CN104361590B - High-resolution remote sensing image registration method with control points distributed in adaptive manner - Google Patents
High-resolution remote sensing image registration method with control points distributed in adaptive manner Download PDFInfo
- Publication number
- CN104361590B CN104361590B CN201410637686.5A CN201410637686A CN104361590B CN 104361590 B CN104361590 B CN 104361590B CN 201410637686 A CN201410637686 A CN 201410637686A CN 104361590 B CN104361590 B CN 104361590B
- Authority
- CN
- China
- Prior art keywords
- image
- control point
- registration
- value
- pixel
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
- G06T7/337—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods involving reference images or patches
-
- 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
- G06T2207/20004—Adaptive image processing
- G06T2207/20008—Globally adaptive
Landscapes
- Engineering & Computer Science (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 high-resolution remote sensing image registration method with control points distributed in an adaptive manner. By the use of the high-resolution remote sensing image registration method, a multi-scale J-image is introduced into image registration. A self-adaptive control points extracting method based on a partitioning strategy is provided to extract the control points of the multi-scale J-image image, thus the defect that the control points only sense specific directional high-frequency information is overcome. A self-adaptive control points extracting strategy is defined so as to limit control point distribution. The control points are subjected to multi-scale matching by NMI (normalized mutual information measure), allowing registration function smooth. Geometric correction is realized by adopting Delaunay triangle local transformation. In the following text, basic principle and key steps of the algorithm are introduced, and three remote-sensing image groups of different types are subjected to experiment and analysis and comparison experiments with various registration methods based on wavelet transform and NSCT (non-subsampled contourlet transform).
Description
Technical field
Belong to the present invention relates to a kind of high-resolution remote sensing image method for registering for controlling point self-adapted distribution, remote sensing image
Registration technique field.
Background technology
The basic framework of 1 image registration
With the extensive use of high-resolution remote sensing image, the image registration research for high-resolution remote sensing image turns into
One important topic of remote sensing application research field.The basic framework of image registration mainly includes the content of the following aspects:
(1) feature space, refers to, with reference to image and the characteristics of image of extraction in image subject to registration, mainly to include that gray scale is special
Levy, point feature, edge feature and not bending moment architectural feature etc..
(2) search space, refer to image registration in geometry compared with the scope of orthochronous transformation and by the way of.Wherein, convert
Scope includes global, part and displacement field three types.Mapping mode is divided into linear change and nonlinear transformation, and linear transformation is again
Rigid body translation, projective transformation and affine transformation three types can be divided into.
(3) similarity measurement, for the criterion that the quality to transformation results is evaluated.Common similarity measurement has mutually
Information, normalized mutual information, Euclidean distance, combination entropy, gradient cross correlation etc..It has been generally acknowledged that feature extraction and similarity measurement
Directly determine the robustness of registration Algorithm.
(4) search strategy, for finding optimal registration parameter in search space, and using similarity measurement as judge
Foundation.Common search strategy is including ant group algorithm, genetic algorithm, parabolic method, Powell methods, Newton method etc..
The classification of 2 image registrations
The also ununified classifying rules of current method for registering images, can be divided mainly into the figure based on gray scale under normal circumstances
As the image registration of registration and feature based.
Method for registering based on gray scale directly sets up the space with reference to image with image subject to registration using the gray value of image
Transformation relation, its basic thought is:Geometric transformation is carried out to image subject to registration first, is then set up and is referred to image and shadow subject to registration
Similarity measurement as between, i.e., define an object function on the basis of image greyscale information is counted.By the object function
As the object function for evaluating the good and bad criterion of registration result and geometrical registration parameter optimization so that Image registration is asked
Topic is converted into the extreme-value problem for seeking the object function.Finally, optimal registration parameter is determined using a series of optimal methods.
Although the method for registering based on gray scale is directly perceived and is easily achieved, calculate usual with template institute during similarity measurement
It is in place to be set to center, thus the peak face for being formed is more gentle, it is difficult to obtain accurate registration result.On the other hand, due to existing
Illumination condition change, the not equal factor of sensor type, larger ash is commonly present in multidate high-resolution remote sensing image registration
The differences such as degree, rotation, the method for registering based on gray scale is poor to the robustness of these disturbing factors.And the registration side of feature based
Method can effectively overcome these problems, therefore the main registration side using feature based of high-resolution remote sensing image registration at present
Method.
By the use of the feature of image as registering foundation, these features should have to gray scale change the method for registering of feature based
Change and the characteristics of spatial alternation is insensitive, so that method has preferably repeatability and stability, common feature includes
Angle point, edge feature etc..The method for registering basic step of feature based is:
(1) image is pre-processed, with reduce or eliminate with reference to the geometric deformation between image and image subject to registration with
And gray-scale deviation, reduce the influence to registration process.
(2) select or define suitable feature, Automatic signature extraction can also be realized using feature extraction operators such as SIFT.
Matched according to Corner Feature (control point), should also make control point that obvious terrain and its features is identified in image, and protect
The certain quantity of card and rational distribution.
(3) select or define suitable matching strategy, the dominating pair of vertices of condition will be met as a pair of same places, obtain institute
The point set of the same name for needing.
(4) set up with reference to the mapping relations between image and image subject to registration, and according to coordinate and reference according to point set of the same name
Image carries out resampling to image subject to registration, obtains final registration result.
In the method for registering of various feature baseds, method for registering application and research based on point feature are the most extensive.Its
Basic thought is to extract the control point in image first, and then feature and the control such as the local gray level based on control point, gradient
Geometrical relationship between point obtains the set of same place pair, so as to realize the complete registration between image.The advantage is that only to control
System point is matched, so as to substantially reduce amount of calculation compared with all pixels are calculated;Control point can in extraction process
The influence of gray difference, geometric deformation, noise and other disturbing factors is effectively reduced, thus with good robustness;Base
In control point similarity measurement to position sensitive, registration accuracy can be effectively improved;Point set of the same name is used directly for meter
Calculate with reference to the spatial transform relation between image and image subject to registration.Based on these advantages, the Image registration method based on point feature
Fast and accurately high precision image registration usually can be realized, is that the image of current main flow is matched somebody with somebody while having wide applicability
Quasi- method, while also represent the developing direction of image registration techniques.
3 have problem
Influence for geometrical registration error to change accuracy of detection, has many scholars and expands deep grinding at present
Study carefully.Research such as Dai et al. shows that the location matches of mistake can cause " crossing change detection " between image.And Verbyla et al. is then
Think, due to inevitably there is residual error in image rectification model, therefore almost without completely accurate Image Matching, this is just
" the puppet change " that result in testing result is inevitable outcome of the geometrical registration error in detection is changed.In Townshend etc.
