CN110060285A - A kind of remote sensing image registration method and system based on SURF algorithm - Google Patents
A kind of remote sensing image registration method and system based on SURF algorithm Download PDFInfo
- Publication number
- CN110060285A CN110060285A CN201910354697.5A CN201910354697A CN110060285A CN 110060285 A CN110060285 A CN 110060285A CN 201910354697 A CN201910354697 A CN 201910354697A CN 110060285 A CN110060285 A CN 110060285A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- image
- registration
- feature
- algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of remote sensing image registration method and system based on SUFR algorithm.The described method includes: uploading benchmark remote sensing images and remote sensing images subject to registration into GPU video memory；With parallel calculation in GPU, feature extraction is carried out to benchmark remote sensing images and remote sensing images subject to registration respectively using SURF algorithm, obtains fisrt feature description and second feature description；Fisrt feature description is the Feature Descriptor of benchmark remote sensing images, and second feature description is the Feature Descriptor of remote sensing images subject to registration；Characteristic matching is carried out to fisrt feature description and second feature description using quick approximate KNN searching algorithm in GPU, obtains characteristic matching image；Characteristic matching image is downloaded in CPU memory, the image after being registrated.The present invention, with parallel calculation, realizes image registration using SURF algorithm in GPU, can be improved the speed of service of remote sensing image registration, and then improves registration efficiency.
Description
Technical field
The present invention relates to image registration techniques fields, more particularly to a kind of remote sensing image registration side based on SUFR algorithm
Method and system.
Background technique
The registration of remote sensing images is the committed step of the applications such as remote sensing images matching, variation monitoring, Multi-spectral image fusion.Nobody
Machine remote sensing images show high-resolution, the trend of big data quantity, and the utilization benefit for promoting remote sensing images is always people's pursuit
Target.In carrying out process of image registration, the selection of feature extraction determines whole registration efficiency, SURF feature descriptor
Although on operation efficiency it is faster than SIFT very much, but still be not able to satisfy practical application to real-time in remote sensing image registration speed
It is required that.
In unmanned aerial vehicle remote sensing image registration field, currently, the method for registering generallyd use is to use SURF in CPU memory
Algorithm carries out the operations such as feature description, feature detection, feature extraction and carries out image registration.Although SURF algorithm changes compared to SIFT
Extraction and describing mode into feature, the extraction and description of feature are completed with a kind of highly efficient mode, and arithmetic speed is high
Lead to image registration low efficiency, characteristic point slowly in SIFT algorithm, but there are still the speed due to feature point description, monitoring, extraction
Matching accuracy is low, image registration effect difference disadvantage, is not able to satisfy the speed of service in Millisecond to the remotely-sensed data of magnanimity
The requirement of registration.
Summary of the invention
Based on this, it is necessary to a kind of remote sensing image registration method and system based on SUFR algorithm is provided, to improve remote sensing
The speed of service of image registration, and then improve registration efficiency.
To achieve the above object, the present invention provides following schemes:
A kind of remote sensing image registration method based on SUFR algorithm, comprising:
Benchmark remote sensing images and remote sensing images subject to registration are uploaded into GPU video memory；
With parallel calculation in the GPU, using SURF algorithm respectively to benchmark remote sensing images and described
Remote sensing images subject to registration carry out feature extraction, obtain fisrt feature description and second feature description；The fisrt feature is retouched
The Feature Descriptor that son is the benchmark remote sensing images is stated, second feature description is the spy of the remote sensing images subject to registration
Sign description；
Fisrt feature description is retouched with the second feature using quick approximate KNN searching algorithm in GPU
Sub- carry out characteristic matching is stated, characteristic matching image is obtained；
The characteristic matching image is downloaded in CPU memory, the image after being registrated.
Optionally, it is described in the GPU with parallel calculation, using SURF algorithm respectively to the benchmark remote sensing
Image and the remote sensing images subject to registration carry out feature extraction, obtain fisrt feature description and second feature description, specifically
Include:
In GPU by the way of parallel computation respectively to the benchmark remote sensing images and the remote sensing images subject to registration into
Row Integral Processing obtains benchmark remote sensing integral image and remote sensing integral image subject to registration；
The benchmark remote sensing integral image and the remote sensing integral image subject to registration are carried out respectively using SURF algorithm special
Sign detection, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter is to the benchmark remote sensing integrogram
Location parameter and scale parameter as carrying out the characteristic point that characteristic point detects；The second feature parameter is to described wait match
Quasi- remote sensing integral image carries out the location parameter and scale parameter for the characteristic point that feature detects；
Fisrt feature principal direction is calculated according to the fisrt feature parameter, it is special to calculate second according to the second feature parameter
Levy principal direction；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images, and the second feature parameter is
The feature principal direction of the remote sensing images subject to registration；
Fisrt feature description is calculated according to the fisrt feature principal direction, calculates the according to the second feature principal direction
Two Feature Descriptors.
Optionally, described to download to the characteristic matching image in CPU memory, the image after being registrated is specific to wrap
It includes:
The characteristic matching image is downloaded in CPU memory, and the characteristic matching image in CPU memory is had an X-rayed
Matrixing；
Image after transformed characteristic matching image to be determined as to registration.
