed to time series SAR image using based on the method for distributional difference, realization is as follows,
In the SAR image distributed model estimation approached based on Edgeworth, the distribution p (x) of time series SAR image
It is expressed as follows,
Wherein, H3(x)=x3-3x,H4(x)=x4-6x2+3,H6(x)=x6-15x4+45x2- 15, k3And k4When respectively indicating
Between sequence SAR image 3 ranks and 4 rank cumulants, α (x) indicate standardized normal distribution, x be time series SAR image pixel；
The source domain image distribution p of time series SAR image centering is calculated using cross-entropy method1(x) and target area image
Distribution p2(x) differential index (di) is as follows,
After regularization processing, the differential index (di) of source domain image and target area image is expressed as follows,
KLDedgeworth=KLedgeworth(x1,x2)+KLedgeworth(x2,x1)
Obtain the differential index (di) figure based on cross entropy；
Step 2, it is based on preset false alarm rate, adaptively obtains detection threshold value IcIt is as follows,
Wherein, PfaFor given false alarm rate, p (t) is the distribution function of the differential index (di) figure of variation detection；
When obtaining detection threshold value IcAfterwards, step 1 gained differential index (di) figure is detected, obtains the two-value of variation detection
Figure, and unchanged sample is obtained from binary map；
Step 3, source domain image X is utilized1The initial SVM classifier of known training set training, the mark of sample will not changed
Label assign target area image X2, as increment sample；When the variable quantity caused by newly-increased increment sample meets condition, the new sample
Originally it is judged as supporting vector, initial supporting vector is added to and concentrates, and relevant classifier parameters are updated；Instead
It, then be not supporting vector, to SVM classifier without contribution, is directly abandoned；
It is updated, is obtained by the parameter of iterative process, the incremental learning and svm classifier model of completing entire increment sample set
It is suitable for target area image X2SVM classifier.
Moreover, the false alarm rate in step 2 in CFAR threshold value is 0.01.
Moreover, whether variable quantity caused by newly-increased increment sample meets condition, according to variation delta αcWhether 0 < is met
Δαc< C realizes that C is preset penalty factor.
Different from single width SAR image, the sample information that do not change that time series SAR image changes in testing result can be right
The classifier parameters of source domain image are updated, without re -training classifier, therefore can effectively improve nicety of grading and
Efficiency, the present invention, which passes through, combines time series variation detection method, SVM and incremental learning based on distributional difference, proposes one
The new distributional difference for the classification of time series SAR image of kind detects and support vector machines Increment Learning Algorithm.In single width figure
As realizing the variation detection of time series SAR image using distributional difference detection method, and combine SVM on the basis of feature representation
And Increment Learning Algorithm, the nicety of grading of SAR image can be effectively improved.
Detailed description of the invention
Fig. 1 is the distributional difference variation overhaul flow chart of time series SAR image of the embodiment of the present invention；
Fig. 2 is the svm classifier flow chart based on incremental learning of time series SAR image of the embodiment of the present invention.
Specific embodiment
The present invention is directed to the characteristics of time series SAR image, combines distributional difference change detecting method, SVM and increasing
Amount study, does not sufficiently change sample information with time series SAR image, proposes and classifies for time series SAR image
Based on distributional difference and supporting vector Increment Learning Algorithm.Unlike single width SAR image, time series SAR image point
Class can be updated classifier using the information for not changing sample, and re -training classifier is not needed.Therefore base can be used
In the change detecting method of distributional difference, the distributed model of time series SAR image is adaptively fitted, according to the difference of distribution
It is different to obtain variation testing result, and unchanged sample is chosen as increment sample, it is added in training set.It uses on this basis
Increment Learning Algorithm makes full use of variation testing result, is quickly trained to the supporting vector in increment sample, saves the time
The space and.Distributional difference change detecting method is in conjunction with supporting vector incremental learning, and the achievable domain from source domain to aiming field is certainly
It adapts to, effectively realizes the classification of target area image.
Below in conjunction with drawings and examples the present invention will be described in detail technical solution.