People carries out the change of difference comparsion to multidate NDVI (Normalized Difference Vegetation Index) image
In detection, at least 40% is caused due to the other registration error of sub-pixel.And the achievement in research of Dai et al. shows, in,
In the change detection of low resolution remote sensing image, only registration accuracy reaches 1/5th pixels can obtain 90% change inspection
Survey precision.And high-resolution remote sensing image change detection is easier to be influenceed by registration error, therefore no matter using which kind of tool
Body change detecting method, accurate geometrical registration is all successful precondition and guarantee.
With the continuous improvement of remote sensing image spatial resolution, remote sensing image provides more abundant characteristic information and is used to
Expression atural object.On the other hand, compared with middle low resolution remote sensing image, high-resolution remote sensing image different time scales (when
Between yardstick be also referred to as temporal resolution, refer to the time interval between the continuous acquisition image of same position) in characters of ground object
As the state difference produced by seasonal variations is more protruded.Simultaneously as sensor internal imaging system and hypsography,
Cloud cover, satellite orbit change, attitude of flight vehicle is unstable etc., and the disturbing factor such as nonlinear distortion for causing is caused to registration
Influence it is also more obvious.For these problems, the method for registering based on point feature can provide an effective solution route.
By based on point feature method for registering basic thought and realize flow as can be seen that control point reasonable selection in feature space and
Accurately mate between control point is to improve the key factor of registration accuracy.
According to the characteristics of high-resolution remote sensing image, it is obtained in that more with matching using the Control point extraction under multiple dimensioned
Reliable point set of the same name, thus be more and more taken seriously.Multiscale analysis instrument common at present mainly becomes including small echo
Change (Wavelet Transform), profile wave convert (Contourlet Transform) [106], non-down sampling contourlet change
(NSCT, Non-Subsampled Contourlet Transform) [107] etc. are changed, for example I.Zavorin et al. proposes one
The high-resolution remote sensing image registration Algorithm based on wavelet pyramid is planted, can obtain good in Multiple Source Sensor Image registration
Effect.The selection such as Z.G.Chen NSCT carries out multiple dimensioned Control point extraction and homotopy mapping, significantly improves registration accuracy.
Although multiscale analysis instrument is conducive to more accurately extracting the feature of image, the method based on wavelet transformation would generally be suddenly
Slightly 45 ° of directional subband information, and without translation invariance;And the method for being based on profile wave convert and NSCT equally exist it is right
Set direction sensitive issue, the i.e. high-frequency sub-band of selection different directions can directly affect the registration accuracy of image, reduce calculation
The stability and robustness of method.On the other hand, when ground control point is selected, following several factors are considered as:Scene should be selected
In there is the position of notable witness marker, intersection, the building border of such as road etc.;It is ensured that the corresponding atural object of same place
Do not change with time and change;Ground control point should be made reasonably to be distributed in whole scene, in order to avoid in control point compact district
It is absorbed in local optimum.Especially in high-resolution remote sensing image, by atmospheric refraction, hypsography, earth curvature etc. it is many because
, generally there is irregular local deformation in scene in the influence of element.Registration is made to ensure registration accuracy and reducing local deformation
Into influence, inner vein feature rich, baroque object should using more same place to being marked;And for line
Reason feature is single, regional area of simple structure extract a small amount of same place to.Existing multiple dimensioned registration Algorithm constraint control
During point distribution, it is usually chosen in and a number of control point is extracted in a certain specific dimensions window, therefore, it is difficult to ensure same place
The reasonable layout of collection.Finally, the Ground control point matching strategy employed in the Image registration of point feature is currently based on generally only to consider
The gray-scale statistical characteristics of regional area, and have ignored the features such as texture, yardstick in raw video, be easily subject to noise etc. because
The interference of element, reduces the reliability of point set of the same name, increased the possibility of error hiding.
The content of the invention
Goal of the invention:For problems of the prior art, the invention provides a kind of point self-adapted distribution of control
High-resolution remote sensing image method for registering, mainly includes following three step:(1) multiple dimensioned Control point extraction.(2) based on NMI
Homotopy mapping.(3) geometric correction based on Delaunay triangle partial transformations.
Technical scheme:A kind of high-resolution remote sensing image method for registering of the point self-adapted distribution of control, by multiple dimensioned J-
Image is incorporated into image registration, first proposed a kind of Self Adaptive Control point extracting method based on partition strategy, is realized
Control point extraction in multiple dimensioned J-image images, the office only sensitive to indivedual direction high-frequency informations so as to overcome control point
Limit.Meanwhile, define Self Adaptive Control point and extract strategy to constrain the distribution at control point.And then surveyed using normalized mutual information
Degree NMI control point is carried out it is multiple dimensioned match, can effectively smooth match somebody with somebody quasi-function.Finally using Delaunay triangles office
Geometric correction is realized in portion's conversion.The general principle and committed step of proposed algorithm are described initially below, and then to three groups not
The remote sensing image of same type is tested and analyzed, and is carried out from the different method for registering based on wavelet transformation and NSCT respectively
Comparative experiments.
The method for registering that JSEG is combined with NMI
Multiple dimensioned Control point extraction
Selection J-image images are controlled an extraction and match as multiscale analysis instrument.
1 calculates J-image image sequences
Using the corresponding local homogeney index J- of each pixel in the window calculation raw video of a certain specific dimensions
Value and as the pixel value of the pixel, can obtain the J-image images of single yardstick.The calculating process of J-value is such as
Under:
Quantization method (Deng Y, Kenney C, Moore M S, the et al.Peer for being proposed using Deng et al. first
group filtering and perceptual color image quantization[C]//Proceedings of
the 1999 IEEE International Symposium on Circuits and Systems,1999,4:21-24.)
Raw video is quantified so as to obtain quantification image.Position z (x, y) for making each pixel in quantification image is pixel z
Pixel value, z (x, y) ∈ Z.Z is the set of all pixels composition in specific dimensions window centered on pixel z.
In quantification image, it is the sum of all pixels centered on z in window to define N, then average m:
Define mpTo belong to all pixels average of same grey level p, Z in windowpTo belong to the institute of gray level p in window
There is the set of pixel, P is the gray level sum in quantification image, then belong to the variance and S of same gray-level pixels in windowW
May be defined as:
Define STIt is the population variance of all pixels in window:
Then J-value is:
J=(ST-SW)/SW (2.4)
Calculate the J-value of z under obstructed yardstick respectively using different size window, and as the pixel value of z, so as to obtain
Multiple dimensioned J-image image sequences.According to the definition of above J-value, in the J-image of a certain yardstick, a certain pixel
Corresponding J-value values are bigger, then be more likely to be at the edge of object;Conversely, being then likely located at the center of object.Meanwhile, J-
Value combines spectral information, texture information and the dimensional information that raw video is included, and insensitive to directional information, because
This can also be used for the matching of same place as description at control point.