Optionally, it is described in GPU by the way of parallel computation respectively to benchmark remote sensing images and described wait match
Quasi- remote sensing images carry out Integral Processing, obtain benchmark remote sensing integral image and remote sensing integral image subject to registration, specifically include:
The first pixel value is calculated using row prefix addition algorithm in GPU；First pixel value is the benchmark remote sensing
The sum of the pixel of every a line pixel in image；
It is summed using column prefix addition algorithm to first pixel value in GPU, obtains benchmark remote sensing integrogram
Picture；
The second pixel value is calculated using row prefix addition algorithm in GPU；Second pixel value is described subject to registration distant
Feel the sum of the pixel of every a line pixel in integral image；
It is summed using column prefix addition algorithm to second pixel value in GPU, obtains remote sensing integral subject to registration
Image.
Optionally, described that the benchmark remote sensing integral image and the remote sensing subject to registration are integrated respectively using SURF algorithm
Image carries out feature detection, obtains fisrt feature parameter and second feature parameter, specifically includes:
Convolution algorithm is carried out to the benchmark remote sensing integral image using box filter, obtains the benchmark remote sensing integral
The scale space of image；
Feature detection is carried out to the scale space of the benchmark remote sensing integral image, obtains fisrt feature parameter；
Convolution algorithm is carried out to the remote sensing integral image subject to registration using box filter, obtains the remote sensing subject to registration
The scale space of integral image；
Feature detection is carried out to the scale space of the remote sensing integral image subject to registration, obtains second feature parameter.
The present invention also provides a kind of remote sensing image registration systems based on SUFR algorithm, comprising:
Image uploading module, for uploading to benchmark remote sensing images and remote sensing images subject to registration in GPU video memory；
Characteristic extracting module, in the GPU with parallel calculation, using SURF algorithm respectively to the base
Quasi- remote sensing images and the remote sensing images subject to registration carry out feature extraction, obtain fisrt feature description and second feature description
Son；Fisrt feature description is the Feature Descriptor of the benchmark remote sensing images, and second feature description is described
The Feature Descriptor of remote sensing images subject to registration；
Characteristic matching module, for being described using quick approximate KNN searching algorithm to the fisrt feature in GPU
It is sub to carry out characteristic matching with second feature description, obtain characteristic matching image；
Download module, for the characteristic matching image to be downloaded in CPU memory, the image after being registrated.
Optionally, the characteristic extracting module, specifically includes:
Integral unit, in GPU by the way of parallel computation respectively to the benchmark remote sensing images and it is described to
It is registrated remote sensing images and carries out Integral Processing, obtain benchmark remote sensing integral image and remote sensing integral image subject to registration；
Characteristic detection unit, for using SURF algorithm respectively to the benchmark remote sensing integral image and described subject to registration distant
Feel integral image and carry out feature detection, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter is to institute
State location parameter and scale parameter that benchmark remote sensing integral image carries out the characteristic point that characteristic point detects；The second feature
Parameter is the location parameter and scale parameter that the characteristic point that feature detects is carried out to the remote sensing integral image subject to registration；
First computing unit, for calculating fisrt feature principal direction according to the fisrt feature parameter, according to described second
Calculation of characteristic parameters second feature principal direction；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images,
The second feature parameter is the feature principal direction of the remote sensing images subject to registration；
Second computing unit, for calculating fisrt feature description according to the fisrt feature principal direction, according to described the
Two feature principal directions calculate second feature description.
Optionally, the download module, specifically includes:
Converter unit, for downloading to the characteristic matching image in CPU memory, and to the feature in CPU memory
Perspective matrix transformation is carried out with image；
Determination unit, for transformed characteristic matching image to be determined as the image after being registrated.
Optionally, the integral unit, specifically includes:
First computation subunit, for calculating the first pixel value using row prefix addition algorithm in GPU；First picture
Element value is the sum of the pixel of every a line pixel in the benchmark remote sensing images；
Second computation subunit, for using column prefix addition algorithm to sum first pixel value in GPU,
Obtain benchmark remote sensing integral image；
Third computation subunit, for calculating the second pixel value using row prefix addition algorithm in GPU；Second picture
Element value is the sum of the pixel of every a line pixel in the remote sensing integral image subject to registration；
4th computation subunit, for using column prefix addition algorithm to sum second pixel value in GPU,
Obtain remote sensing integral image subject to registration.
Optionally, the characteristic detection unit, specifically includes:
5th computation subunit, for carrying out convolution algorithm to the benchmark remote sensing integral image using box filter,
Obtain the scale space of the benchmark remote sensing integral image；
First detection sub-unit carries out feature detection for the scale space to the benchmark remote sensing integral image, obtains
Fisrt feature parameter；
6th computation subunit, for carrying out convolution fortune to the remote sensing integral image subject to registration using box filter
It calculates, obtains the scale space of the remote sensing integral image subject to registration；
Second detection sub-unit carries out feature detection for the scale space to the remote sensing integral image subject to registration, obtains
To second feature parameter.
Compared with prior art, the beneficial effects of the present invention are:
The invention proposes a kind of remote sensing image registration method and system based on SUFR algorithm.This method in GPU with
Parallel calculation carries out feature extraction to benchmark remote sensing images and remote sensing images subject to registration respectively using SURF algorithm, obtains
To fisrt feature description and second feature description；Fisrt feature description is the Feature Descriptor of benchmark remote sensing images, the
Two Feature Descriptors are the Feature Descriptor of remote sensing images subject to registration；Quick approximate KNN searching algorithm pair is used in GPU
Fisrt feature description carries out characteristic matching with second feature description, obtains characteristic matching image；It will be under characteristic matching image
It is downloaded in CPU memory, the image after being registrated.The present invention, with parallel calculation, is realized in GPU using SURF algorithm
Feature extraction realizes characteristic matching using quick approximate KNN searching algorithm, can be improved the operation speed of remote sensing image registration
Degree, and then improve registration efficiency.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without any creative labor, can also be according to these attached drawings
Obtain other attached drawings.