The embodiment of the present invention can be used computer software technology and realize automatic flow operation, this is described in detail step by step below
The classification process of inventive embodiments:
Step 1, detection is changed to time series SAR image using based on the method for distributional difference.Fig. 1 is variation
Overhaul flow chart.In the SAR image distributed model estimation approached based on Edgeworth expansion, time series SAR image
Distribution p (x) indicates are as follows:
Wherein Hermite multinomial H3(x)=x3- 3x, H4(x)=x4-6x2+ 3, H6(x)=x6-15x4+45x2- 15, k3
And k43 ranks and the 4 rank cumulants of time series SAR image are respectively indicated, α (x) indicates standardized normal distribution, and x is time series
The pixel of SAR image.It when being unfolded using Edgeworth and calculating Gross entropy, is carried out once in sample data in window
It calculates to quadruplicate cumulative amount, rather than calculates on the entire image.When it is implemented, for different experimental datas, this
Field technical staff can select suitable sliding window size according to variation testing result, and sliding window size used in the examples is 21 ×
21 pixels.
To the source domain image X of time-series image centering1Distribution p1(x) and target area image X2Distribution p2(x) it calculates
Cross entropy KLedgeworth(x1,x2):
When specific calculating, abbreviation can be carried out to above formula and obtained:
Wherein, x1,x2It is source domain image X respectively1, target area image X2Pixel；
Have each intermediate variable as follows:
kx；1=μx；1, μx；1~μx；4Respectively indicate 1~4 rank moment of the orign of image；Enable x
=x1,x2,x1',x2' when calculation it is identical.
c2=m2+n2, c3=m3+3mn2, c4=m4+6m2n2+3n4, c6=m6+15m4n2+45m2n4+15n6；
In order to solve the problems, such as that cross entropy is asymmetric, embodiment carries out regularization processing, the difference of two width SAR images to above formula
Different exponential representation are as follows:
KLDedgeworth=KLedgeworth(x1,x2)+KLedgeworth(x2,x1)
The differential index (di) figure based on cross entropy can be obtained, can equally be approached to obtain corresponding distribution letter by Edgeworth expansion
Number p (t), t is the point in differential index (di) figure, and then can obtain image change region according to step 2 gained CFAR threshold value.Two width
The corresponding point of every two has a differential index (di) in SAR image, this differential index (di) is the window centered on the two points
Upper calculating.If SAR image size is m × n, then the differential index (di) figure size finally obtained is also m × n.
Step 2, it is based on preset false alarm rate, according to the adaptive acquisition detection threshold value I of clutter distributed modelc:
Wherein, PfaFor preset false alarm rate, p (t) is the distribution function of the differential index (di) figure of variation detection, IcFor detection threshold
Value, i.e. CFAR threshold value (false alarm rate threshold value).When obtaining detection threshold value IcAfterwards, differential index (di) figure is detected, obtains variation inspection
Binary map is surveyed, and obtains unchanged sample from binary map.When it is implemented, those skilled in the art can voluntarily preset
The value of false alarm rate, false alarm rate is 0.01 in embodiment.
Step 3, Fig. 2 is time series SAR image classification process figure.Utilize source domain image X1Known training set training just
The SVM classifier of beginning assigns the label of unchanged sample to target area image X2, as increment sample.Sample is with some
The feature vector extracted in window centered on pixel.As newly-increased sample (xc,yc) caused by variation delta αcMeet condition
0 < Δ αcWhen < C (x indicates that sample, y indicate classification, and C is preset penalty factor, and those skilled in the art can voluntarily preset),
The new samples are judged as supporting vector, are added to initial supporting vector and concentrate, and carry out more to relevant classifier parameters
Newly；Conversely, not being then supporting vector, to SVM classifier without contribution, directly abandon.
The training of SVM classifier is the Optimization Learning of Lagrange multiplier α in decision function, in the feelings of Nonlinear separability
Under condition, convex quadratic objective function is indicated are as follows:
Wherein, W is convex quadratic objective function, and i, j are the subscript of sample, symmetric positive definite nuclear matrix Qij=yiyjK(xi,xj),
K is kernel function, xiIt is i-th of the sample indicated with vector, yiIt is corresponding classification marker, xjIt is j-th of the sample indicated with vector
This, yjIt is corresponding classification marker.
It is available when meeting KTT (Karush-Kuhn-Tucker) condition:
Wherein, f (xi) it is decision function, b is offset, giIt is W about αiPartial derivative.
