CN107025664A - Domestic remote sensing satellite rapid registering method based on parallel computation - Google Patents
Domestic remote sensing satellite rapid registering method based on parallel computation Download PDFInfo
- Publication number
- CN107025664A CN107025664A CN201710247014.7A CN201710247014A CN107025664A CN 107025664 A CN107025664 A CN 107025664A CN 201710247014 A CN201710247014 A CN 201710247014A CN 107025664 A CN107025664 A CN 107025664A
- Authority
- CN
- China
- Prior art keywords
- image
- multispectral
- remote sensing
- panchromatic
- registration
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10036—Multispectral image; Hyperspectral image
-
- 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
- G06T2207/10041—Panchromatic image
Landscapes
- Image Processing (AREA)
Abstract
The invention belongs to satellite image processing technology field, more particularly to the domestic remote sensing satellite rapid registering method based on parallel computation, including methods such as a variety of images, it is inadequate that the present invention solves number of control points, skewness is determined in control, by cloud, mist, the influence such as uneven illumination is larger, it is too slow with Quasi velosity, it is not inconsistent the requirement that text handles on the day of day data and provides public service, the handling process method of domestic satellite remote sensing date is less, there is rotation in each wave band of such as China-Brazil Earth Resources Satellite series, translation, the problem of in terms of scaling, with procedure, the processing of each scape satellite image is rapidly completed full-automaticly, the coordinate and Online release map view of unified multi-source satellite image, the atural object deformation problems that different phase different inclination angles are caused, multi-source remote sensing satellite image registration accuracy is high, make full use of hardware resource, improve the advantageous effects of the treatment effeciency of the large-scale parallel registration of mass remote sensing image.
Description
Technical field
The invention belongs to satellite image processing technology field, more particularly to the domestic remote sensing satellite based on parallel computation is quick
Method for registering.
Background technology
The development of internet and mobile Internet, allows remote sensing satellite application to incorporate in popular life, remote sensing spatial information
Application develop also along the direction constantly brought forth new ideas.In face of the active demand and fast development of public's remote sensing, it is necessary to make full use of
Multi-sensor satellite remote sensing collects ability, the full-automatic rapid treating technology of remote sensing, releases remote sensing service towards the public, realizes
The Public Value of remote sensing, the geographical space attribute of uniqueness of remote sensing image causes its application widely.Different resolution
Remote sensing image has its different application category, but we have found unavoidably during understanding in depth, should in middle low resolution
In, domestic remote sensing image occupies most of market;And be but difficult to see domestic satellite in the market for needing high score image
The figure of remote sensing image.However, as the increasing high score satellite of China goes up to the sky, this phenomenon has been taken on a new look.With
The increase of quantity delivered, the market share of domestic satellite-remote-sensing image will be increasing, and the domestic high-definition remote sensing of magnanimity is defended
The satellite image data such as are painted in sing data, including high score series, resource series, environment series, day, how rapidly to produce and reach the standard grade
Issue, as the problem serviced at this stage by the remote sensing of remote sensing image data release towards the public.In process of production, image
Registration is an essential committed step.Multi-source Remote Sensing Images registration is to being derived from different time, different sensors or not
The process for being matched, being superimposed with two width or multiple image of the same scenery at visual angle.In remote sensing application field, image registration is
Realize image co-registration, change detection, the necessary process of image rectification；Towards popular field, image registration is multi- source Remote Sensing Data data
Optimization in the precision of geographical position is there is provided the experience of more accurate and visual and optimal video vision, and image registration is generally comprised
Two steps, the first step is the sufficient amount of registration control points of selection（GCPS）, these control point requested numbers are enough and are distributed
More uniform, second step is that, as reference picture, another width is as registering image, matching somebody with somebody using the width in two images to be matched somebody with somebody
Registering to compare and to analyze with reference picture progress after quasi- image conversion, wherein the first step is extremely fine, due to reality
The satellite remote-sensing image breadth of border processing is very big, and operator has to select substantial amounts of control point, Er Qiebi between two images
It must select quite accurate, if manually carrying out calibration process by professional image software, not only lose time and people
Power, and by the image of people's subjective factor, it is impossible to the precision of registration is effectively ensured, therefore, large format remote sensing images in recent years
Fast accurate registration has obtained the concern and research of scholars,
In order to overcome this shortcoming, the method that scholars propose autoregistration is broadly divided into based on gray scale, transform domain and feature
Three major types, but in actual applications, it is usually present following several shortcomings：Number of control points not enough, control determine skewness, by
It is larger, too slow with Quasi velosity to influences such as cloud, mist, uneven illuminations, it is not inconsistent text and is handled on the day of day data and public service is provided
Requirement, domestic satellite remote sensing date handling process method it is less, such as China-Brazil Earth Resources Satellite series each wave band deposit
The problem of in terms of rotation, translation, scaling.