Fig. 1 is a kind of flow chart of the remote sensing image registration method based on SUFR algorithm of the embodiment of the present invention；
Fig. 2 is first group of remote sensing image registration result figure of the embodiment of the present invention；
Fig. 3 is second group of remote sensing image registration result figure of the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, with reference to the accompanying drawing and specific real
Applying mode, the present invention is described in further detail.
Term is explained:
Box filter (Box filter), quick approximate KNN searching algorithm (Fast Library for
Approximate Nearest Neighbors, Flann).
Fig. 1 is a kind of flow chart of the remote sensing image registration method based on SUFR algorithm of the embodiment of the present invention.
Referring to Fig. 1, the remote sensing image registration method based on SUFR algorithm of embodiment includes:
Step S1: benchmark remote sensing images and remote sensing images subject to registration are uploaded into GPU video memory.
Specifically, the benchmark remote sensing images and the remote sensing images subject to registration are to be sent to GPU video memory by host memory
In.
Step S2: with parallel calculation in the GPU, using SURF algorithm respectively to the benchmark remote sensing images
Feature extraction is carried out with the remote sensing images subject to registration, obtains fisrt feature description and second feature description.
Fisrt feature description is the Feature Descriptor of the benchmark remote sensing images, and the second feature describes son and is
The Feature Descriptor of the remote sensing images subject to registration.
The step S2, specifically includes:
21) in GPU by the way of parallel computation respectively to the benchmark remote sensing images and the remote sensing figure subject to registration
As carrying out Integral Processing, benchmark remote sensing integral image and remote sensing integral image subject to registration are obtained.
The GPU prefix addition algorithm for specifically using Harris in this step, obtains benchmark remote sensing integral image and wait match
Quasi- remote sensing integral image.It is specific:
Firstly, calculating the first pixel value, calculation formula using row prefix addition algorithm in GPU are as follows:
Sx(x, j) indicates the first pixel value, i.e., the sum of the pixel of every a line pixel, I (x, y) table in benchmark remote sensing images
Show that benchmark remote sensing images, (x, y) indicate the coordinate of pixel in benchmark remote sensing images,；Then column prefix addition is used in GPU
Algorithm sums to first pixel value, obtains benchmark remote sensing integral image, calculation formula are as follows:
Wherein S (x, y) is benchmark remote sensing integral image.In row or column prefix addition algorithm, to the behaviour of each row or column
It is completely independent, so this method for completing benchmark remote sensing images integral in two steps is very suitable to come using parallel computation
It speeds up to realize, is able to ascend the speed of operation.
The process for obtaining remote sensing integral image subject to registration is identical as the process for obtaining benchmark remote sensing integral image, is also divided into
Two steps are completed.The second pixel value is calculated using row prefix addition algorithm in GPU；Second pixel value is described subject to registration distant
Feel the sum of the pixel of every a line pixel in integral image；Then use column prefix addition algorithm to second picture in GPU
Plain value is summed, and remote sensing integral image subject to registration can be obtained.
22) the benchmark remote sensing integral image and the remote sensing integral image subject to registration are carried out respectively using SURF algorithm
Feature detection, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter is to integrate to the benchmark remote sensing
Image carries out the location parameter and scale parameter for the characteristic point that characteristic point detects；The second feature parameter be to it is described to
Registration remote sensing integral image carries out the location parameter and scale parameter for the characteristic point that feature detects.It specifically includes:
A, convolution algorithm is carried out to the benchmark remote sensing integral image using box filter, obtains the benchmark remote sensing product
The scale space of partial image；Feature detection is carried out to the scale space of the benchmark remote sensing integral image, obtains fisrt feature ginseng
Number.Specifically, SURF algorithm is when feature detects building scale space, each tower tomographic image generating process is independent from each other.Cause
This can open one for the generating process of every layer of every width target image among the process of building for carrying out scale space
GPU Parallel Kernel (Workgroup), wherein the size of Workgroup is determined by the size of target image, and each kernel function is main
The calculating of Hessian matrix and Hessian receptance function value is completed, and then obtains the scale space of integral image.Scale space
After building, algorithm obtains scale parameter by scale space, and algorithm carries out the accurate positioning of characteristic point, obtains location parameter.Its
In, in key point position fixing process, SURF algorithm is to carry out threshold decision first, acquires the set of sampling point to be processed, then
The judgement that extreme point is successively carried out using GPU core function carries out size comparison to the extreme value of 26 neighborhood points, to obtain feature
The location parameter of point.
B, convolution algorithm is carried out to the remote sensing integral image subject to registration using box filter, obtained described subject to registration distant
Feel the scale space of integral image；Feature detection is carried out to the scale space of the remote sensing integral image subject to registration, obtains second
Characteristic parameter.The detailed process for obtaining second feature parameter is identical with the above-mentioned process for obtaining fisrt feature parameter.