As some new sample (xc,yc) be added when, caused variable quantity needs to meet KKT condition:
WhereinIndicate supporting vector set, lSFor the number of supporting vector.Sample (xc,yc) cause
Increment Delta αcMeet with offset b:
Δ b=β Δ αc,Δαj=βjΔαc
Wherein, Δ b is the variable quantity of offset b, and β is sensitivity coefficient.Initialize the inverse matrix R of Jacobian:
Then sensitivity coefficient is calculated
Wherein, β is the sensitivity coefficient of b,It is supporting vector s1Sensitivity coefficient ...
ΔαcCalculation formula are as follows:
Δαc=gc/γc
Intermediate variable gc、γcIt seeks as follows:
Wherein, when i, j are not indicated especially, arbitrary sample is indicated；J ∈ S indicates that j is the supporting vector in sample.
As increment Delta αcMeet 0 < Δ α of conditioncWhen < C, newly-increased sample (xc,yc) it is judged as supporting vector, it is added to
Initial supporting vector is concentrated, and is updated to relevant classifier parameters, conversely, not being then supporting vector, to svm classifier
Device is directly abandoned without contribution.In addition, when supporting vector collection is added in newly-increased sample, the more new formula of R are as follows:
The incremental learning that newly-increased sample is then completed by above parameter renewal process, by being held to each newly-increased sample
The parameter of the iterative process of the above incremental learning of row, the incremental learning and svm classifier model of completing entire increment sample set updates,
It obtains being suitable for target area image X2SVM classifier.
This mode can be promoted the use of, such as:
For X1, X2: X1As source domain image, X2As target area image, it is changed detection, then randomly chooses X1Often
Part (such as 10%) does not change sample (it is recommended that not selecting marginal portion as far as possible) in one kind, with their feature vector to SVM into
Row updates, and completes X2Classification.
For X2, X3: X2As source domain image, X3As target area image.The known X from previous step2Distribution p2(x) and
Classification results obtain p further according to Edgeworth expansion3(x), it is changed detection；Then X is randomly choosed2Every a kind of middle part
Divide (such as 10%) not change sample (it is recommended that not selecting marginal portion as far as possible), SVM is carried out more again with their feature vector
Newly, and X is completed3Classification.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (3)
1. a kind of timing SAR image classification method based on distributional difference and incremental learning, which is characterized in that including following step
Suddenly,
Step 1, detection being changed to time series SAR image using based on the method for distributional difference, realization is as follows,
In the SAR image distributed model estimation approached based on Edgeworth, the distribution p (x) of time series SAR image is indicated
It is as follows,
Wherein, H3(x)=x3-3x,H4(x)=x4-6x2+3,H6(x)=x6-15x4+45x2- 15, k3And k4Respectively indicate time sequence
3 ranks of column SAR image and 4 rank cumulants, α (x) indicate that standardized normal distribution, x are the pixel of time series SAR image；
The source domain image distribution p of time series SAR image centering is calculated using cross-entropy method1(x) and aiming field image distribution p2
(x) differential index (di) is as follows,
In order to solve the problems, such as that cross entropy is asymmetric, after regularization processing, the differential index (di) table of source domain image and target area image
Show it is as follows,
KLDedgeworth=KLedgeworth(x1,x2)+KLedgeworth(x2,x1)
Obtain the differential index (di) figure based on cross entropy；
Step 2, it is based on preset false alarm rate, detection threshold value I is adaptively obtained according to clutter distributed modelcIt is as follows,
Wherein, PfaFor given false alarm rate, p (t) is the distribution function of the differential index (di) figure of variation detection；
When obtaining detection threshold value IcAfterwards, step 1 gained differential index (di) figure is detected, obtain variation detection binary map, and from
Unchanged sample is obtained in binary map；
Step 3, source domain image X is utilized1The initial SVM classifier of known training set training, the label that will do not change sample assigns
Target area image X2, as increment sample；
If W is convex quadratic objective function, i, j are the subscript of sample, symmetric positive definite nuclear matrix Qij=yiyjK(xi,xj), K is core letter
Number, xiIt is i-th of the sample indicated with vector, yiIt is corresponding classification marker, xjIt is j-th of the sample indicated with vector, yjIt is phase
Answer classification marker；
It is obtained when meeting KTT condition:
Wherein, f (xi) it is decision function, b is offset, giIt is W about αiPartial derivative, C be preset penalty factor；
As some new sample (xc,yc) be added when, caused variable quantity needs to meet KKT condition, sample (xc,yc) caused by
Increment Delta αcMeet Δ b=β Δ α with offset bc, wherein Δ b is the variable quantity of offset b, and β is sensitivity coefficient, by initial
The inverse matrix for changing Jacobian obtains；
When the variable quantity caused by newly-increased increment sample meets condition, which is judged as supporting vector, is added to just
The supporting vector of beginning is concentrated, and is updated to relevant classifier parameters；Conversely, not being then supporting vector, to SVM classifier
Without contribution, directly abandon；
It is updated, is suitble to by the parameter of iterative process, the incremental learning and svm classifier model of completing entire increment sample set
In target area image X2SVM classifier.