The content of the invention
The present invention provides the domestic remote sensing satellite rapid registering method based on parallel computation, to solve in above-mentioned background technology
Number of control points not enough, determine skewness, influenceed larger, too slow with Quasi velosity by cloud, mist, uneven illumination etc., is not inconsistent by control
Text handles on the day of day data and provides the requirement of public service, the handling process method of domestic satellite remote sensing date is less, than
The problem of such as the serial each wave band of China-Brazil Earth Resources Satellite in the presence of in terms of rotation, translation, scaling.
Technical problem solved by the invention is realized using following technical scheme：Domestic remote sensing based on parallel computation is defended
Star rapid registering method, the method for registering includes：
When individually handling non-multi-source remote sensing satellite multispectral image, the non-multi-source remote sensing satellite multispectral image is through long-range mistake
Journey invocation protocol RPC processing forms image before the multispectral registration of non-multi-source remote sensing satellite, and the non-multi-source remote sensing satellite is multispectral
Image combines through task distribution and after converting calibration process and refers to image output before registering；
When individually processing multi-source remote sensing satellite multispectral image, the multi-source remote sensing satellite multispectral image is adjusted through remote process
Shadow before image before the multispectral registration of multi-source remote sensing satellite, the multispectral registration of multi-source remote sensing satellite is formed with agreement RPC processing
Image output is referred to as being combined after task distribution and conversion calibration process after secondary registration process；
When handling non-optical multispectral and panchromatic image, the multispectral image and panchromatic image RPC processing of respectively hanging oneself are formed
Image is fused before image and panchromatic fusion before image before non-optical multispectral and panchromatic fusion, the non-optical Multi-spectral image fusion
Image before non-optical multispectral and panchromatic registration is formed after processing, the described non-optical multispectral and panchromatic preceding image of registration is through task
Combined after distribution and conversion calibration process and refer to image output；
When handling multispectral optics and panchromatic image, the optics multispectral image and panchromatic image are respectively hung oneself after RPC processing
Image before the multispectral and panchromatic fusion of optics is formed after matched processing again, image and panchromatic is melted before the optics Multi-spectral image fusion
Image before image before multispectral and panchromatic registration, the multispectral and panchromatic registration of the optics is formed before closing after the fused processing of image
Combined through task distribution and after converting calibration process and refer to image output.
Further, the matching treatment includes：
Image before multispectral and panchromatic matching is formed, the matched processing of image before image and panchromatic matching before the multispectral matching
And handled in the method for B-spline elastic registrating for resource series satellite.
Further, the secondary registration process step includes：
Image and multispectral reference image, which are distributed and converted through task, after the multi-source remote sensing satellite multispectral image registration calibrates
Combined after processing and refer to image output；
If conversion calibration process is exported successfully, effective image feature data storage is retained after cloud removing in Grid square
In storehouse；
If converting calibration process output failure, enter failure queue, the then effective shadow of reservation stored in grid data bank
The calibration secondary carry out task distribution of output data of matching affine transformation and conversion calibration as characteristic and in failure queue
Exported after processing.
Further, task distribution and conversion calibration process include the distribution of concurrent processing framework MapReduce tasks and
Algorithmic match affine transformation calibrates SIFT＆Ransac.
Further, the RPC processing is that possess thick matching precision after RPC combinations digital complex demodulation is handled and eliminate
Terrain error.
Further, the multispectral image include high score series 16m multispectral images, environment series 30m multispectral images,
Multi-source remote sensing satellite multispectral image.
Further, described multispectral and panchromatic image includes high score series 8m multispectral and 2m panchromatic images, high score series
3.2m is multispectral and 0.8m panchromatic images, multispectral optics and panchromatic image, day paint that serial 10m is multispectral and the panchromatic shadows of 2.5m
Picture.
Further, the algorithmic match affine transformation calibration includes：Piecemeal will be carried out with registering image read in caching, according to
Piecemeal image overlap area with registration extends out certain limit and read in caching, caching with registering piecemeal image after stream process
Internal memory is read in successively, and the piecemeal image with registration in internal memory is synchronously transmitted when kernel function is run, and the band after output is matched somebody with somebody
Accurate piecemeal image is calculated and accelerated to metric space structure and characteristic point through unifiedly calculating equipment framework CUDA texture features
It is described and the result for matching distance matrix is read in and buffered, and the result data that distance matrix is matched is exported to main frame end memory
The result of distance matrix matching carries out excluding Exceptional point and carries out affine transformation or spline function through multiple threads function
Conversion, the image after the data correction after conversion is exported to complete registration.