23) fisrt feature principal direction is calculated according to the fisrt feature parameter, calculates the according to the second feature parameter
Two feature principal directions；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images, the second feature ginseng
Number is the feature principal direction of the remote sensing images subject to registration.By the detailed process of calculation of characteristic parameters feature principal direction are as follows:
Due to principal direction calculate with Feature Descriptor (feature vector) seek it is inconsistent in the configuration of thread block, so
Respectively completed with a GPU function call.Wherein the GPU of principal direction is calculated carries out according to two steps of original SURF: every first
A thread calculates the Haar response of a sampled point and is stored in shared memory；Then, per thread calculates a fanlight
Interior vector sum；The maximum vector of mould is stored in by the size for finally comparing the mould of 16 fanlight vector sums of fanlight neighborhood
Shared memory obtains the maximum in shared memory in local modulus maximum value, and the angle of the maximum is the master of this feature
Direction.
24) fisrt feature description is calculated according to the fisrt feature principal direction, according to the second feature principal direction meter
Calculate second feature description.The process for calculating Feature Descriptor by feature principal direction is specific:
Feature vector, that is, Feature Descriptor is generated by the calculated feature principal direction of previous step.SURF feature vector and
Row calculating process is more intuitive, and the feature vector calculating with SURF algorithm is consistent, and has 16 regions since SURF describes son,
16 threads can be used to calculate the feature vector of a characteristic point, per thread calculates a region.The 4*4 around characteristic point
Haar wavelet transformation is calculated in a square subregions domain, and four dimensional vectors can be obtained to indicate each subregion, calculate horizontal direction value
The sum of, the sum of horizontal direction absolute value, the sum of vertical direction, then they are added by the sum of vertical direction absolute value, final to obtain
Feature Descriptor of the vector tieed up to 4*4*4=64 as image.
Step S3: using quick approximate KNN searching algorithm to fisrt feature description and described the in GPU
Two Feature Descriptors carry out characteristic matching, obtain characteristic matching image.It is specific:
Feature is carried out with second feature description to fisrt feature description obtained in the previous step in GPU
Match.The SURF Feature Descriptor extracted is matched using quick approximate KNN searching algorithm in GPU, in order to
Detection speed is improved, before calling matching function, first trains a matcher.Training process can use Flann first
Based Matcher function optimizes, and is characterized descriptor (descriptor) and establishes index tree, in this way will be in a large amount of numbers of matching
According to when, speed is promoted can be very fast.It is as follows using the quick matched partial code of approximate KNN searching algorithm in GPU:
Ptr<cv::cuda::DescriptorMatcher>matcher=
cv::cuda::DescriptorMatcher::createFlannBasedMatcher(surf.defaultNorm
())；
vector<DMatch>matches；
matcher->match(descriptors1GPU,descriptors2GPU,matches)；// to what is described
Feature point image carries out matching and is put into inside container matches
Step S4: the characteristic matching image is downloaded in CPU memory, the image after being registrated.
The step S4 is specifically included:
The characteristic matching image is downloaded in CPU memory, and the characteristic matching image in CPU memory is had an X-rayed
Matrixing；Image after transformed characteristic matching image to be determined as to registration.It is exactly using the purpose that perspective matrix converts
Image by characteristic matching image projection to a new plane, after being registrated.IDE used in the step is Visual
Studio 2017, exploitation environment are (the open source computer vision library) opencv3.4.1.For calculating perspective transform in opencv3
The function of matrix is cv::getPerspectiveTransform (), its calling form of C++ interface is as follows:
Cv::Mat cv::getPerspectiveTransform//return 3x3 perspective transformation matrix
(const cv::Point2f*src, four apex coordinates of // source images (point array)
The coordinate (point array) on four vertex on const cv::Point2f*dst//target image
)；
By calling the cv::getPerspectiveTransform () function, final remote sensing image registration knot is obtained
Fruit, that is, the image after being registrated.
The remote sensing image registration method based on SUFR algorithm in the present embodiment is verified below.
Fig. 2 is first group of remote sensing image registration result figure of the embodiment of the present invention, wherein (a) in Fig. 2 is partially first group
Remote sensing images subject to registration, (b) in Fig. 2 are partially benchmark remote sensing images, and (c) in Fig. 2 is partially for using traditional SURF algorithm
Image after registration, (d) in Fig. 2 is partially for using the image after the algorithm registration in the present embodiment.Fig. 3 is that the present invention is implemented
Second group of remote sensing image registration result figure of example, wherein (a) in Fig. 3 is partially second group of remote sensing images subject to registration, in Fig. 3
It (b) is partially benchmark remote sensing images, (c) in Fig. 3 is partially the image after being registrated using traditional SURF algorithm, (d) in Fig. 3
Part is using the image after the algorithm registration in the present embodiment.
By two methods of the result images of observation comparison, we can intuitively observe the GPU using the present embodiment
The result accuracy for accelerating SURF method more preferably to obtain the effect of unmanned aerial vehicle remote sensing image registration is also higher.And it uses single
SURF algorithm is registrated remote sensing images, and obtained result has apparent distortion, deformation, is not able to satisfy in accuracy and actually answers
Needs.
Table 1 is that tradition Surf algorithm and the present embodiment GPU accelerate SURF algorithm operation efficiency comparing result.
Table 1
Upper table proposes the runing time of two kinds of algorithms with the progress unmanned aerial vehicle remote sensing image registration of both algorithms is respectively adopted
The number for the characteristic point got has done detailed statistics.It is distant in the unmanned plane of equal resolution to this two index comprehensive comparisons
The feature points extracted in sense image using algorithm in the present embodiment are 16 times of SURF algorithm or so, and are calculated than SURF
Under the feature points that 16 times of fado, the runing time for carrying out remote sensing image registration has lacked nearly 3 times.Just because of the present embodiment
The number of characteristic point that can be extracted based on the GPU SURF algorithm accelerated uses this implementation 16 times more than SURF algorithm
The method of example carries out the effect of remote sensing image registration also superior to single SURF algorithm, can satisfy unmanned aerial vehicle remote sensing image registration
Real-time, accuracy, the requirement of robustness can promote high-resolution unmanned aerial vehicle remote sensing image in remote Sensing Image Analysis field
Utilization benefit.
The remote sensing image registration method based on SUFR algorithm of the present embodiment has the advantage that
(1) integral image is generated using prefix addition algorithm in GPU: being totally divided into the parallel side of row and column in GPU
Formula progress pixel value summation add parallel to every a line (or each column) respectively when calculating the sum of the pixel value of row (or column)
The addition of the sum of row and column pixel value is finally obtained the integral image of whole image by method operation.The method of this parallel computation is big
The speed for calculating integral image is improved greatly.
(2) characteristic value is calculated in GPU, obtains characteristic parameter: every in a certain neighborhood of pixel points when calculating characteristic value
The calculating of a characteristic value is all independent, therefore GPU can be used and carry out parallel computation, accelerates the solving speed of characteristic value.
(3) feature principal direction is calculated in GPU: being sought since principal direction is calculated with feature vector (Feature Descriptor)
It is inconsistent in the configuration of thread block, so respectively being completed with a GPU function call.The a certain picture of parallel computation in GPU for the first time
The Haar wavelet transformation of all pixels point in vegetarian refreshments and its neighborhood.All pixels in neighborhood of pixel points are calculated in GPU for the second time
Direction vector, find out the maximum feature principal direction as this feature point, improve the speed of service.
(4) feature vector is acquired in GPU: the GPU calculating process of SURF feature vector is more intuitive, since SURF is described
Son has 16 regions, and 16 threads can be used to calculate the feature vector of a feature, and per thread calculates a region, improves
The speed of service.
(5) characteristic matching is carried out in GPU: in GPU to extracted SURF description son using Flann algorithm into
Row matching, calls Flann Based Matcher function to optimize acceleration, is characterized descriptor in GPU video memory
(descriptor) index tree is established, very fast speed will be promoted when matching mass data in this way.
The present invention also provides a kind of remote sensing image registration systems based on SUFR algorithm, comprising:
Image uploading module, for uploading to benchmark remote sensing images and remote sensing images subject to registration in GPU video memory.
Characteristic extracting module, in the GPU with parallel calculation, using SURF algorithm respectively to the base
Quasi- remote sensing images and the remote sensing images subject to registration carry out feature extraction, obtain fisrt feature description and second feature description
Son；Fisrt feature description is the Feature Descriptor of the benchmark remote sensing images, and second feature description is described
The Feature Descriptor of remote sensing images subject to registration.
Characteristic matching module, for being described using quick approximate KNN searching algorithm to the fisrt feature in GPU
It is sub to carry out characteristic matching with second feature description, obtain characteristic matching image.
Download module, for the characteristic matching image to be downloaded in CPU memory, the image after being registrated.
As an alternative embodiment, the characteristic extracting module, specifically includes:
Integral unit, in GPU by the way of parallel computation respectively to the benchmark remote sensing images and it is described to
It is registrated remote sensing images and carries out Integral Processing, obtain benchmark remote sensing integral image and remote sensing integral image subject to registration.
Characteristic detection unit, for using SURF algorithm respectively to the benchmark remote sensing integral image and described subject to registration distant
Feel integral image and carry out feature detection, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter is to institute
State location parameter and scale parameter that benchmark remote sensing integral image carries out the characteristic point that characteristic point detects；The second feature
Parameter is the location parameter and scale parameter that the characteristic point that feature detects is carried out to the remote sensing integral image subject to registration.
First computing unit, for calculating fisrt feature principal direction according to the fisrt feature parameter, according to described second
Calculation of characteristic parameters second feature principal direction；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images,
The second feature parameter is the feature principal direction of the remote sensing images subject to registration.
Second computing unit, for calculating fisrt feature description according to the fisrt feature principal direction, according to described the
Two feature principal directions calculate second feature description.
As an alternative embodiment, the download module, specifically includes:
Converter unit, for downloading to the characteristic matching image in CPU memory, and to the characteristic matching figure in CPU
As carrying out perspective matrix transformation.
Determination unit, for transformed characteristic matching image to be determined as the image after being registrated.
As an alternative embodiment, the integral unit, specifically includes:
First computation subunit, for calculating the first pixel value using row prefix addition algorithm in GPU；First picture
Element value is the sum of the pixel of every a line pixel in the benchmark remote sensing images.
Second computation subunit, for using column prefix addition algorithm to sum first pixel value in GPU,
Obtain benchmark remote sensing integral image.
Third computation subunit, for calculating the second pixel value using row prefix addition algorithm in GPU；Second picture
Element value is the sum of the pixel of every a line pixel in the remote sensing integral image subject to registration.
4th computation subunit, for using column prefix addition algorithm to sum second pixel value in GPU,
Obtain remote sensing integral image subject to registration.
As an alternative embodiment, the characteristic detection unit, specifically includes:
5th computation subunit, for carrying out convolution algorithm to the benchmark remote sensing integral image using box filter,
Obtain the scale space of the benchmark remote sensing integral image.
First detection sub-unit carries out feature detection for the scale space to the benchmark remote sensing integral image, obtains
Fisrt feature parameter.