2 Control point extractions and distribution constraint
Definition and feature according to J-value, it is proposed that a kind of Self Adaptive Control point extracting method based on partition strategy,
Its core concept is the more control points of extracted region enriched in textural characteristics, realizes that flow is as follows:
Step1:The J-image images of each yardstick are divided into various sizes of subgraph, certain yardstick neutron image is big
It is small identical with certain window size that is calculating yardstick J-image images, but retain angle point, it is the square of N × N pixels
Window.
Step2:The J-value values of view picture image are calculated using formula (2.4) in raw video, T is defined asa。TaReflection
The overall homogeneous degree of raw video.
Step3:In a certain yardstick J-image, respectively by the J-valve and T of each subgraph center pixelaCompared
Compared with if being more than Ta, then it is assumed that the subgraph inner vein feature rich, complex structure should extract more control point, take subgraph
J-value maximum preceding K pixel is used as control point as in;Conversely, then extracting a small amount of control point, J- in subgraph is taken
Value maximum preceding L pixel meets as control pointThe value of K and L can be according to image actual size and shadow
As the complexity of texture manually sets.
Step4:In all yardstick J-image images repeat Step3 operation, realize control point multiple dimensioned extraction and
Distribution constraint.
Ground control point matching based on NMI
Complete it is multiple dimensioned under Control point extraction after, it is necessary to further with similarity measurement to each control point one by one
Differentiate and obtain point set of the same name.Similarity measurement between control point generally uses the gray scale of the regional area centered on control point to unite
Meter characteristic is used as similarity measure.Therefore, the control point similarity measurement based on topography's gray feature can be used
The general method for registering images based on gray feature.
Mutual information (MI, Mutual Information) is wide variety of a kind of reversibility phase in current image registration
Estimate like property, the features such as with without pre-segmentation, strong robustness and high precision.Due to that need not be pre-processed, even if in figure
In having an application scenario of excalation as data, good registration result can be still obtained, simultaneously can be used for multi-source remote sensing
In the registration of image.On the other hand, mutual information similarity measure is excessively sensitive to the application scenario of image overlapping region, and works as
When overlapping region is reduced, it is meant that less pixel take part in the statistics of mutual information, so as to reduce the reliability of registration.For gram
Take normalized mutual information measure (NMI, the Normalized Mutual of the propositions such as this limitation, Studholme C
Information) can effectively smooth with quasi-function, mutual information makes to the sensitivity of overlapping region in reducing image registration
The relation that object function more accurately reflects between registration parameter and mutual information, so as to improve registration accuracy.Therefore, herein
Selection NMI extracts the same place pair under each yardstick as control point similarity measurement, obtains global point set of the same name.
1 calculates NMI similarity measurements
In a certain yardstick J-image, using the control point region in reference image and image subject to registration as two
Vector C and D, calculate the NMI between control point.The window size for being used is specific with what calculating current scale J-image was used
Window size is identical.Mutual information is described with entropy first, entropy refers to the general characteristic characterized to information source on average, can be led to
Cross and possible temporal information content weighted average is obtained.Define entropy H and mutual information I (C, D) such as formula (2.5), (2.6) institute
Show:
I (C, D)=H (C)+H (D)-H (C, D) (2.6)
Wherein, PiRepresent certain stochastic variable be possible in the probability that occurs of i-th kind of possible situation.Then H (C), H (D),
H (C, D) is respectively with reference to the corresponding entropy of same place and their combination entropy in image and image subject to registration.According to formula
(2.5) understand:
Wherein PC(c) and PDD () is respectively probability distribution when C and D is completely independent, PCDIt is the joint probability point of C and D
Cloth.NMI is defined as follows:
In with reference to the same yardstick J-image of image and image subject to registration, centered on control point, to calculate current chi
The certain window for spending J-value is regional area spanning subgraph picture, calculates the NMI between subgraph so as to measure the phase between control point
Like property.Although it is pointed out that the process for calculating NMI is only united to the local gray level feature of control point region
Meter, but because the gray value of each pixel in J-image images is the corresponding J-value values of each pixel in raw video, because
This it can be seen from the definition of J-value, the ash of regional area where the NMI similarity measurement concentrated expressions that are calculated control point
The similitude of degree, texture and spectral signature, so that relatively reliable.
The 2 maximum bi-directional matching strategies based on NMI
To realize the accurate match between control point, it is proposed that a kind of maximum bi-directional matching strategy based on NMI, matching process
It is as follows:
Step1:In with reference to same yardstick J-image of the image with image subject to registration, NMI is maximum and more than threshold value TNMI
Dominating pair of vertices as a pair of same places, and then obtain point set of the same name 1.
Step2:With Step1 calculated directions conversely, by a certain control point in image subject to registration respectively with reference to institute in image
There is control point to be compared, obtain point set of the same name 2.
Step3:Compare point set of the same name 1 and point set of the same name 2, using identical same place to being obtained most as under current scale
Whole point set of the same name.
Step4:Step1 to Step3 is repeated, all yardstick J-image are differentiated, and the same place that will be obtained is to complete
Portion is mapped to the same position in raw video.
Step5:It is further using classical RANSAC (Random Sample Consensus) algorithm in raw video
Error hiding phenomenon is eliminated, final point set of the same name is obtained.
Image registration based on Delaunay triangles
Because atural object huge number in high-resolution remote sensing image, texture information enrich, generally there is substantial amounts of local shape
Become, thus simple use be difficult to ensure that registering precision based on global polynomial mapping function, and local triangle fractal transform is adopted
Local mapping function is used, this problem is can effectively solve the problem that.Compared with the common triangulation network, what Rognant et al. was proposed
Delaunay triangulation network has the advantages that topological relation remains constant, i.e., no matter constructed since where, all control points
The Delaunay triangulation network lattice of construction are all consistent.Thus, herein using Delaunay triangulation network constructive geometry mapping function,
Resampling is carried out to image subject to registration.Mapping function is:
Wherein, (x, y) is the coordinate with reference to local triangle in image, and (x', y') is correspondence part in image subject to registration
The coordinate of triangle, due to build three of triangle control point coordinates, it is known that thus can directly obtain it is undetermined in (2.11) formula
Coefficient.
The present invention uses above-mentioned technical proposal, has the advantages that:It is many with based on the tradition such as wavelet transformation and NSCT
The method for registering of dimensional analysis instrument is compared, and the present invention, as multiscale analysis instrument, overcomes control using J-image sequences
The influence that point is caused using different directions high-frequency sub-band when extracting to registration result.At the same time, proposed based on piecemeal plan
Self Adaptive Control point slightly extracts accurate extraction and the distribution constraint that strategy realizes control point.And it is based on the multiple dimensioned same of NMI
Famous cake matching strategy fully utilizes spectrum, texture and scale feature in raw video, can effectively eliminate error hiding phenomenon.
