CN108154495A - Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks - Google Patents
Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks Download PDFInfo
- Publication number
- CN108154495A CN108154495A CN201711146227.7A CN201711146227A CN108154495A CN 108154495 A CN108154495 A CN 108154495A CN 201711146227 A CN201711146227 A CN 201711146227A CN 108154495 A CN108154495 A CN 108154495A
- Authority
- CN
- China
- Prior art keywords
- dccae
- networks
- network
- remote sensing
- relative spectral
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks, step (1) chooses training sample；Step (2) determines DCCAE network parameters, obtains enabling to the network parameter of relevance maximum and reconstructed error minimum, builds trained DCCAE networks；Two width multidate complete images are input in trained DCCAE networks by step (3) so that two images are aligned in latent space relative spectral, obtain its expression in latent space.Compared with prior art, deep neural network is applied in across the sensor remote sensing images relative spectral alignment algorithms of multidate by the present invention for the first time, effectively increases the accuracy rate of spectrum alignment.
Description
Technical field
The present invention relates to image processing field, more particularly, to a kind of based on depth canonical correlation autoencoder network
Across the sensor remote sensing images relative spectral alignment algorithms of multidate.
Background technology
Variation detection (change detection) refers to that handle the remote sensing images of different phase the same areas becomes to extract
Change information, and the process either quantitatively or qualitatively analyzed is made to feature difference or space structure difference.Variation detection is remote sensing
As the key technology of monitoring surface condition, great function is played in various fields for field important research direction.Remote sensing figure
Land use and vegetation situation can be analyzed as variation detects, evaluation and test disaster development trend is estimated, identifies forest cover change,
The covering of the surfaces such as urban sprawl and layout variation tendency and Evolution etc. are shown, with urgent scientific practice demand and extensively
Wealthy actual application prospect.
By multiple sensors in the processing of different time the image collected and the method being integrated into the remote sensing system of standard
It is highly important.First, the fusion feature extracted from complementary data has more information content；Secondly, from difference
It is identical similar to identical sensing wavelength glazing spectral resolution that the collected multi-temporal image of sensor can substantially reduce some
Deng constraint, the temporal resolution of this research may be increased in this way.In addition, satellite operating needs the regular hour, use is single
Sensor table sample condition needs certain period with shooting same geographic area.It can be by arbitrary multisensor and across sensing
The integrated of device image is further reduced the response time of image processing system, such as after calamity needed for management and hazards entropy
Response time.Moment sensor type is enriched, and quantity is more, makes full use of resource, is carried out with the image that different sensors obtain
Variation detection is a good selection.
It is a boundless field to be changed detection after across sensor remote sensing images relative spectral alignment, is sent out in recent years
The deep learning (Deep Learning) that exhibition is got up is provided for a brand-new direction.Come from current achievement in research
It sees, has achieved preferable achievement, show excellent performance of the deep learning in terms of Remote Sensing Imagery Change Detection.But by
In the diversity and complexity across sensor Multitemporal Remote Sensing Images, the research of this respect is still seldom, and deep learning is in remote sensing
Application in Image Change Detection field also has very big excavation space.
So far, in the paper and document at home and abroad published there is not yet by depth canonical correlation own coding net
Network is applied in Remote Sensing Imagery Change Detection.
Invention content
Based on the above-mentioned prior art, the present invention proposes a kind of multidate based on DCCAE networks across sensor remote sensing images
Relative spectral alignment algorithm using a preprocessing means, can will be mapped to same latent space across sensor remote sensing images.
A kind of across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks of the present invention, should
Algorithm includes the following steps：
Step 1 chooses training sample, i.e., randomly selects a certain number of not changed pictures according to ground truth
Member (pixel that i.e. label is 0) is as training sample；
Step 2, the weights and bias for determining to calculate each layer of f, g, p, q network in DCCAE networks, obtaining can
So that the network parameter of relevance maximum and reconstructed error minimum, builds trained DCCAE networks, specifically includes：It will instruction
Practice sample input DCCAE networks, extract the nonlinear characteristic of image X and Y with two neural network f and g respectively, will extract
Relevance between the nonlinear characteristic f (X) and g (Y) that go out maximizes, while respectively with reconstruct autoencoder network p and q come weight
Two images of structure, using gradient descent method optimal reconfiguration error and the relevance of character representation, loss function is expressed as：
It should
Formula represents to maximize the relevance of character representation simultaneously and minimizes the reconstructed error of two images；
Constraints formula：
ui Tf(X)g(Y)Tvj=0, for i ≠ j,
Wherein, Wf, Wg, Wp, WqThe weights of f, g, p, q network, U=[u are represented respectively1,...,uL] and V=[v1,...,vL]
It represents the mapping function of canonical correlation analysis, neural network f and the g result exported is mapped to latent space, 0 tables of λ ＞ respectively
Show balance parameters, N represents the number of sample point, (rx,ry) ＞ 0 represent sample covariance estimation regularization parameter；By more
Secondary training calculates the weights and bias of each layer of f, g, p, q network in DCCAE networks, builds trained DCCAE nets
Network；
Step 3 inputs piece image X and the second width image Y respectively to the f networks in trained DCCAE networks
In g networks, network output result is f (X) and g (Y), and f (X) and g (Y) are the expression of X and Y in latent space, realize two
Person is aligned in the relative spectral of latent space.
Compared with prior art, the present invention has the following effects that：
Obtain one be similar to pretreatment as a result, by waveband channels number is different, spatial resolution is different more
Phase remote sensing images are mapped in same latent space, in this space, can use any change detection algorithm, have
It is widely applied prospect.
Description of the drawings
Across the sensor remote sensing images relative spectral alignment algorithms of the multidate based on DCCAE networks that Fig. 1 is the present invention are whole
Body flow chart；
Fig. 2 is depth canonical correlation autoencoder network (DCCAE) network；
Fig. 3 is L5VSALI experimental result comparison diagrams.
Specific embodiment
Depth canonical correlation autoencoder network (DCCAE) is applied to across sensor remote sensing images variations and detected by the present invention
The preprocessing process to image in.In order to make full use of sensor image resource, the remote sensing that will be obtained from different sensors
Image passes through in DCCAE network mappings to same latent space, so as to be calculated in this space using any variation detection
Method.
The pretreatment network DCCAE is by two autoencoder networks and linear CCA Algorithm constitutions.The coding of autoencoder network
Network and decoding network are individual two networks.Input linear CCA algorithms be two coding networks output feature.
Two width Multitemporal Remote Sensing Images are inputted first, are randomly selected and a certain number of not become according to ground truth
The pixel of change is as training sample；Secondly, training sample is inputted into DCCAE networks, obtains enabling to relevance maximum simultaneously
And the network parameter of reconstructed error minimum.Then, complete image is inputted into network so that two images are in latent space phase
Spectrum is aligned, obtains its expression in latent space.Finally by treated, image is input in CVA variation detection models,
Extract change information.Variation testing result is assessed, and export result.
Process is as follows：
Step 1 chooses training sample：A certain number of not changed pixels are randomly selected according to ground truth
(pixel that i.e. label is 0) preserves the location index value of sample point as training sample.The number of sample point is respectively
50th, 100,250 and 500, it chooses 10 times respectively.
Step 2, the weights and bias for determining to calculate each layer of f, g, p, q network in DCCAE networks, obtaining can
So that the network parameter of relevance maximum and reconstructed error minimum, builds trained DCCAE networks：Training sample is defeated
Enter DCCAE networks, extract the nonlinear characteristic of image X and Y with two neural network f and g respectively, it is non-thread by what is extracted
Property feature f (X) and g (Y) between relevance maximize, while reconstruct two images with reconstruct autoencoder network p and q respectively.
The relevance of optimization reconstructed error and character representation is needed, loss function is：
The formula represents to maximize the relevance of character representation simultaneously and minimizes the reconstructed error of two images.
ui Tf(X)g(Y)Tvj=0, for i ≠ j,
Constraints formula：
ui Tf(X)g(Y)Tvj=0, for i ≠ j,
Wherein, Wf, Wg, Wp, WqThe weights of f, g, p, q network, U=[u are represented respectively1,...,uL] and V=[v1,...,vL]
It represents the mapping function of canonical correlation analysis, neural network f and the g result exported is mapped to latent space, 0 tables of λ ＞ respectively
Show balance parameters, N represents the number of sample point, (rx,ry) ＞ 0 represent sample covariance estimation regularization parameter；By more
Secondary training calculates the weights and bias of each layer of f, g, p, q network in DCCAE networks, builds trained DCCAE nets
Network.
Further loss function is optimized by batch gradient descent method (Batch GradientDescent), is sought
Find optimal parameter.Because the training set data amount of remote sensing images is smaller, directly passed through using BGD to all samples
Calculate the direction to solve gradient.Each the corresponding gradient of parameter θ is：
Each the value of parameter θ is：
Wherein, hθ(x) it is the function to be fitted,It is loss function, θ is the parameter to be iteratively solved, and m is trained
The record strip number of collection, j are the numbers of parameter.
By repeatedly training, the weights and bias of each layer of DCCAE networks are calculated, build trained DCCAE nets
Network.
Step 3 inputs piece image X and the second width image Y respectively to the f networks in trained DCCAE networks
In g networks, network output result is f (X) and g (Y).F (X) and g (Y) is the expression of X and Y in latent space, realizes two
Person is aligned in the relative spectral of latent space.Two identical with spatial resolution for waveband channels number g (Y) width remote sensing images, institute
Existing change detection algorithm can be used.
Two width multidate complete images are input in trained DCCAE networks by step 3 so that two images are latent
In space, relative spectral is aligned, and obtains its expression in latent space：Piece image X is transformed to f (X), the second width image Y
It is transformed to g (Y).G (Y) is the waveband channels number two width remote sensing images identical with spatial resolution, it is possible to using existing
Change detection algorithm.
Step 4 performs change detection algorithm：Depth canonical correlation autoencoder network is applied across sensor remote sensing images
In change detection algorithm, the remote sensing images that waveband channels number is different, spectral resolution is inconsistent are mapped to same potential sky
Between, realize relative spectral alignment.In this space, any existing change detection algorithm can be used.It detects to occur special
The pixel position of fixed variation (such as fire).The present invention is tested using the Change vector Analysis method (CVA) of standard.Also may be used
To use the better change detection algorithm of effect.The radiation variation when change detection algorithm considers different between phase images, side
It overweights calculating and analyzes the difference between each wave band, to determine the intensity of variation and directional characteristic.By the wave band picture of two phases
First value makes the difference to obtain change vector, and size represents the intensity of variation, and direction represents each pixel in the side that each wave band changes
To.Between two phases being determined according to the threshold value of separation, the regional change/unchanged position.
Change vector Analysis method is expressed as with mathematical way：
Wherein, G=(g1,g2,...,gn)TWith H=(h1,h2,...,hn)TIt is represented respectively from different phase the same areas
Two images in the spectral value of same position, giAnd hiTwo phases are represented respectively.
Depth canonical correlation autoencoder network (DCCAE) is used to across sensor remote sensing figure to proposed by the present invention below
It is verified as doing pretreatment.It meanwhile will be at the handling result of the algorithm and current leading algorithm KCCA by emulation experiment
The result of reason is compared, and carrys out the validity of verification algorithm.The parameter used in experiment for：In DCCAE networks, herein will
Four deep neural networks f, g, p and q's is set as：Three hidden layers and an output layer.Wherein each hidden layer has 28
Neuron, output layer have 7 neurons.Each layer is using Sigmoid functions as activation primitive.Initial learning rate is set as
0.01, batch processing size is set as 50, random initializtion weight.
The present invention gives the performance for quantifying variation detection using 2 parameters：Overall classification accuracy (OA) and Kappa coefficient
(Kappa).OA values are equal to the ratio for the pixel number and total pixel number correctly classified, and describe classification well on the whole
Precision.Kappa coefficients are the methods of another characterization nicety of grading, equal to point using certain algorithm classification and completely random
Class generates the ratio of wrong reduction, can be represented with equation below：I.e.：
Fig. 2 is that test image by the DCCAE algorithms that MAD and this patent use across sensor remote sensing images in advance locate
The mapping graph of reason and variation detection figure.From comparison as can be seen that in three groups of experiments, by the pretreated figures of DCCAE
Significantly the variation testing result for obtaining image more pretreated than process MAD is good for the variation testing result of picture.
It can be seen from Tables 1 and 2 in same sensor experiment, the image energy after DCCAE network mappings
It is enough to carry out accurately variation detection, it is linear to extract characterization method MAD's although result has got well a little than original image
Testing result is poor.In L5T1vs.ALIT2 across in sensor experiment, result by DCCAE network processes in close proximity to
The testing result for the original image that same sensor obtains, and only it is somewhat below the same sensor handled with DCCAE
Image testing result.The result of MAD is lower than DCCAE again, this shows to have realized when only calculating data are non-linear
The relative spectral alignment of effect ground.It is more complicated due to testing for the last one experiment, so accuracy of detection is relatively low.In this feelings
Under condition, MAD cannot detect the information of variation, this is because MAD does not build a large amount of universal differences well
It is molded as.DCCAE networks constantly learn to extract the nonlinear characteristic of image by great amount of samples, will not changed area
Domain is aligned in latent space.It is used in combination to obtain the angle and amplitude that change more accurately as a result, in CVA algorithms,
To exclude the change information unrelated with vegetation.
Table 1, variation testing result data analysis table (OA values)
| Test serial number | Original method | MAD | DCCAE |
| L5vs.L5 | 86.401 | 81.137 | 95.507 |
| L5vs.ALI | 85.507 | 94.833 | |
| L5vs.L8 | 71.5781 | 88.895 |
Table 2 changes testing result data analysis table (Kappa coefficients)
| Test serial number | Original method | MAD | DCCAE |
| L5vs.L5 | 0.632 | 0.414 | 0.771 |
| L5vs.ALI | 0.457 | 0.752 | |
| L5vs.L8 | 0.169 | 0.512 |
Claims (1)
1. across the sensor remote sensing images relative spectral alignment algorithms of a kind of multidate based on DCCAE networks, which is characterized in that should
Algorithm includes the following steps：
Step (1) chooses training sample, i.e., randomly selects a certain number of not changed pixels according to groundtruth
(pixel that i.e. label is 0) is as training sample；
Step (2), the weights and bias for determining to calculate each layer of f, g, p, q network in DCCAE networks, are enabled to
The network parameter of relevance maximum and reconstructed error minimum, builds trained DCCAE networks, specifically includes：It will training sample
This input DCCAE networks extract the nonlinear characteristic of image X and Y with two neural network f and g respectively, non-by what is extracted
Relevance between linear character f (X) and g (Y) maximizes, while reconstructs two figures with reconstruct autoencoder network p and q respectively
Picture, using gradient descent method optimal reconfiguration error and the relevance of character representation, loss function is expressed as：
The formula
Represent the relevance for maximizing character representation simultaneously and the reconstructed error for minimizing two images；
Constraints formula：
ui Tf(X)g(Y)Tvj=0, fori ≠ j,
Wherein, Wf,Wg,Wp,WqThe weights of f, g, p, q network, U=[u are represented respectively1,...,uL] and V=[v1,...,vL] represent
The mapping function of canonical correlation analysis, is mapped to latent space by neural network f and the g result exported respectively, and λ ＞ 0 represent flat
Weigh parameter, and N represents the number of sample point, (rx,ry) ＞ 0 represent sample covariance estimation regularization parameter；By repeatedly instructing
Practice, calculate the weights and bias of each layer of f, g, p, q network in DCCAE networks, build trained DCCAE networks；
Step (3) inputs piece image X and the second width image Y respectively to the f networks and g in trained DCCAE networks
In network, network output result is f (X) and g (Y), and f (X) and g (Y) are the expression of X and Y in latent space, realizes the two and exists
The relative spectral alignment of latent space.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711146227.7A CN108154495A (en) | 2017-11-17 | 2017-11-17 | Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711146227.7A CN108154495A (en) | 2017-11-17 | 2017-11-17 | Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108154495A true CN108154495A (en) | 2018-06-12 |
Family
ID=62468006
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711146227.7A Pending CN108154495A (en) | 2017-11-17 | 2017-11-17 | Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108154495A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113487546A (en) * | 2021-06-25 | 2021-10-08 | 中南大学 | Feature-output space double-alignment change detection method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198483A (en) * | 2013-04-07 | 2013-07-10 | 西安电子科技大学 | Multiple time phase remote sensing image registration method based on edge and spectral reflectivity curve |