6th computation subunit, for carrying out convolution fortune to the remote sensing integral image subject to registration using box filter
It calculates, obtains the scale space of the remote sensing integral image subject to registration.
Second detection sub-unit carries out feature detection for the scale space to the remote sensing integral image subject to registration, obtains
To second feature parameter.
The remote sensing image registration system based on SUFR algorithm of the present embodiment is used with parallel calculation in GPU
SURF algorithm realizes feature extraction, realizes characteristic matching using quick approximate KNN searching algorithm, can be improved remote sensing images
The speed of service of registration, and then improve registration efficiency.
For the system disclosed in the embodiment, since it is corresponded to the methods disclosed in the examples, so the ratio of description
Relatively simple, reference may be made to the description of the method.
Used herein a specific example illustrates the principle and implementation of the invention, and above embodiments are said
It is bright to be merely used to help understand method and its core concept of the invention；At the same time, for those skilled in the art, foundation
Thought of the invention, there will be changes in the specific implementation manner and application range.In conclusion the content of the present specification is not
It is interpreted as limitation of the present invention.
Claims (10)
1. a kind of remote sensing image registration method based on SUFR algorithm characterized by comprising
Benchmark remote sensing images and remote sensing images subject to registration are uploaded into GPU video memory；
With parallel calculation in the GPU, using SURF algorithm respectively to benchmark remote sensing images and described wait match
Quasi- remote sensing images carry out feature extraction, obtain fisrt feature description and second feature description；Fisrt feature description
For the Feature Descriptor of the benchmark remote sensing images, second feature description is that the feature of the remote sensing images subject to registration is retouched
State son；
Son is described with the second feature to fisrt feature description using quick approximate KNN searching algorithm in GPU
Characteristic matching is carried out, characteristic matching image is obtained；
The characteristic matching image is downloaded in CPU memory, the image after being registrated.
2. a kind of remote sensing image registration method based on SUFR algorithm according to claim 1, which is characterized in that it is described
With parallel calculation in the GPU, using SURF algorithm respectively to the benchmark remote sensing images and the remote sensing subject to registration
Image carries out feature extraction, obtains fisrt feature description and second feature description, specifically includes:
The benchmark remote sensing images and the remote sensing images subject to registration are accumulated respectively by the way of parallel computation in GPU
Divide processing, obtains benchmark remote sensing integral image and remote sensing integral image subject to registration；
Feature inspection is carried out to the benchmark remote sensing integral image and the remote sensing integral image subject to registration respectively using SURF algorithm
It surveys, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter be to the benchmark remote sensing integral image into
The location parameter and scale parameter for the characteristic point that row characteristic point detects；The second feature parameter is to described subject to registration distant
Sense integral image carries out the location parameter and scale parameter for the characteristic point that feature detects；
Fisrt feature principal direction is calculated according to the fisrt feature parameter, calculates second feature master according to the second feature parameter
Direction；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images, and the second feature parameter is described
The feature principal direction of remote sensing images subject to registration；
Fisrt feature description is calculated according to the fisrt feature principal direction, it is special to calculate second according to the second feature principal direction
Sign description.
3. a kind of remote sensing image registration method based on SUFR algorithm according to claim 1, which is characterized in that described to incite somebody to action
The characteristic matching image downloads in CPU memory, and the image after being registrated specifically includes:
The characteristic matching image is downloaded in CPU memory, and perspective matrix is carried out to the characteristic matching image in CPU memory
Transformation；
Image after transformed characteristic matching image to be determined as to registration.
4. a kind of remote sensing image registration method based on SUFR algorithm according to claim 2, which is characterized in that it is described
The benchmark remote sensing images and the remote sensing images subject to registration are carried out at integral respectively by the way of parallel computation in GPU
Reason, obtains benchmark remote sensing integral image and remote sensing integral image subject to registration, specifically includes:
The first pixel value is calculated using row prefix addition algorithm in GPU；First pixel value is the benchmark remote sensing images
In every a line pixel the sum of pixel；
It is summed using column prefix addition algorithm to first pixel value in GPU, obtains benchmark remote sensing integral image；
The second pixel value is calculated using row prefix addition algorithm in GPU；Second pixel value is the remote sensing product subject to registration
The sum of the pixel of every a line pixel in partial image；
It is summed using column prefix addition algorithm to second pixel value in GPU, obtains remote sensing integral image subject to registration.
5. a kind of remote sensing image registration method based on SUFR algorithm according to claim 2, which is characterized in that described to adopt
Feature detection is carried out to the benchmark remote sensing integral image and the remote sensing integral image subject to registration respectively with SURF algorithm, is obtained
Fisrt feature parameter and second feature parameter, specifically include:
Convolution algorithm is carried out to the benchmark remote sensing integral image using box filter, obtains the benchmark remote sensing integral image
Scale space；
Feature detection is carried out to the scale space of the benchmark remote sensing integral image, obtains fisrt feature parameter；
Convolution algorithm is carried out to the remote sensing integral image subject to registration using box filter, obtains the remote sensing integral subject to registration
The scale space of image；
Feature detection is carried out to the scale space of the remote sensing integral image subject to registration, obtains second feature parameter.
6. a kind of remote sensing image registration system based on SUFR algorithm characterized by comprising
Image uploading module, for uploading to benchmark remote sensing images and remote sensing images subject to registration in GPU video memory；
Characteristic extracting module, in the GPU with parallel calculation, it is distant to the benchmark respectively using SURF algorithm
Feel image and the remote sensing images subject to registration carry out feature extraction, obtains fisrt feature description and second feature description；Institute
The Feature Descriptor that fisrt feature description is the benchmark remote sensing images is stated, second feature description is described subject to registration
The Feature Descriptor of remote sensing images；
Characteristic matching module, in GPU using quick approximate KNN searching algorithm to the fisrt feature description son with
Second feature description carries out characteristic matching, obtains characteristic matching image；
Download module, for the characteristic matching image to be downloaded in CPU memory, the image after being registrated.
7. a kind of remote sensing image registration system based on SUFR algorithm according to claim 6, which is characterized in that the spy
Extraction module is levied, is specifically included:
Integral unit, in GPU by the way of parallel computation respectively to benchmark remote sensing images and described subject to registration
Remote sensing images carry out Integral Processing, obtain benchmark remote sensing integral image and remote sensing integral image subject to registration；
Characteristic detection unit, for long-pending to the benchmark remote sensing integral image and the remote sensing subject to registration respectively using SURF algorithm
Partial image carries out feature detection, obtains fisrt feature parameter and second feature parameter；The fisrt feature parameter is to the base
Quasi- remote sensing integral image carries out the location parameter and scale parameter for the characteristic point that characteristic point detects；The second feature parameter
For the location parameter and scale parameter for carrying out the characteristic point that feature detects to the remote sensing integral image subject to registration；
First computing unit, for calculating fisrt feature principal direction according to the fisrt feature parameter, according to the second feature
Parameter calculates second feature principal direction；The fisrt feature principal direction is the feature principal direction of the benchmark remote sensing images, described
Second feature parameter is the feature principal direction of the remote sensing images subject to registration；
Second computing unit, it is special according to described second for calculating fisrt feature description according to the fisrt feature principal direction
It levies principal direction and calculates second feature description.
8. a kind of remote sensing image registration system based on SUFR algorithm according to claim 6, which is characterized in that under described
Module is carried, is specifically included:
Converter unit, for downloading to the characteristic matching image in CPU memory, and to the characteristic matching figure in CPU memory
As carrying out perspective matrix transformation；
Determination unit, for transformed characteristic matching image to be determined as the image after being registrated.
9. a kind of remote sensing image registration system based on SUFR algorithm according to claim 7, which is characterized in that the product
Sub-unit specifically includes:
First computation subunit, for calculating the first pixel value using row prefix addition algorithm in GPU；First pixel value
For the sum of the pixel of every a line pixel in the benchmark remote sensing images；
Second computation subunit is obtained for using column prefix addition algorithm to sum first pixel value in GPU
Benchmark remote sensing integral image；
Third computation subunit, for calculating the second pixel value using row prefix addition algorithm in GPU；Second pixel value
For the sum of the pixel of every a line pixel in the remote sensing integral image subject to registration；
4th computation subunit is obtained for using column prefix addition algorithm to sum second pixel value in GPU
Remote sensing integral image subject to registration.
10. a kind of remote sensing image registration system based on SUFR algorithm according to claim 7, which is characterized in that described
Characteristic detection unit specifically includes:
5th computation subunit is obtained for carrying out convolution algorithm to the benchmark remote sensing integral image using box filter
The scale space of the benchmark remote sensing integral image；
First detection sub-unit carries out feature detection for the scale space to the benchmark remote sensing integral image, obtains first
Characteristic parameter；
6th computation subunit is obtained for carrying out convolution algorithm to the remote sensing integral image subject to registration using box filter
To the scale space of the remote sensing integral image subject to registration；
Second detection sub-unit carries out feature detection for the scale space to the remote sensing integral image subject to registration, obtains the
Two characteristic parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910354697.5A CN110060285A (en) | 2019-04-29 | 2019-04-29 | A kind of remote sensing image registration method and system based on SURF algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910354697.5A CN110060285A (en) | 2019-04-29 | 2019-04-29 | A kind of remote sensing image registration method and system based on SURF algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110060285A true CN110060285A (en) | 2019-07-26 |
Family
ID=67321613
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910354697.5A Pending CN110060285A (en) | 2019-04-29 | 2019-04-29 | A kind of remote sensing image registration method and system based on SURF algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110060285A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796104A (en) * | 2019-11-01 | 2020-02-14 | 深圳市道通智能航空技术有限公司 | Target detection method and device, storage medium and unmanned aerial vehicle |
| CN111861883A (en) * | 2020-06-23 | 2020-10-30 | 燕山大学 | Multi-channel video splicing method based on synchronous integral SURF algorithm |