2. the timing SAR image classification method based on distributional difference and incremental learning as described in claim 1, it is characterised in that:
False alarm rate in step 2 in CFAR threshold value is 0.01.
3. the timing SAR image classification method based on distributional difference and incremental learning, feature exist as claimed in claim 1 or 2
In: whether variable quantity caused by newly-increased increment sample meets condition, according to variation delta αcWhether 0 < Δ α is metc< C is real
Existing, C is preset penalty factor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510292485.0A CN104866869B (en) | 2015-05-29 | 2015-05-29 | Timing SAR image classification method based on distributional difference and incremental learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510292485.0A CN104866869B (en) | 2015-05-29 | 2015-05-29 | Timing SAR image classification method based on distributional difference and incremental learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104866869A CN104866869A (en) | 2015-08-26 |
| CN104866869B true CN104866869B (en) | 2018-12-14 |
Family
ID=53912690
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510292485.0A Expired - Fee Related CN104866869B (en) | 2015-05-29 | 2015-05-29 | Timing SAR image classification method based on distributional difference and incremental learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104866869B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105224949B (en) * | 2015-09-23 | 2018-11-13 | 电子科技大学 | SAR image terrain classification method based on cross-cutting transfer learning |
| CN107346433B (en) * | 2016-05-06 | 2020-09-18 | 华为技术有限公司 | Text data classification method and server |
| CN106897705B (en) * | 2017-03-01 | 2020-04-10 | 上海海洋大学 | Ocean observation big data distribution method based on incremental learning |
| CN109564632A (en) * | 2017-05-24 | 2019-04-02 | 深圳配天智能技术研究院有限公司 | A kind of visible detection method, equipment, system and the device with store function |
| CN107153180B (en) * | 2017-06-15 | 2020-02-07 | 中国科学院声学研究所 | Target signal detection method and system |
| CN107358257B (en) * | 2017-07-07 | 2019-07-16 | 华南理工大学 | Under a kind of big data scene can incremental learning image classification training method |
| CN107506799B (en) * | 2017-09-01 | 2020-04-24 | 北京大学 | Deep neural network-based mining and expanding method and device for categories of development |
| CN108229589B (en) * | 2018-02-09 | 2021-10-15 | 天津师范大学 | Foundation cloud picture classification method based on transfer learning |
| JP6839116B2 (en) * | 2018-02-09 | 2021-03-03 | 日本電信電話株式会社 | Learning device, estimation device, learning method, estimation method and computer program |
| CN108537227A (en) * | 2018-03-21 | 2018-09-14 | 华中科技大学 | A kind of offline false distinguishing method of commodity based on width study and wide-angle micro-image |
| CN109214386B (en) * | 2018-09-14 | 2020-11-24 | 京东数字科技控股有限公司 | Method and apparatus for generating image recognition model |
| CN109685366A (en) * | 2018-12-24 | 2019-04-26 | 中国人民解放军32181部队 | Equipment health state evaluation method based on mutation data |
| CN111259784B (en) * | 2020-01-14 | 2023-02-07 | 西安理工大学 | SAR image change detection method based on transfer learning and active learning |
| CN117133039B (en) * | 2023-09-01 | 2024-03-15 | 中国科学院自动化研究所 | Image fake identification model training method, image fake identification device and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101551856A (en) * | 2009-05-22 | 2009-10-07 | 西安电子科技大学 | SAR target recognition method based on sparse least squares support vector machine |
| CN101833093A (en) * | 2009-03-11 | 2010-09-15 | 中国科学院电子学研究所 | Method for detecting automatic variation of satellite-borne multiple phase synthetic aperture radar image |
| CN104318242A (en) * | 2014-10-08 | 2015-01-28 | 中国人民解放军空军工程大学 | High-efficiency SVM active half-supervision learning algorithm |
| CN104463218A (en) * | 2014-12-16 | 2015-03-25 | 重庆邮电大学 | sEMG self-adaptive mode recognition method based on on-line SVM and application of method on intelligent wheelchair |
-
2015