Experiment proves that such Control point extraction and matching strategy ensure that the same place of acquisition to the texture and knot according to regional area
Structure feature is different and be reasonably distributed.Finally, image registration is realized by introducing Delaunay triangle partial transformations.Experiment
Using three groups of different types of high-resolution remote sensing images, and enter with two kinds of method for registering based on traditional multiscale transform analysis tool
Go and compared.Experiment proves to propose that algorithm has registration accuracy higher, and is particularly suited for there is great amount of images in scene
The application scenario of distortion, can provide reliable image registration pretreatment for follow-up change detection.
Brief description of the drawings
Fig. 1 is the method flow diagram of the embodiment of the present invention;
Fig. 2 is 9 × 9 pixel windows of the embodiment of the present invention;
Fig. 3 is 18 × 18 pixel windows of the embodiment of the present invention;
Fig. 4 is the reference image #1 of the embodiment of the present invention
Fig. 5 is the image #2 subject to registration of the embodiment of the present invention;
84 pairs of same places that Fig. 6 is extracted for the method for the embodiment of the present invention;
Fig. 7 is the 43 pairs of same places extracted based on small wave converting method;
Fig. 8 is the 53 pairs of same places extracted based on NSCT methods;
Fig. 9 is the method registration result of the embodiment of the present invention;
Figure 10 is based on small wave converting method registration result;
Figure 11 is based on NSCT method registration results;
Figure 12 is the reference image #3 of the embodiment of the present invention;
Figure 13 is the image #4 subject to registration of the embodiment of the present invention;
61 pairs of same places that Figure 14 is extracted for the method for the embodiment of the present invention;
Figure 15 is the 43 pairs of same places extracted based on small wave converting method;
Figure 16 is the 53 pairs of same places extracted based on NSCT methods;
Figure 17 is the method registration result of the embodiment of the present invention;
Figure 18 is based on small wave converting method registration result;
Figure 19 is based on NSCT method registration results;
Figure 20 is the reference image #5 of the embodiment of the present invention;
Figure 21 is the image #6 subject to registration of the embodiment of the present invention;
30 pairs of same places that Figure 22 is extracted for the method for the embodiment of the present invention;
Figure 23 is the 23 pairs of same places extracted based on small wave converting method;
Figure 24 is the 26 pairs of same places extracted based on NSCT methods;
Figure 25 is the method registration result of the embodiment of the present invention;
Figure 26 is based on small wave converting method registration result;
Figure 27 is based on NSCT method registration results.
Specific embodiment
With reference to specific embodiment, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention
Rather than limitation the scope of the present invention, after the present invention has been read, those skilled in the art are to various equivalences of the invention
The modification of form falls within the application appended claims limited range.
As shown in figure 1, the high-resolution remote sensing image method for registering of point self-adapted distribution is controlled, mainly including following three
Step:(1) multiple dimensioned Control point extraction.(2) homotopy mapping based on NMI.(3) based on Delaunay triangle partial transformations
Geometric correction.
Multiple dimensioned J-image is incorporated into image registration, be first proposed a kind of based on the self-adaptive controlled of partition strategy
Point extracting method processed, realizes the Control point extraction in multiple dimensioned J-image images, so as to overcome control point only to indivedual sides
To the sensitive limitation of high-frequency information.Meanwhile, define Self Adaptive Control point and extract strategy to constrain the distribution at control point.And then adopt
Control point is carried out with normalized mutual information measure NMI it is multiple dimensioned match, can effectively smooth with quasi-function.Finally adopt
Geometric correction is realized with Delaunay triangle partial transformations.The general principle and key of proposed algorithm are described initially below
Step, and then three groups of different types of remote sensing images are tested and analyzed, and respectively with based on wavelet transformation and NSCT
Different method for registering compare experiment.
The method for registering that JSEG is combined with NMI
Multiple dimensioned Control point extraction
Selection J-image images are controlled an extraction and match as multiscale analysis instrument.
1 calculates J-image image sequences
Using the corresponding local homogeney index J- of each pixel in the window calculation raw video of a certain specific dimensions
Value and as the pixel value of the pixel, can obtain the J-image images of single yardstick.The calculating process of J-value is such as
Under:
Quantization method (Deng Y, Kenney C, Moore M S, the et al.Peer for being proposed using Deng et al. first
group filtering and perceptual color image quantization[C]//Proceedings of
the 1999 IEEE International Symposium on Circuits and Systems,1999,4:21-24.)
Raw video is quantified so as to obtain quantification image.Position z (x, y) for making each pixel in quantification image is pixel z
Pixel value, z (x, y) ∈ Z.Z is the set of all pixels composition in specific dimensions window centered on pixel z.For example, adopting
With the window shown in Fig. 2 as the first yardstick, the certain window size centered on pixel z is 9 × 9 pixels.Then in the second chi
In degree, centered on z 18 × 18 pixel windows are used, as shown in Figure 3.Meanwhile, it is further to reduce amount of calculation, in Fig. 3 only
The pixel that "+" is represented is used to calculate the J-value of pixel z under current scale, and the rest may be inferred for the J-value for calculating under each yardstick.
To ensure the uniformity of all directions as far as possible, the angle point in window is removed.
In quantification image, it is the sum of all pixels centered on z in window to define N, then average m:
Define mpTo belong to all pixels average of same grey level p, Z in windowpTo belong to the institute of gray level p in window
There is the set of pixel, P is the gray level sum in quantification image, then belong to the variance and S of same gray-level pixels in windowW
May be defined as:
Define the population variance that ST is all pixels in window:
Then J-value is:
J=(ST-SW)/SW (2.4)
Calculate the J-value of z under obstructed yardstick respectively using different size window, and as the pixel value of z, so as to obtain
Multiple dimensioned J-image image sequences.According to the definition of above J-value, in the J-image of a certain yardstick, a certain pixel
Corresponding J-value values are bigger, then be more likely to be at the edge of object;Conversely, being then likely located at the center of object.Meanwhile, J-
Value combines spectral information, texture information and the dimensional information that raw video is included, and insensitive to directional information, because
This can also be used for the matching of same place as description at control point.
2 Control point extractions and distribution constraint
Definition and feature according to J-value, it is proposed that a kind of Self Adaptive Control point extracting method based on partition strategy,
Its core concept is the more control points of extracted region enriched in textural characteristics, realizes that flow is as follows:
Step1:The J-image images of each yardstick are divided into various sizes of subgraph, certain yardstick neutron image is big
It is small identical with certain window size that is calculating yardstick J-image images, but retain angle point, it is the square of N × N pixels
Window.
Step2:The J-value values of view picture image are calculated using formula (2.4) in raw video, T is defined asa。TaReflection
The overall homogeneous degree of raw video.
Step3:In a certain yardstick J-image, respectively by the J-valve and T of each subgraph center pixelaCompared
Compared with if being more than Ta, then it is assumed that the subgraph inner vein feature rich, complex structure should extract more control point, take subgraph
J-value maximum preceding K pixel is used as control point as in;Conversely, then extracting a small amount of control point, J- in subgraph is taken
Value maximum preceding L pixel meets as control pointThe value of K and L can be according to image actual size and shadow
As the complexity of texture manually sets.