| CN113724300A (en) * | 2020-05-25 | 2021-11-30 | 北京达佳互联信息技术有限公司 | Image registration method and device, electronic equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103426186A (en) * | 2013-09-05 | 2013-12-04 | 山东大学 | Improved SURF fast matching method |
| CN104794703A (en) * | 2015-03-23 | 2015-07-22 | 中国科学技术大学先进技术研究院 | Real-time stereo matching system and method based on ZNCC algorithm |
| CN105069743A (en) * | 2015-07-28 | 2015-11-18 | 中国科学院长春光学精密机械与物理研究所 | Detector splicing real-time image registration method |
| CN106204576A (en) * | 2016-07-05 | 2016-12-07 | 董超超 | A kind of image registration device described based on Patch properties |
| CN106971404A (en) * | 2017-03-20 | 2017-07-21 | 西北工业大学 | A kind of robust SURF unmanned planes Color Remote Sensing Image method for registering |
| CN107689058A (en) * | 2017-09-01 | 2018-02-13 | 哈尔滨理工大学 | A kind of image registration algorithm based on SURF feature extractions |
-
2019
- 2019-04-29 CN CN201910354697.5A patent/CN110060285A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103426186A (en) * | 2013-09-05 | 2013-12-04 | 山东大学 | Improved SURF fast matching method |
| CN104794703A (en) * | 2015-03-23 | 2015-07-22 | 中国科学技术大学先进技术研究院 | Real-time stereo matching system and method based on ZNCC algorithm |
| CN105069743A (en) * | 2015-07-28 | 2015-11-18 | 中国科学院长春光学精密机械与物理研究所 | Detector splicing real-time image registration method |
| CN106204576A (en) * | 2016-07-05 | 2016-12-07 | 董超超 | A kind of image registration device described based on Patch properties |
| CN106971404A (en) * | 2017-03-20 | 2017-07-21 | 西北工业大学 | A kind of robust SURF unmanned planes Color Remote Sensing Image method for registering |
| CN107689058A (en) * | 2017-09-01 | 2018-02-13 | 哈尔滨理工大学 | A kind of image registration algorithm based on SURF feature extractions |
Non-Patent Citations (1)
| Title |
|---|
| CSHILIN: ""高维数据的快速最近邻算法FLANN"", 《HTTPS//BLOG.CSDN.NET/CSHILIN/ARTICLE/DETAILS/52107580?UTM_SOURCE=BLOGXGWZ2》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796104A (en) * | 2019-11-01 | 2020-02-14 | 深圳市道通智能航空技术有限公司 | Target detection method and device, storage medium and unmanned aerial vehicle |
| CN113724300A (en) * | 2020-05-25 | 2021-11-30 | 北京达佳互联信息技术有限公司 | Image registration method and device, electronic equipment and storage medium |
| CN111861883A (en) * | 2020-06-23 | 2020-10-30 | 燕山大学 | Multi-channel video splicing method based on synchronous integral SURF algorithm |
| CN111861883B (en) * | 2020-06-23 | 2022-06-14 | 燕山大学 | Multi-channel video splicing method based on synchronous integral SURF algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20210342990A1 (en) | Image coordinate system transformation method and apparatus, device, and storage medium | |
| CN110866977B (en) | Augmented reality processing method, device, system, storage medium and electronic equipment | |
| CN110288547A (en) | Method and apparatus for generating image denoising model | |
| CN110246181B (en) | Anchor point-based attitude estimation model training method, attitude estimation method and system | |
| CN111598993B (en) | Three-dimensional data reconstruction method and device based on multi-view imaging technology | |
| CN111054080B (en) | Method, device and equipment for intelligently detecting perspective plug-in and storage medium thereof | |
| CN110060285A (en) | A kind of remote sensing image registration method and system based on SURF algorithm | |
| CN109242961A (en) | A kind of face modeling method, apparatus, electronic equipment and computer-readable medium | |
| CN108876723B (en) | Method for constructing color background of gray target image | |
| CN108876804B (en) | Matting model training and image matting method, device and system and storage medium | |
| CN104537705B (en) | Mobile platform three dimensional biological molecular display system and method based on augmented reality | |
| CN110717494A (en) | Android mobile terminal indoor scene three-dimensional reconstruction and semantic segmentation method | |
| CN110276317A (en) | A kind of dimension of object detection method, dimension of object detection device and mobile terminal | |
| CN105046213A (en) | Method for augmenting reality | |
| CN107507226A (en) | A kind of method and device of images match | |
| WO2024060978A1 (en) | Key point detection model training method and apparatus and virtual character driving method and apparatus | |
| CN111062362A (en) | Face living body detection model, method, device, equipment and storage medium | |
| CN114627244A (en) | Three-dimensional reconstruction method and device, electronic equipment and computer readable medium | |
| CN113298871B (en) | Map generation method, positioning method, system thereof, and computer-readable storage medium | |
| JP3863014B2 (en) | Object detection apparatus and method | |
| AU2018101640A4 (en) | A system and method for image processing | |
| CN116091871A (en) | Physical countermeasure sample generation method and device for target detection model | |
| Lee et al. | Surface IR reflectance estimation and material recognition using ToF camera | |
| CN113593297B (en) | Parking space state detection method and device | |
| CN115760888A (en) | Image processing method, image processing device, computer and readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information |
Inventor after: Lei Tianjie Inventor after: Zhang Yazhen Inventor after: Cheng Ziyi Inventor after: Qin Jing Inventor after: Gong Adu Inventor after: Li Yang Inventor after: Xu Jing Inventor before: Lei Tianjie Inventor before: Lu Jingxuan Inventor before: Pang Zhiguo Inventor before: Wang Weiping Inventor before: Yang Dielong Inventor before: Zhang Ju Inventor before: Li Shican Inventor before: Qin Jing Inventor before: Feng Jie Inventor before: Li Shuguang Inventor before: Feng Wei Inventor before: Yuan Man Inventor before: Yang Huichen Inventor before: Zhao Chun Inventor before: Gong Adu Inventor before: Li Aili Inventor before: Zhou Yuanjing Inventor before: Wang Yang Inventor before: Liu Zhongwei Inventor before: Wan Jinhong Inventor before: Xu Jing Inventor before: Cheng Hui Inventor before: Zhang Yazhen Inventor before: Li Yang Inventor before: Wang Jiabao Inventor before: Li Xiangyu Inventor before: Cheng Ziyi Inventor before: Huang Jintao |
|
| CB03 | Change of inventor or designer information |