- 2015-05-29 CN CN201510292485.0A patent/CN104866869B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101833093A (en) * | 2009-03-11 | 2010-09-15 | 中国科学院电子学研究所 | Method for detecting automatic variation of satellite-borne multiple phase synthetic aperture radar image |
| CN101551856A (en) * | 2009-05-22 | 2009-10-07 | 西安电子科技大学 | SAR target recognition method based on sparse least squares support vector machine |
| CN104318242A (en) * | 2014-10-08 | 2015-01-28 | 中国人民解放军空军工程大学 | High-efficiency SVM active half-supervision learning algorithm |
| CN104463218A (en) * | 2014-12-16 | 2015-03-25 | 重庆邮电大学 | sEMG self-adaptive mode recognition method based on on-line SVM and application of method on intelligent wheelchair |
Non-Patent Citations (1)
| Title |
|---|
| 基于增量学习向量SVM方法的图像分割应用;陈沅涛 等;《南京理工大学学报》;20140228;第38卷(第1期);第6-11页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104866869A (en) | 2015-08-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104866869B (en) | Timing SAR image classification method based on distributional difference and incremental learning | |
| Roy et al. | Spectral–spatial morphological attention transformer for hyperspectral image classification | |
| Huang et al. | A new building extraction postprocessing framework for high-spatial-resolution remote-sensing imagery | |
| CN110728197B (en) | Single-tree-level tree species identification method based on deep learning | |
| CN107784657A (en) | A kind of unmanned aerial vehicle remote sensing image partition method based on color space classification | |
| CN108537102A (en) | High Resolution SAR image classification method based on sparse features and condition random field | |
| CN107545571A (en) | A kind of image detecting method and device | |
| CN107452022A (en) | A kind of video target tracking method | |
| CN111507296A (en) | Intelligent illegal building extraction method based on unmanned aerial vehicle remote sensing and deep learning | |
| CN106023159B (en) | Facilities vegetable leaf portion scab image partition method and system | |
| Ling et al. | Learning-based superresolution land cover mapping | |
| CN102163285A (en) | Cross-domain video semantic concept detection method based on active learning | |
| CN109784205A (en) | A kind of weeds intelligent identification Method based on multispectral inspection image | |
| Genitha et al. | Classification of satellite images using new fuzzy cluster centroid for unsupervised classification algorithm | |
| CN108388828A (en) | A kind of seashore wetland land cover pattern information extracting method of comprehensive multi- source Remote Sensing Data data | |
| Han et al. | Object-wise joint-classification change detection for remote sensing images based on entropy query-by fuzzy ARTMAP | |
| Niroumand Jadidi et al. | A novel approach to super resolution mapping of multispectral imagery based on pixel swapping technique | |
| Liu et al. | Contour-oriented cropland extraction from high resolution remote sensing imagery using richer convolution features network | |
| Zhang et al. | Contour structural profiles: An edge-aware feature extractor for hyperspectral image classification | |
| Nogueira et al. | Coffee crop recognition using multi-scale convolutional neural networks | |
| CN110097067A (en) | It is a kind of based on layer into the Weakly supervised fine granularity image classification method of formula eigentransformation | |
| Kuchkorov et al. | Perspectives of deep learning based satellite imagery analysis and efficient training of the U-Net architecture for land-use classification | |
| Lyu et al. | A deep information based transfer learning method to detect annual urban dynamics of Beijing and Newyork from 1984–2016 | |
| CN110378405A (en) | The Hyperspectral Remote Sensing Imagery Classification method of Adaboost algorithm based on transfer learning | |
| CN107992863A (en) | Multiresolution grain worm species visual identity method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate 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: 20181214 Termination date: 20190529 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |