CN108401565B - Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms - Google Patents
Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms Download PDFInfo
- Publication number
- CN108401565B CN108401565B CN201518002421.3A CN201518002421A CN108401565B CN 108401565 B CN108401565 B CN 108401565B CN 201518002421 A CN201518002421 A CN 201518002421A CN 108401565 B CN108401565 B CN 108401565B
- Authority
- CN
- China
- Prior art keywords
- point
- image
- registration
- pseudo
- remote sensing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of remote sensing image registration methods based on improved KAZE features and Pseudo-RANSAC algorithms.By the way that the line feature in Harr-like features is added and to corner characteristics, constructs new KAZE features, improves the characterization ability to Image neighborhood.Meanwhile reducing the error hiding rate of initial matching point pair.In addition, obtaining candidate matches point to collection using Delaunay Triangulation, recycles Pseudo-RANSAC algorithms to carry out " exterior point " and reject, further obtain stable transformation model parameter.Simulation result shows that the present invention to existing compared with high-gray level, translation, rotation, the remote sensing images for scaling difference, has stronger adaptability, while having higher registration accuracy height and preferable robustness on multi-source Remote Sensing Images data set.
Description
Technical field
The invention belongs to digital image processing methods, and in particular to one kind based on improved KAZE features and
The remote sensing image registration method of Pseudo-RANSAC algorithms.
Background technology
Remote sensing is the only resource for being capable of providing dynamic in global range at present and observing data, has continuity spatially
With temporal sequentiality.Meanwhile modern Remote Sensing Technical be going into one quickly, a variety of earth observations are provided in time
The new stage of data, the remotely-sensed data information that various sensors obtain is more rich and varied, this is resource investigation, environment
Monitoring etc. provides abundant data, and to constitute global change research due, environmental monitoring and assessment, disaster dynamic is supervised
It surveys and the multi-level application such as prevention.
Remotely-sensed data information is broadly divided into image format and non-image forms.Information is obtained from remote sensing images and is subject to profit
With being the final purpose of remote sensing work.With the fast development of digital image processing techniques, a variety of image processing techniques
Application on remote sensing images has been significantly greatly increased the type and quantity for obtaining information from remote sensing images, while also having extended
The application fields of remote sensing images.In these image processing techniques, image registration is both an independent research direction,
It is the basis of a variety of image processing techniques, such as image mosaic, image co-registration, image change detection, image oversubscription again
Resolution reconstruction etc..
Existing method for registering images is broadly divided into two classes:The image of method for registering images and feature based based on region
Method for registering.Method for registering images based on region is broadly divided into the method for registering images based on half-tone information and is based on frequency
The method for registering images of domain information.Method for registering images based on region, the image to there is smaller rotation, translation can
To obtain more accurate registration result, but for there are the registration between the image of larger geometric deformation, registration accuracy can
It can not reach requirement, or even can not be completed image registration task.In addition, be registrated for large-sized remote sensing images,
Requirement based on the method for registering in region to time and space is huge.Therefore, the remote sensing image registration of feature based obtains
It is widely applied.
Feature based image registration algorithm extraction image notable feature as registration primitive, point feature be exactly wherein it
One.Common feature point detecting method has:Harris Corner Detection Algorithms, FAST Corner Detection Algorithms, SIFT are special
Sign point detection algorithm, SURF feature point detection algorithms etc..Wherein SIFT feature detection algorithm and SURF characteristic points
Detection algorithm due to its good characterization ability, be typically used to image registration and splicing, image recognition, image classification,
The applications such as three-dimensional reconstruction.But SIFT and SURF algorithm are all based on linear gaussian pyramid and carry out multi-resolution decomposition
Come eliminate noise and extraction remarkable characteristic, this sacrifice local accuracy be cost, be easy to cause obscurity boundary and
Loss in detail.Document " Alcantarilla, Pablo Fem á ndez, Adrien Bartoli, and Andrew J.Davison.
" KAZE features. " Computer Vision-ECCV 2012.Springer Berlin Heidelberg, 2012.
214-227. " carries out Nonlinear Scale decomposition using additive operator splitting algorithm, and the edge for preferably maintaining image is special
Sign, and it is extracted the feature more stable than SIFT, this provides a kind of new mode for images match.Due to
KAZE features are responded only with Harr small echos horizontally and vertically, and statistical nature is too simple, can not
The local characteristics for accurately describing image cause feature error hiding to increase.In addition, initial matching point to collection there may be
" exterior point ", traditional RANSAC (RANdom SAmple Consensus, random sampling are consistent) algorithms due to
Iterations excessively cause inefficiency.Importantly, RANSAC algorithms random selection initial point can influence model
The stability of parameter so that registration Algorithm robustness reduces.
In conclusion traditional KAZE features and RANSAC algorithms exist centainly in terms of remote sensing image registration
Limitation, cannot obtain higher registration accuracy and robustness.
Invention content
In order to avoid the shortcomings of the prior art, the present invention propose it is a kind of based on improved KAZE features and
The remote sensing image registration method of Pseudo-RANSAC algorithms overcomes traditional KAZE features and RANSAC algorithms to exist
Limitation in terms of remote sensing image registration, the robustness for improving image registration accuracy, enhancing algorithm.
A kind of remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms, it is special
Sign is that steps are as follows:
Step 1:Benchmark image and image subject to registration are carried out using Additive Operator Splitting algorithms non-linear
Diffusing filter, and Nonlinear Scale sky is constructed using variable conduction method of diffusion respectively to benchmark image and image subject to registration
Between;Extreme point is detected in two Nonlinear Scale Space Theories, obtained extreme point is approached using the second Taylor series formula
Sub-pixel location is obtained, image and the precise and stable characteristic point position of image subject to registration on the basis of sub-pixel location;
Step 2:Centered on characteristic point, reference axis is rotated into principal direction first, one is chosen according to principal direction
20σi×20σiSquare window, and window is divided into 4 × 4 regions, each area size is 5 σi×5σiWindow;
It calculates the Harr small echos per sub-regions with respect to principal direction to respond, every sub-regions are all 1.5 σ with a Gaussian kerneli's
Gaussian function is weighted;At this point, forming a seed point d per sub-regionsv, description is that 4*4*8=128 is tieed up
dv=(∑ Lx, ∑ | Lx|, ∑ Ly, ∑ | Ly|, ∑ L1, ∑ | L1|, ∑ Ld, ∑ | Ld|)
Wherein:dvEach dimension be followed successively by horizontal direction Haar small echos response and its response absolute value and, hang down
Histogram to Harr small echos response and its response absolute value and, the absolute value of line characteristic response and its response and, diagonally
Response and its response absolute value and;
The improvement KAZE Feature Descriptors of benchmark image and image subject to registration are respectively labeled as VrAnd Vs;
Step 3:To each feature V of benchmark imageri, distance is found in the characteristic set of image subject to registration recently
Adjacent featureSecondary neighbour's featureJudge that matching condition threshold value T is:
Ranging from the 0.5~0.8 of the threshold value T obtains initial matching point to collection;
Step 4:Error hiding pair is rejected using Pseudo-RANSCA algorithms, and affine change is obtained by least square method
Matrix is changed, process is as follows:
A. random selection set MPkIn 3 pairs of corresponding dot pairs, ensure 3 pairs of points it is not conllinear in respective space, it is no
Then choose again;
B. according to the corresponding dot pair of selection, with the parameter of least-squares estimation transformation model H;
C. in remaining point to middle selection i-th pair matching double points (Pi, Pi'), if meeting formula | | Pi-T(Pi′)||2≤ ε,
The definition of middle T () such as formula (15), then by this to candidate matches point to being defined as " interior point ";
D. both the above step is repeated, until iterations reach maximum sampling number Nmax.Before iteration stopping,
" interior point " is updated with probability ω every time;
E. most points is counted out in selection to collection, calculating transformation model parameter is as final result, i.e., best
Transformation matrixWherein:a1, a2, a3, a4, a5, a6It is the six-freedom degree parameter of plane affine transformation,
Also it is to solve for the parameter of transformation model;
T (P)=HP wherein, P=(x, y, 1)T
The parameter setting is as follows:Distance threshold ε=1, confidence level p=0.99 update " interior point " probability ω=0.4, most
Big sampling number Nmax=log (1-p)/log (1- ω4);
Step 5:Image subject to registration is mapped to reference space by affine transformation, and is obtained finally using interpolation algorithm
Registration result.
The T selections 0.65.
A kind of remote sensing images based on improved KAZE features and Pseudo-RANSAC algorithms proposed by the present invention are matched
Quasi- method, KAZE features are that Alcantarilla was proposed in 2012, are protected well using Nonlinear Scale decomposition
Held the edge feature of image, and all layers of KAZE features with original image resolution ratio having the same.By
It is responded only with Harr small echos horizontally and vertically in KAZE features, statistical nature is too simple, therefore
The new feature point description period of the day from 11 p.m. to 1 a.m is being constructed, line feature in Harr-like features is added and to corner characteristics, is improving spy
The characterization ability of sign.Concentration may be deposited due to obtaining initial matching point using improvement KAZE features progress characteristic matching
At " exterior point ", the present invention carries out subdivision to initial matching point using Delaunay Triangulation to collection, at the beginning of obtaining candidate
Then beginning matching double points collection carries out model parameter estimation to it using Pseudo-RANSAC algorithms, not only increases
Image registration accuracy, and enhance the stability of registration Algorithm.
The beneficial effects of the invention are as follows:KAZE feature point detection algorithms are improved, image side not only can be preferably kept
Edge feature, and the feature more stable than SIFT can be extracted.Simultaneously as adding Harr-like feature sets
In line feature and to corner characteristics, can preferably characterize the neighborhood information of image in this way so that initial matching point to collection
More stablize.In addition, Mismatching point rejecting is carried out using Pseudo-RANSAC algorithms, on the basis of pretreated
It selects Space Consistency and the relatively good initial matching point of continuity to subset, directly effectively reduces model iteration time
Number shortens the calculating time, and improves the stability and accuracy of model parameter estimation.Therefore, for remote sensing
Image registration problem, the present invention have higher registration accuracy and stronger robustness.
Description of the drawings
Fig. 1 is that the present invention is based on the remote sensing images flow charts for improving KAZE features and Pseudo-RANSAC algorithms.
Specific implementation mode
In conjunction with embodiment, attached drawing, the invention will be further described:
1) the remote sensing images L (benchmark image or image subject to registration) for being directed to input utilizes Nonlinear diffusion filtering method structure
Nonlinear Scale Space Theory is built, is described with following nonlinear partial differential equation:
Wherein, div andRespectively represent divergence and gradient operator.C (x, y, t) indicates propagation function, and x and y are indicated respectively
The coordinate position of image level and vertical direction, t indicate chronomere.C (x, y, t) is defined as follows:
Original image L is indicated by the gradient image after Gaussian smoothing, σ indicates the Gaussian smoothing factor (present invention
σ=1.0).Function g () is as follows:
The value of parameter k is gradient image70% percentile of histogram on value.
Following equation is obtained to formula (1) discretization:
Wherein, Al() indicates conductance matrix of the image on dimension l, and m is tieed up in total, and the matrix is diagonal dominance
Triple diagonal matrix.By obtaining formula (5) to formula (4) abbreviation, it indicates i+1 tomographic image in Nonlinear Scale Space Theory
Li+1, and such linear system can pass through Thomas algorithm rapid solvings.
I indicates original graph, tiThe chronomere for indicating the i-th tomographic image, the scale factor σ with respective layeriRelationship is as follows:
Wherein, scale factor σiIt is defined as follows:
σi(o, s)=σ 02o+s/SO ∈ [0 ..., O-1], s ∈ [0 ..., S-1], i ∈ [0 ..., OS-1] (7)
Nonlinear Scale Space Theory shares O groups, every group S layers (generally 3~5 layers, 4 layers are taken in of the invention).O and s is indicated
Corresponding with i group of number and the number of plies.The group number of Nonlinear Scale Space Theory is determined by picture size:
O=log 2 { min (M, N) } (8)
M and N indicates the line number and columns of image.
By above step, an O group, the Nonlinear Scale Space Theory of every group of S tomographic image can be obtained.
2) extreme point is selected in Nonlinear Scale Space Theory, it is 26 points in 3 × 3 neighborhood of present image domain and scale domain
Maximum or minimum.Obtained extreme point is approached to obtain sub-pixel location using the second Taylor series formula:
Assuming that i layers of scaling function is Li(x, y, σi), second order Taylor expansion is as follows:
Wherein,It is the offset of characteristic point, T indicates transposition, and the position of characteristic point should be's
Extreme point obtains, and is obtained to the derivation of above formula both sides:
In order to make the characteristic point detected that there is stability, the non-linear space functional value at extreme point is utilizedIt rejects
The characteristic point of low contrast.Formula (10) is substituted into formula (9), takes first two can obtain:
To all extreme points, ifThis feature point just remains, and otherwise abandons.It thus can be with
Obtain more accurate degree characteristic point position.
3) after characteristic point is accurately positioned, for chronomere tiCorresponding scale parameter σiCharacteristic point, with feature
Centered on point, reference axis is rotated into principal direction first, 20 σ are chosen according to principal directioni×20σiSquare window,
And window is divided into 4 × 4 regions, each area size is 5 σi×5σiWindow, calculate relatively main per sub-regions
The Harr small echos in direction respond, and every sub-regions are all 1.5 σ with a Gaussian kerneliGaussian function be weighted.This
When, a seed point d is formed per sub-regionsv, such a Feature Descriptor is 4*4*8=128 dimensions.
dv=(∑ Lx, ∑ | Lx|, ∑ Ly, ∑ | Ly|, ∑ Ll, ∑ | Ll|, ∑ Ld, ∑ | Ld|) (12)
Wherein, dvEach dimension be followed successively by horizontal direction Haar small echos response and its response absolute value and, Vertical Square
To Harr small echos response and its response absolute value and, the absolute value of line characteristic response and its response and, to angular response
And its response absolute value and.
4) the improvement KAZE Feature Descriptors of benchmark image and image subject to registration are respectively labeled as VrAnd Vs.To benchmark
Each feature V of imageri, found apart from arest neighbors feature in the characteristic set of image subject to registrationSecondary neighbour
FeatureJudge that matching condition is:
Normal conditions, ranging from the 0.5~0.8 of threshold value T, the present invention in T selection 0.65, to obtain initial matching point pair
Collection.
5) it uses Delaunay Triangulation to carry out subdivision to collection to initial matching point, realizes to initial matching point to collection
Screening, it is specific as follows:Assuming that a pair of of same place P and P ' of the initial matching point to concentration, the adjacent triangle of each
Number of vertices be not N0With N '0, the point P adjacent with P1, P2..., PmThe point P adjacent with P '1', P2' ..., Pm' be respectively
Initial point matching double points share m pairs.
Choose N0With N '0In it is maximum be denoted as N, enable β=m/N.
As β > β0When, construct following vectorial V=[v1, v2..., vm], wherein viIt is defined as follows:
Calculate the variances sigma of vector V.As σ < σ0When, then P and P ' this space structure consistency is met to same place,
So data element to corresponding dot pair as new matching double points collection.According to the method described above, initial matching point pair is traversed
Every a pair of of the same place concentrated, is added to the corresponding dot pair for meeting above-mentioned condition new data set, finally obtains new
Input of the matching double points subset as Pseudo-RANSAC.In the present invention, β0=0.25, σ0=0.50.
After Delaunay Triangulation, recalculate the variances sigma of each vector V, and preserve this to same place and
The point pair of its neighborhood is stored in corresponding set MP.Variance is ranked up, selects minimum variance σ corresponding
Set MP.If the matching logarithm of set MP is less than three pairs, the corresponding set of time small variance is selected, successively
The corresponding set of selection.If final choice variances sigmakCorresponding set MPk, Pseudo-RANSAC algorithm steps are as follows:
A) random selection set MPkIn 3 pairs of corresponding dot pairs, ensure 3 pairs of points it is not conllinear in respective space, otherwise weigh
It is new to choose;
B) according to the corresponding dot pair of selection, with the parameter of least-squares estimation transformation model H;
C) in remaining point to middle selection i-th pair matching double points (Pi, Pi'), if meeting formula | | Pi-T(Pi′)||2≤ ε, wherein
T () definition such as formula (15), then by this to candidate matches point to being defined as " interior point ";
D) both the above step is repeated, until iterations reach maximum sampling number Nmax.Before iteration stopping, often
It is secondary that " interior point " is updated with probability ω;
E) most points is counted out in selection to collection, calculates transformation model parameter as final result, i.e., best transformation
MatrixHere a1, a2, a3, a4, a5, a6It is the six-freedom degree parameter of plane affine transformation, and
Solve the parameter of transformation model.
T (P)=HP wherein, P=(x, y, 1)T (15)
Parameter setting is as follows in the present invention:Distance threshold ε=1, confidence level p=0.99 update " interior point " probability ω=0.4,
Maximum sampling number Nmax=log (1-p)/log (1- ω4)。
Image subject to registration is mapped to reference space by affine transformation, and is finally matched using bilinear interpolation algorithm
Quasi- result.
Claims (2)
1. a kind of shaking sense method for registering images, feature based on improved KAZE features and Pseudo-RANSAC algorithms
It is that steps are as follows:
Step 1:Benchmark image and image subject to registration are carried out using Additive Operator Splitting algorithms non-linear
Diffusing filter, and Nonlinear Scale sky is constructed using variable conduction method of diffusion respectively to benchmark image and image subject to registration
Between;Extreme point is detected in two Nonlinear Scale Space Theories, obtained extreme point is approached using the second Taylor series formula
Sub-pixel location is obtained, image and the precise and stable characteristic point position of image subject to registration on the basis of sub-pixel location;
Step 2:Centered on characteristic point, reference axis is rotated into principal direction first, one is chosen according to principal direction
20σi×20σiSquare window, and window is divided into 4 × 4 regions, each area size is 5 σi×5σiWindow;
It calculates the Harr small echos per sub-regions with respect to principal direction to respond, every sub-regions are all 1.5 σ with a Gaussian kerneli's
Gaussian function is weighted;At this point, forming a seed point d per sub-regionsv, description is that 4*4*8=128 is tieed up
dv=(∑ Lx, ∑ | Lx|, ∑ Ly, ∑ | Ly|, ∑ Ll, ∑ | Ll|, ∑ Ld, ∑ | Ld|)
Wherein:σiFor chronomere tiCorresponding scale parameter, dvEach dimension be followed successively by the Harr of horizontal direction
Small echo response and its absolute value responded and the absolute value of the response of Harr small echos and its response of vertical direction and line spy
The absolute value of sign response and its response and, to the absolute value of angular response and its response and;
The improvement KAZE Feature Descriptors of benchmark image and image subject to registration are respectively labeled as VrAnd Vs;
Step 3:To each feature V of benchmark imageri, distance is found in the characteristic set of image subject to registration recently
Adjacent featureSecondary neighbour's featureJudge that matching condition threshold value T is:
Ranging from the 0.5~0.8 of the threshold value T obtains initial matching point to collection;
Step 4:Mismatching point pair is rejected using Pseudo-RANSAC algorithms, and is obtained by least square method affine
Transformation matrix, process are as follows:
A. random selection set MPkIn 3 pairs of corresponding dot pairs, ensure 3 pairs of corresponding dot pairs in respective space not its
Otherwise line is chosen again;
B. according to the corresponding dot pair of selection, with the parameter of least-squares estimation transformation model H;
C. i-th pair matching double points (P is selected in remaining corresponding dot pairi, Pi'), if meeting formula
||Pi-T(Pi′)||2≤ ε, wherein T () definition such as formula (15), then by this to candidate matches point to being defined as " interior point ";
D. both the above step is repeated, until iterations reach maximum sampling number Nmax;Before iteration stopping,
" interior point " is updated with probability ω every time;
E. it counts out most matching double points in selection, calculates transformation model parameter as final result, i.e., most preferably
Transformation matrixWherein:a1, a2, a3, a4, a5, a6It is the six-freedom degree ginseng of plane affine transformation
Number, is also to solve for the parameter of transformation model;
T (P)=HP wherein, P=(x, y, 1)T (15)
The parameter setting is as follows:Distance threshold ε=1, confidence level p=0.99 update " interior point " probability ω=0.4, most
Big sampling number Nmax=log (1-p)/log (1- ω4);
Step 5:Image subject to registration is mapped to reference space by affine transformation, and is obtained finally using interpolation algorithm
Registration result.
2. the remote sensing images according to claim 1 based on improved KAZE features and Pseudo-RANSAC algorithms are matched
Quasi- method, it is characterised in that:The T selections 0.65.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201518002421.3A CN108401565B (en) | 2015-05-28 | 2015-05-28 | Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201518002421.3A CN108401565B (en) | 2015-05-28 | 2015-05-28 | Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108401565B true CN108401565B (en) | 2017-12-15 |
Family
ID=63078679
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201518002421.3A Active CN108401565B (en) | 2015-05-28 | 2015-05-28 | Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108401565B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108109153A (en) * | 2018-01-12 | 2018-06-01 | 西安电子科技大学 | SAR image segmentation method based on SAR-KAZE feature extractions |
CN108680900A (en) * | 2018-05-21 | 2018-10-19 | 武汉科技大学 | A kind of ESPRIT sound localization methods based on RANSAC |
CN109544608A (en) * | 2018-03-22 | 2019-03-29 | 广东电网有限责任公司清远供电局 | A kind of unmanned plane Image Acquisition feature registration method |
CN111104983A (en) * | 2019-12-20 | 2020-05-05 | 河北科技大学 | AKAZA-based feature matching method for satellite remote sensing image pair |
CN112381864A (en) * | 2020-12-08 | 2021-02-19 | 兰州交通大学 | Multi-source multi-scale high-resolution remote sensing image automatic registration technology based on antipodal geometry |
CN113470085A (en) * | 2021-05-19 | 2021-10-01 | 西安电子科技大学 | Image registration method based on improved RANSAC |
-
2015
- 2015-05-28 CN CN201518002421.3A patent/CN108401565B/en active Active
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108109153A (en) * | 2018-01-12 | 2018-06-01 | 西安电子科技大学 | SAR image segmentation method based on SAR-KAZE feature extractions |
CN109544608A (en) * | 2018-03-22 | 2019-03-29 | 广东电网有限责任公司清远供电局 | A kind of unmanned plane Image Acquisition feature registration method |
CN109544608B (en) * | 2018-03-22 | 2023-10-24 | 广东电网有限责任公司清远供电局 | Unmanned aerial vehicle image acquisition characteristic registration method |
CN108680900A (en) * | 2018-05-21 | 2018-10-19 | 武汉科技大学 | A kind of ESPRIT sound localization methods based on RANSAC |
CN111104983A (en) * | 2019-12-20 | 2020-05-05 | 河北科技大学 | AKAZA-based feature matching method for satellite remote sensing image pair |
CN112381864A (en) * | 2020-12-08 | 2021-02-19 | 兰州交通大学 | Multi-source multi-scale high-resolution remote sensing image automatic registration technology based on antipodal geometry |
CN113470085A (en) * | 2021-05-19 | 2021-10-01 | 西安电子科技大学 | Image registration method based on improved RANSAC |
CN113470085B (en) * | 2021-05-19 | 2023-02-10 | 西安电子科技大学 | Improved RANSAC-based image registration method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108401565B (en) | Remote sensing image registration method based on improved KAZE features and Pseudo-RANSAC algorithms | |
CN110929607B (en) | Remote sensing identification method and system for urban building construction progress | |
CN104008538B (en) | Based on single image super-resolution method | |
US9342870B2 (en) | Tree-based linear regression for denoising | |
CN105869146B (en) | SAR image change detection based on conspicuousness fusion | |
CN108960404B (en) | Image-based crowd counting method and device | |
Ma et al. | Accurate dense stereo matching based on image segmentation using an adaptive multi-cost approach | |
CN107170001A (en) | Method and apparatus for carrying out registration to image | |
Qin et al. | Spatiotemporal inferences for use in building detection using series of very-high-resolution space-borne stereo images | |
Chen et al. | An improved multi-resolution hierarchical classification method based on robust segmentation for filtering ALS point clouds | |
Zhu et al. | Stereo matching algorithm with guided filter and modified dynamic programming | |
Ravanbakhsh et al. | A comparative study of DEM registration approaches | |
Chen et al. | A color-guided, region-adaptive and depth-selective unified framework for Kinect depth recovery | |
CN112907557A (en) | Road detection method, road detection device, computing equipment and storage medium | |
CN106934398B (en) | Image de-noising method based on super-pixel cluster and rarefaction representation | |
Song et al. | Small UAV-based multi-temporal change detection for monitoring cultivated land cover changes in mountainous terrain | |
Song et al. | Small UAV based multi-viewpoint image registration for monitoring cultivated land changes in mountainous terrain | |
Janardhana Rao et al. | Image inpainting method with improved patch priority and patch selection | |
CN108447084B (en) | Stereo matching compensation method based on ORB characteristics | |
CN108509835B (en) | PolSAR image ground object classification method based on DFIC super-pixels | |
CN110517300A (en) | Elastic image registration algorithm based on partial structurtes operator | |
Chen et al. | Change detection of multispectral remote-sensing images using stationary wavelet transforms and integrated active contours | |
Yan et al. | HR optical and SAR image registration using uniform optimized feature and extend phase congruency | |
Hao et al. | Superpixel-based active contour model for unsupervised change detection from satellite images | |
CN109726679B (en) | Remote sensing classification error spatial distribution mapping method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
GR03 | Grant of secret patent right | ||
GR03 | Grant of secret patent right | ||
DC01 | Secret patent status has been lifted | ||
DC01 | Secret patent status has been lifted |