Step4:In all yardstick J-image images repeat Step3 operation, realize control point multiple dimensioned extraction and
Distribution constraint.
Ground control point matching based on NMI
Complete it is multiple dimensioned under Control point extraction after, it is necessary to further with similarity measurement to each control point one by one
Differentiate and obtain point set of the same name.Similarity measurement between control point generally uses the gray scale of the regional area centered on control point to unite
Meter characteristic is used as similarity measure.Therefore, the control point similarity measurement based on topography's gray feature can be used
The general method for registering images based on gray feature.
Mutual information (MI, Mutual Information) is wide variety of a kind of reversibility phase in current image registration
Estimate like property, the features such as with without pre-segmentation, strong robustness and high precision.Due to that need not be pre-processed, even if in figure
In having an application scenario of excalation as data, good registration result can be still obtained, simultaneously can be used for multi-source remote sensing
In the registration of image.On the other hand, mutual information similarity measure is excessively sensitive to the application scenario of image overlapping region, and works as
When overlapping region is reduced, it is meant that less pixel take part in the statistics of mutual information, so as to reduce the reliability of registration.For gram
Take normalized mutual information measure (NMI, the Normalized Mutual of the propositions such as this limitation, Studholme C
Information) can effectively smooth with quasi-function, mutual information makes to the sensitivity of overlapping region in reducing image registration
The relation that object function more accurately reflects between registration parameter and mutual information, so as to improve registration accuracy.Therefore, herein
Selection NMI extracts the same place pair under each yardstick as control point similarity measurement, obtains global point set of the same name.
1 calculates NMI similarity measurements
In a certain yardstick J-image, using the control point region in reference image and image subject to registration as two
Vector C and D, calculate the NMI between control point.The window size for being used is specific with what calculating current scale J-image was used
Window size is identical.Mutual information is described with entropy first, entropy refers to the general characteristic characterized to information source on average, can be led to
Cross and possible temporal information content weighted average is obtained.Define entropy H and mutual information I (C, D) such as formula (2.5), (2.6) institute
Show:
I (C, D)=H (C)+H (D)-H (C, D) (2.6)
Wherein, PiRepresent certain stochastic variable be possible in the probability that occurs of i-th kind of possible situation.Then H (C), H (D),
H (C, D) is respectively with reference to the corresponding entropy of same place and their combination entropy in image and image subject to registration.According to formula
(2.5) understand:
Wherein PC(c) and PDD () is respectively probability distribution when C and D is completely independent, PCDIt is the joint probability point of C and D
Cloth.NMI is defined as follows:
In with reference to the same yardstick J-image of image and image subject to registration, centered on control point, to calculate current chi
The certain window for spending J-value is regional area spanning subgraph picture, calculates the NMI between subgraph so as to measure the phase between control point
Like property.Although it is pointed out that the process for calculating NMI is only united to the local gray level feature of control point region
Meter, but because the gray value of each pixel in J-image images is the corresponding J-value values of each pixel in raw video, because
This it can be seen from the definition of J-value, the ash of regional area where the NMI similarity measurement concentrated expressions that are calculated control point
The similitude of degree, texture and spectral signature, so that relatively reliable.
The 2 maximum bi-directional matching strategies based on NMI
To realize the accurate match between control point, it is proposed that a kind of maximum bi-directional matching strategy based on NMI, matching process
It is as follows:
Step1:In with reference to same yardstick J-image of the image with image subject to registration, NMI is maximum and more than threshold value TNMI
Dominating pair of vertices as a pair of same places, and then obtain point set of the same name 1.
Step2:With Step1 calculated directions conversely, by a certain control point in image subject to registration respectively with reference to institute in image
There is control point to be compared, obtain point set of the same name 2.
Step3:Compare point set of the same name 1 and point set of the same name 2, using identical same place to being obtained most as under current scale
Whole point set of the same name.
Step4:Step1 to Step3 is repeated, all yardstick J-image are differentiated, and the same place that will be obtained is to complete
Portion is mapped to the same position in raw video.
Step5:It is further using classical RANSAC (Random Sample Consensus) algorithm in raw video
Error hiding phenomenon is eliminated, final point set of the same name is obtained.
Image registration based on Delaunay triangles
Because atural object huge number in high-resolution remote sensing image, texture information enrich, generally there is substantial amounts of local shape
Become, thus simple use be difficult to ensure that registering precision based on global polynomial mapping function, and local triangle fractal transform is adopted
Local mapping function is used, this problem is can effectively solve the problem that.Compared with the common triangulation network, what Rognant et al. was proposed
Delaunay triangulation network has the advantages that topological relation remains constant, i.e., no matter constructed since where, all control points
The Delaunay triangulation network lattice of construction are all consistent.Thus, herein using Delaunay triangulation network constructive geometry mapping function,
Resampling is carried out to image subject to registration.Mapping function is:
Wherein, (x, y) is the coordinate with reference to local triangle in image, and (x', y') is correspondence part in image subject to registration
The coordinate of triangle, due to build three of triangle control point coordinates, it is known that thus can directly obtain it is undetermined in (2.11) formula
Coefficient.
Experimental result and analysis
In order to verify the validity of put forward algorithm, herein respectively to three groups of high-definition remote sensing shadows from different sensors
As data set is tested.In addition, experimental result will be based on the method for registering (document of wavelet transformation with use respectively
Zavorin I,Le Moigne J.Use of multi-resolution wavelet feature pyramids for
automatic registration of multi-sensor imagery[J].Image Processing,IEEE
Transactions on,2005,14(6):770-782.) and the method for registering based on NSCT (document Zhigang C,
Fuchang Y,Fu S.Registration technique for high-resolution remote sensing
images based on non-subsampled contourlet transform[J].Acta Optica Sinica,
2009,29(10):2744-2750.) it is compared.Wherein, the method for registering based on wavelet transformation is using global fitting of a polynomial
Registering image is obtained, the method for registering based on NSCT then builds mapping function using Delaunay triangulation network.The method for being proposed
Experiment parameter is set by multigroup experimental selection optimal value, and other method is related to parameter to offer setting with reference to original text.
The evaluation index for being used is same place to image after quantity, registration and the relative phase relationship number with reference to image
(Comparative Correlation)ρcorr, root-mean-square error (Root-Mean-Square Error) RMSEAnd normalization
Root-mean-square error (Normalized-Root-Mean-Square Error) NRMSE, it is defined as follows:
Relative phase relationship number,
Root-mean-square error,
Normalization root-mean-square error,
Wherein, N represents the number of pixels of image after registration, xiIt is pixel z in image after registrationi(i=1,2...N) picture
Element value,It is ziIn the pixel value with reference to respective pixel in image.In each evaluation index, relative phase relationship number is bigger, explanation
Two width images are closer to ideal value is 1.Root-mean-square error and normalization root-mean-square error reflect the dispersion degree of sample, ginseng
Numerical value is smaller, and two images are closer to then registration accuracy is higher.