| CN106203256A (en) * | 2016-06-24 | 2016-12-07 | 山东大学 | A kind of low resolution face identification method based on sparse holding canonical correlation analysis |
| CN106529604A (en) * | 2016-11-24 | 2017-03-22 | 苏州大学 | Adaptive image tag robust prediction method and system |
-
2017
- 2017-11-17 CN CN201711146227.7A patent/CN108154495A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103198483A (en) * | 2013-04-07 | 2013-07-10 | 西安电子科技大学 | Multiple time phase remote sensing image registration method based on edge and spectral reflectivity curve |
| CN106203256A (en) * | 2016-06-24 | 2016-12-07 | 山东大学 | A kind of low resolution face identification method based on sparse holding canonical correlation analysis |
| CN106529604A (en) * | 2016-11-24 | 2017-03-22 | 苏州大学 | Adaptive image tag robust prediction method and system |
Non-Patent Citations (4)
| Title |
|---|
| HICHEM SAHBI: "Misalignment resilient CCA for interactive satellite image change detection", 《2016 23RD INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION》 * |
| MICHELE VOLPI ET AL: "Multi-sensor change detection based on nonlinear canonical correlations", 《2013 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM》 * |
| MICHELE VOLPI ET AL: "Spectral alignment of multi-temporal cross-sensor images with automated kernel canonical correlation analysis", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 * |
| WEIRAN WANG ET AL: "On Deep Multi-View Representation Learning: Objectives and Optimization", 《HTTPS://ARXIV.ORG/ABS/1602.01024》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113487546A (en) * | 2021-06-25 | 2021-10-08 | 中南大学 | Feature-output space double-alignment change detection method |
| CN113487546B (en) * | 2021-06-25 | 2024-04-02 | 中南大学 | Feature-output space double-alignment change detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yao et al. | Nonconvex-sparsity and nonlocal-smoothness-based blind hyperspectral unmixing | |
| Wu et al. | Unsupervised change detection in multitemporal VHR images based on deep kernel PCA convolutional mapping network | |
| Gu et al. | Integration of spatial–spectral information for resolution enhancement in hyperspectral images | |
| Yang et al. | Blind spectral unmixing based on sparse nonnegative matrix factorization | |
| Liu et al. | Enhancing spectral unmixing by local neighborhood weights | |
| CN111080629A (en) | Method for detecting image splicing tampering | |
| Zhang et al. | Automatic radiometric normalization for multitemporal remote sensing imagery with iterative slow feature analysis | |
| Mustapha et al. | Comparison of neural network and maximum likelihood approaches in image classification | |
| CN108596108A (en) | Method for detecting change of remote sensing image of taking photo by plane based on the study of triple semantic relation | |
| CN108197650A (en) | The high spectrum image extreme learning machine clustering method that local similarity is kept | |
| CN107145836A (en) | Hyperspectral image classification method based on stack boundary discrimination self-encoding encoder | |
| Li et al. | Bayesian Markov chain random field cosimulation for improving land cover classification accuracy | |
| He et al. | Multi-spectral remote sensing land-cover classification based on deep learning methods | |
| CN103426158A (en) | Method for detecting two-time-phase remote sensing image change | |
| Khoshboresh-Masouleh et al. | A Deep Learning Method for Near‐Real‐Time Cloud and Cloud Shadow Segmentation from Gaofen‐1 Images | |
| Li et al. | Spear and shield: Attack and detection for CNN-based high spatial resolution remote sensing images identification | |
| Acito et al. | CWV-Net: A deep neural network for atmospheric column water vapor retrieval from hyperspectral VNIR data | |
| CN110135309A (en) | Based on the shared SAR image change detection indicated of depth | |
| CN108229426B (en) | Remote sensing image change vector change detection method based on difference descriptor | |
| Fırat et al. | Hybrid 3D convolution and 2D depthwise separable convolution neural network for hyperspectral image classification | |
| El-Arafy et al. | Using edge detection techniques and machine learning classifications for accurate lithological discrimination and structure lineaments extraction: a comparative case study from Gattar area, Northern Eastern Desert of Egypt | |
| CN108154495A (en) | Across the sensor remote sensing images relative spectral alignment algorithms of multidate based on DCCAE networks | |
| CN109785302B (en) | Space-spectrum combined feature learning network and multispectral change detection method | |
| Zhang et al. | Cascaded attention-induced difference representation learning for multispectral change detection | |
| CN104851090B (en) | Image change detection method and device |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180612 |