The experimental result of data set 1 and analysis
Selection data one group image #1, #2 are tested as data set 1, such as Fig. 4, shown in Fig. 5.With reference to image #1 with treat
Registering image #2 is respectively the air remote sensing DOM (Digital Orthophoto Map) of the acquisition of in March, 2009 and 2 months 2012
Data, location is NanJing City, Jiangsu Province,China Hohai University Jiangning school district, and spatial resolution is 0.5m, and image size is
512 × 512 pixels.
Such as Fig. 4, shown in Fig. 5, because two phase images are the DOM shadows that are synthesized by the aviation image of several different visual angles
Picture, substantially, so as to cause a large amount of local deformations, this phenomenon is especially baroque artificial for each regional area visual angle difference
It is more prominent in target and skyscraper, it is therefore desirable to which that more control points are corrected.In an experiment, calculate J-value's
Window is set to 20 × 20 pixels, 10 × 10 pixels and 5 × 5 pixels, that is, calculate three J-image images of yardstick.Setting
Parameter K=6, L=3, TNMI=0.85.Method for registering based on wavelet transformation and NSCT is equally decomposed using three-level.Three kinds of methods
The same place of extraction is concentrated, and same place is to being distributed as shown in Figure 6 to 8.
The registration result that three kinds of methods are obtained is as shown in Fig. 9~Figure 11.
As shown in Fig. 4~Figure 11, can be drawn the following conclusions according to visual analysis:(1) method proposed by the present invention is extracted
Same place to most, and according to regional area structure reasonable layout different from textural characteristics inside image.Due to herein
The Self Adaptive Control point extracting method based on partition strategy is proposed, for inner vein feature rich, baroque artificial
The regions such as building, a large amount of control points marked the minutia of object, local caused by the factors such as visual angle difference so as to reduce
Influence caused by deformation;The regions such as, road that structure single uniform for inner vein, lake surface, a small amount of control point is accurate
Position where marked object.(2) because the method for registering based on wavelet transformation and NSCT is in Control point extraction and same place
Single threshold value is all respectively adopted in matching process to be differentiated, and have ignored local area texture feature and structure in image
Complexity, thus extract control point distribution more disperse, it is difficult to ensure optimal.(3) registration based on wavelet transformation
Method as a result of global fitting of a polynomial geometric correction method, thus to a large amount of locality deformation regions in registration result
Calibration result is poor.(4) method for registering based on NSCT employs the geometric correction method based on triangle conversion, thus to figure
Local deformation as in is more sensitive.But a small amount of control point is difficult to fully mark and the region of local deformation occurs in different phase shadows
Shape and architectural difference as in, so as to influence final registration accuracy.On the other hand, if changing threshold value merely to increase extraction
Control point and the quantity of same place pair, can cause the exponential growth of amount of calculation and the increase of error hiding phenomenon again.
It is the performance of further quantitative analysis algorithm, precision evaluation result is as shown in table 1.It can be seen from Table 1 that:(1)
Method proposed by the present invention has extracted more same places pair, and registration accuracy is apparently higher than other two kinds of algorithms, and visual
Analysis result is consistent.(2) the inventive method is compared compared with other two methods, because the J-image sequences for using are used as multiple dimensioned
Analysis tool, in the absence of the set direction of high-frequency sub-band, thus will not ignore the high-frequency information in a direction, so that extract
Control point has more preferable accuracy.(3) high-resolution remote sensing image registration larger for local deformation, using based on triangle
The geometric correction method of fractal transform is better than global polynomial fitting method.
The comparing of the inventive method of table 2.1 and other method
The experimental result of data set 2 and analysis
Data set 2 use spatial resolution for the SPOT 5 of 5m it is panchromatic-Multi-spectral image fusion image, with reference to image #3, wait to match somebody with somebody
Quasi- image #4 (such as Figure 12, shown in Figure 13), size is 512 × 512 pixels.Fusion wave band include panchromatic wave-band and it is red, green,
Near infrared band.The #3 and #4 acquisition times are respectively in June, 2004 and in July, 2008, and location is Chinese Shanghai.
Compared with data set 1, the width image space of data set 2 two differentiates lower, visual angle difference very little, and ground species are rich
Richness, background is more complicated.In an experiment, the window for calculating J-value is set to 10 × 10 pixels, 8 × 8 pixels and 5 × 5 pictures
Element, that is, calculate three J-image images of yardstick.Setup parameter K=6, L=3, TNMI=0.9.Based on wavelet transformation and NSCT
Method for registering equally using three-level decompose.The same place concentration that three kinds of methods are extracted, same place is to distribution such as Figure 14~Figure 16
It is shown.
The registration result that three kinds of methods are obtained is as shown in Figure 17~Figure 19.
Precision evaluation result is as shown in table 2:
The comparing of the present invention of table 2 and other method
By Figure 14~Figure 19 and table 2 as can be seen that visual analysis result and precision evaluation result and the basic phase of data set 1
Together, this paper algorithm validities are further demonstrated.In addition, three kinds of registration accuracies of algorithm have been compared with data set 1 in data set 2
Improve, main cause is that visual angle difference is smaller in data set 2, so as to the local deformation region caused is less.By comparison sheet 1 and
Table 2 is as can be seen that the bearing calibration based on global fitting of a polynomial by local deformation only in less image registration is influenceed
Preferable effect can be obtained.And the bearing calibration of triangle conversion is used except the accuracy requirement to same place pair is higher
Outward, the quantity of same place pair and distribution also can produce material impact to registration accuracy.
The experimental result of data set 3 and analysis
In preceding two groups of experiments, data set 1 is aviation remote sensing image, and data set 2 is satellite remote-sensing image.Further to test
Card proposes effect of the algorithm to high resolution SAR image registration, the ASAR (Advanced that selection ENVISAT satellites shoot
Synthetic Aperture Radar) tested as data set 3 with reference to image #5, image #6 subject to registration, such as Figure 20, scheme
Shown in 21.The location of data set 3 is Wuhan District of China, and spatial resolution is 10m, acquisition time be respectively 2 months 2008 and
In August, 2009, size is 512 × 512 pixels.
Compare image #5, #6 it can be found that main the composition land, water body, lake of the view data of data set 3 is constituted.To the greatest extent
Pipe image collection time interval only three months, but waters border there occurs significant changes, and lacking in scene has significantly positioning
The object of mark, while in the presence of a large amount of coherent speckle noises, these factors all increased difficulty to the alignment in geographical position.In experiment
In, the window for calculating J-value is set to 10 × 10 pixels, 8 × 8 pixels and 5 × 5 pixels, that is, calculate three J- of yardstick
Image images.Setup parameter K=6, L=3, TNMI=0.85.Method for registering based on wavelet transformation and NSCT equally uses three
Level is decomposed.The same place of three kinds of method extractions concentrates same place to distribution as shown in Figure 22~Figure 24.
The registration result that three kinds of methods are obtained is as shown in Figure 25~Figure 27.
The comparing of the present invention of table 3 and other method
By Figure 22~Figure 27 and table 3 as can be seen that in high resolution SAR image registration, proposition method of the present invention is carried
The same place for taking at most and is distributed more uniformly, while each precision index is also superior to other two kinds of algorithms to quantity.Meanwhile, pass through
Compare registration accuracy of three kinds of algorithms in the experiment of three group data sets as can be seen that a large amount of coherent spots for existing are made an uproar in SAR images
Sound does not significantly reduce three kinds of registration accuracies of algorithm.This has also further demonstrated that the remote sensing image based on multiscale analysis instrument
Method for registering has good robustness to noise jamming.
Claims (1)
1. a kind of high-resolution remote sensing image method for registering for controlling point self-adapted distribution, it is characterised in that:Mainly include many chis
Degree Control point extraction, three steps of the Ground control point matching based on NMI and the image registration based on Delaunay triangles;
Multiple dimensioned Control point extraction
Selection J-image images are controlled an extraction and match as multiscale analysis instrument;
1 calculates J-image image sequences
Using the corresponding local homogeney index J-value of each pixel in the window calculation raw video of a certain specific dimensions simultaneously
As the pixel value of the pixel, the J-image images of single yardstick are obtained, the calculating process of J-value is as follows:
Raw video is quantified so as to obtain quantification image using quantization method first;Make each pixel in quantification image
Position z (x, y) be the pixel value of pixel z, z (x, y) ∈ Z, Z are all pictures in specific dimensions window centered on pixel z
The set of element composition;
In quantification image, it is the sum of all pixels centered on z in window to define N, then average m:
Define mpTo belong to all pixels average of same grey level p, Z in windowpTo belong to all pictures of gray level p in window
The set of element, P is the gray level sum in quantification image, then belong to the variance and S of same gray-level pixels in windowWDefinition
For:
Define STIt is the population variance of all pixels in window:
Then J-value is:
J=(ST-SW)/SW (2.4)
Calculate the J-value of z under different scale respectively using different size window, and as the pixel value of z, so as to obtain many chis
Degree J-image image sequences;According to the definition of above J-value, in the J-image of a certain yardstick, the correspondence of a certain pixel
J-value values it is bigger, then be more likely to be at the edge of object;Conversely, being then likely located at the center of object;Meanwhile, J-value
Spectral information, texture information and dimensional information that raw video is included are combined, and it is insensitive to directional information, therefore also make
For description at control point is used for the matching of same place;
2 Control point extractions and distribution constraint
Self Adaptive Control point extracting method based on partition strategy, realizes that flow is as follows:
Step1-1:The J-image images of each yardstick are divided into various sizes of subgraph, certain yardstick neutron image size
It is identical with the certain window size for calculating yardstick J-image images, but retain angle point, it is the square window of N × N pixels
Mouthful;
Step1-2:The J-value values of view picture image are calculated using formula (2.4) in raw video, T is defined asa;TaReflect
The overall homogeneous degree of raw video;
Step1-3:In a certain yardstick J-image, respectively by the J-valve and T of each subgraph center pixelaIt is compared,
If being more than Ta, then it is assumed that the subgraph inner vein feature rich, complex structure should extract more control point, take subgraph
Middle J-value maximum preceding K pixel is used as control point;Conversely, then extracting a small amount of control point, J- in subgraph is taken
Value maximum preceding L pixel meets as control pointThe value of K and L is according to image actual size and image line
The complexity setting of reason;
Step1-4:In all yardstick J-image images repeat Step1-3 operation, realize control point multiple dimensioned extraction and
Distribution constraint;
Ground control point matching based on NMI
1 calculates NMI similarity measurements
In a certain yardstick J-image, using with reference to the control point region in image and image subject to registration as two vector C
And D, calculate the NMI between control point;The certain window chi that the window size for being used is used with calculating current scale J-image
It is very little identical;Mutual information is described with entropy first;Define entropy H and mutual information I (C, D) such as formula (2.5), (2.6) shown:
H=- ΣiPilog2Pi (2.5)
I (C, D)=H (C)+H (D)-H (C, D) (2.6)
Wherein, PiRepresent certain stochastic variable be possible in the probability that occurs of i-th kind of possible situation;Then H (C), H (D), H (C,
D) it is respectively with reference to the corresponding entropy of same place and their combination entropy in image and image subject to registration;Can according to formula (2.5)
Know:
H (C)=- ∑cPC(c)log2PC(c) (2.7)
H (D)=- ∑dPD(d)log2PD(d) (2.8)
H (C, D)=- ∑c,dPCD(c,d)log2PCD(c,d) (2.9)
Wherein PC(c) and PDD () is respectively probability distribution when C and D is completely independent, PCDIt is the joint probability distribution of C and D;NMI
It is defined as follows:
In with reference to the same yardstick J-image of image and image subject to registration, centered on control point, to calculate current scale J-
The certain window of value is regional area spanning subgraph picture, and the NMI calculated between subgraph is similar between control point so as to measure
Property;
The 2 maximum bi-directional matching strategies based on NMI
To realize the accurate match between control point, it is proposed that a kind of maximum bi-directional matching strategy based on NMI, matching process is such as
Under:
Step1:In with reference to same yardstick J-image of the image with image subject to registration, NMI is maximum and more than threshold value TNMIControl
System point obtains point set of the same name 1 to as a pair of same places;
Step2:With Step1 calculated directions conversely, by a certain control point in image subject to registration respectively with reference to all controls in image
System point is compared, and obtains point set of the same name 2;
Step3:Compare point set of the same name 1 and point set of the same name 2, using identical same place to final same as what is obtained under current scale
Name point set;
Step4:Step1 to Step3 is repeated, all yardstick J-image are differentiated, and by the same place of acquisition to all reflecting
It is mapped to the same position in raw video;
Step5:Error hiding phenomenon is further eliminated using classical RANSAC algorithms in raw video, obtains final of the same name
Point set;
Image registration based on Delaunay triangles
Using Delaunay triangulation network constructive geometry mapping function, resampling is carried out to image subject to registration;Mapping function is:
Wherein, a0, a1, a2, b0, b1, b2It is that, with reference to the undetermined coefficient in mapping function between image and image subject to registration, (x, y) is
With reference to the coordinate of local triangle in image, (x', y') is the coordinate of correspondence local triangle in image subject to registration, due to building
Three of triangle control point coordinates, it is known that thus directly obtain undetermined coefficient in (2.11) formula.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410637686.5A CN104361590B (en) | 2014-11-12 | 2014-11-12 | High-resolution remote sensing image registration method with control points distributed in adaptive manner |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410637686.5A CN104361590B (en) | 2014-11-12 | 2014-11-12 | High-resolution remote sensing image registration method with control points distributed in adaptive manner |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104361590A CN104361590A (en) | 2015-02-18 |
CN104361590B true CN104361590B (en) | 2017-05-24 |
Family
ID=52528848
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410637686.5A Expired - Fee Related CN104361590B (en) | 2014-11-12 | 2014-11-12 | High-resolution remote sensing image registration method with control points distributed in adaptive manner |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104361590B (en) |
Families Citing this family (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104992459A (en) * | 2015-07-27 | 2015-10-21 | 北京邮电大学世纪学院 | Method and apparatus for generating symmetrical pattern |
CN106296717B (en) * | 2016-08-25 | 2019-03-19 | 中国电子科技集团公司第十研究所 | Multiband SAR images match object localization method |
CN107341824B (en) * | 2017-06-12 | 2020-07-28 | 西安电子科技大学 | Comprehensive evaluation index generation method for image registration |
CN107705329B (en) * | 2017-10-24 | 2021-03-16 | 武汉大学 | High-resolution optical satellite staring image registration method based on geometric constraint |
CN108267455B (en) * | 2018-03-08 | 2023-11-28 | 陕西科技大学 | Device and method for detecting defects of printed characters of plastic film |
CN108629190B (en) * | 2018-03-23 | 2020-08-25 | 国网陕西省电力公司电力科学研究院 | Geographic information data decryption method |
CN108387899B (en) * | 2018-04-17 | 2020-07-31 | 南京师范大学 | Ground control point automatic selection method in synthetic aperture radar interferometry |
CN108919265B (en) * | 2018-07-16 | 2022-05-06 | 北京航空航天大学 | Space-borne TomosAR space baseline sequence design method based on maximum mutual information quantity |
CN109191501B (en) * | 2018-08-07 | 2021-08-24 | 宁波市测绘和遥感技术研究院 | Automatic geometric registration method for multi-source high-spatial-resolution images |
CN110084841A (en) * | 2019-04-29 | 2019-08-02 | 优乐圈(武汉)科技有限公司 | A kind of weighting guidance figure filtering Stereo Matching Algorithm based on LOG operator |
CN111639662A (en) * | 2019-12-23 | 2020-09-08 | 珠海大横琴科技发展有限公司 | Remote sensing image bidirectional matching method and device, electronic equipment and storage medium |
CN113516693B (en) * | 2021-05-21 | 2023-01-03 | 郑健青 | Rapid and universal image registration method |
CN113486928B (en) * | 2021-06-16 | 2022-04-12 | 武汉大学 | Multi-view image alignment method based on rational polynomial model differentiable tensor expression |
CN114202705B (en) * | 2022-02-16 | 2022-05-20 | 清华大学 | Spectral feature time sequence construction method and system |
CN115077399B (en) * | 2022-07-12 | 2023-03-24 | 浙江威罗德汽配股份有限公司 | Manufacturing method of stamping part |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103514606A (en) * | 2013-10-14 | 2014-01-15 | 武汉大学 | Heterology remote sensing image registration method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8755636B2 (en) * | 2011-09-14 | 2014-06-17 | Mediatek Inc. | Method and apparatus of high-resolution image reconstruction based on multi-frame low-resolution images |
-
2014
- 2014-11-12 CN CN201410637686.5A patent/CN104361590B/en not_active Expired - Fee Related
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103514606A (en) * | 2013-10-14 | 2014-01-15 | 武汉大学 | Heterology remote sensing image registration method |
Non-Patent Citations (4)
Title |
---|
A Implicit Similarity Method for Multi-modal Image Registrtaion;Peng Ye et al.;《2011 International Workshop on Multi-Platform/Multi-Sensor Remote Sensing and Mapping》;20110110;第1-4页 * |
Multi-sensor Image Automatic Registration Using Mutual Information;JIANG Jing et al.;《International Conference on Graphic and Image Processing》;20130320;第8768卷;第1-7页 * |
基于SIFT点特征和Canny边缘特征匹配的多源遥感影像配准研究;王万同 等;《计算机科学》;20110731;第38卷(第7期);第287-289页 * |
基于非采样Contourlet变换高分辨率遥感图像配准;陈志刚 等;《光学学报》;20091031;第29卷(第10期);第2744-2750页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104361590A (en) | 2015-02-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104361590B (en) | High-resolution remote sensing image registration method with control points distributed in adaptive manner | |
CN104867126B (en) | Based on point to constraint and the diameter radar image method for registering for changing region of network of triangle | |
CN104536009B (en) | Above ground structure identification that a kind of laser infrared is compound and air navigation aid | |
CN104376330B (en) | Polarimetric SAR Image Ship Target Detection method based on super-pixel scattering mechanism | |
CN105787937B (en) | A kind of high resolution remote sensing image road change detecting method based on OSM | |
CN107067415B (en) | A kind of object localization method based on images match | |
CN107610164B (en) | High-resolution four-number image registration method based on multi-feature mixing | |
CN103729853B (en) | High score remote sensing image building under three-dimension GIS auxiliary damages detection method | |
CN109146889A (en) | A kind of field boundary extracting method based on high-resolution remote sensing image | |
CN109871823B (en) | Satellite image ship detection method combining rotating frame and context information | |
CN107341781B (en) | SAR image correction method based on improved phase consistency feature vector base map matching | |
CN103235810B (en) | Remote sensing image reference mark data intelligence search method | |
CN110033499B (en) | Landsat OLI image-based water body mapping method | |
CN111008664B (en) | Hyperspectral sea ice detection method based on space-spectrum combined characteristics | |
CN106295503A (en) | The high-resolution remote sensing image Ship Target extracting method of region convolutional neural networks | |
CN104834931A (en) | Improved SIFT algorithm based on wavelet transformation | |
CN107688777A (en) | A kind of urban green space extracting method for cooperateing with multi-source Remote Sensing Images | |
CN104680151B (en) | A kind of panchromatic remote sensing image variation detection method of high-resolution for taking snow covering influence into account | |
CN103226826A (en) | Method for detecting changes of remote sensing image of visual attention model based on local entropy | |
CN106340005A (en) | High-resolution remote sensing image unsupervised segmentation method based on scale parameter automatic optimization | |
CN108364279A (en) | Determine the method that stationary orbit remote sensing satellite is directed toward deviation | |
CN117218201A (en) | Unmanned aerial vehicle image positioning precision improving method and system under GNSS refusing condition | |
CN106886988A (en) | A kind of linear goal detection method and system based on unmanned aerial vehicle remote sensing | |
CN102902982B (en) | Based on the SAR image texture classifying method of observation vector difference | |
CN105631860A (en) | Local sorted orientation histogram descriptor-based image correspondence point extraction method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into 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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170524 Termination date: 20211112 |