Further, the reference image is world Fig. 9 m image；
Beneficial effects of the present invention are：
1 this patent includes series 16m multispectral images, environment series 30m multispectral images, multi-source remote sensing using multispectral image
Satellite 9m multispectral images and multispectral and panchromatic image include high score series 8m multispectral and 2m panchromatic images, high score series
3.2m is multispectral and 0.8m panchromatic images, multispectral optics and panchromatic image, day paint that serial 10m is multispectral and 2.5m panchromatic images
Deng the technological means of the processing of satellite image, due to according to different domestic satellite images, using multinomial Rigid Registration and sample
Bar function non-rigid registration is combined so that domestic satellite data improves matching somebody with somebody for multi-source different resolution in registration process
Quasi- precision, in combination with concurrent processing framework MapReduce, CPU multithreading and the SIFT of many stream handle parallel computations of GPU
+ Ransac feature matching methods, with solving the influence of weather and atural object deformation according to the mode of SIFT feature, and flow
Cheng Hua, it is full-automatic be rapidly completed the processing of each scape satellite image with realize work as day data on the day of the requirement reached the standard grade of processing.
2 this patents use the technological means for world Fig. 9 m image with reference to image, because the geography of WGS-84 ellipsoids is sat
Mark system, the satellite image data for the use of the resolution ratio of day map being 9m with reference to image data source pass through the spy with registering image
Levy and calculate and the feature calculation with the geographical overlapping region of registering image and the reference image extended out, statistical match control point, progress
Registration task, it is necessary to according to the characteristic of different data sources, build whole vertical different under the support of the domestic remote sensing satellite data of magnanimity
Computation model is handled, handles and closes to complete the rapid registering of multi-source data, reaches that Quick thread issue provides public service
It is required that, its advantageous effects is advantageous for the coordinate and Online release map view of unified multi-source satellite image.
3 this patents are used when handling multispectral optics and panchromatic image, the optics multispectral image and panchromatic image
Image before the multispectral and panchromatic fusion of optics is formed after matched processing again after the RPC that respectively hangs oneself processing, the optics is multispectral to be melted
Close and form image before multispectral and panchromatic registration before preceding image and panchromatic fusion after the fused processing of image, the optics is multispectral
With reference to the technological means with reference to image output after distributing through task with image before panchromatic registration and convert calibration process, due to resource
The optical image of series includes number 02C of resource and resource three, and China-Brazil Earth Resources Satellite is China's first generation mode transmission
Ball landsat, due to multispectral image and panchromatic image breadth not into it is whole than, shoot that inclination angle is different, shooting time have compared with
Big difference, so need first to match multispectral and panchromatic image before fusion treatment is carried out, and with B-spline elasticity
The method of registration is handled for resource series satellite, and the atural object deformation caused with solving different phase different inclination angles is asked
Topic.
When 4 this patents use multi-source remote sensing satellite multispectral image, the multi-source remote sensing satellite multispectral image is through long-range
Invocation of procedure agreement RPC processing forms image before the multispectral registration of multi-source remote sensing satellite, and the multi-source remote sensing satellite is multispectral to match somebody with somebody
With reference to the technological means with reference to image output after secondary registration process after image is distributed through task and converts calibration process before accurate,
In multi-source remote sensing satellite image registration process, due to secondary registration approach, reference is used as using world Fig. 9 m image
Image, its phase typically with current gap 1-3, some areas feature changes than more serious, and have one in resolution ratio
Determine difference, there is certain probability can not complete registering or registration accuracy by way of Feature Points Matching relatively low so that multi-source is distant
Feel satellite image registration accuracy high.
5 this patents are included using algorithmic match affine transformation calibration：Piecemeal will be carried out with registering image read in caching,
Certain limit is extended out according to the piecemeal image overlap area with registration and reads in caching, in caching the registering piecemeal image of band through stream at
Internal memory is read in after reason successively, the piecemeal image with registration in internal memory is synchronously transmitted when kernel function is run, after output
Through unifiedly calculating, equipment framework CUDA texture features are built piecemeal image with registration to metric space and characteristic point is calculated simultaneously
The result for accelerating and matching distance matrix reads in buffering, and the result data that distance matrix is matched is exported to main frame end memory,
The result of the distance matrix matching carries out excluding Exceptional point and carries out affine transformation or batten through multiple threads function
Functional transformation, the image after the data correction after conversion is exported to complete the technological means of registration, due to unified calculation equipment
Framework（Compute Unified Device Architecture, CUDA）It is being developed for video card for NVIDIA companies production
General-purpose computations system, its equipment framework is described as single instrction multithreading by NVIDIA（Single Instruction
Multiple Thread, SIMT）, this equipment possesses the internal storage structure system and a variety of links mechanism of complexity, is program
Maximized broadband is provided to support.CUDA particular architectures cause it to be advantageous to handle view data, can greatly carry
Hi-vision data-handling efficiency.Pointwise computing is essentially calculating when SIFT feature is extracted with distance matrix matching, is conducive to
GPU computings, because GPU does not introduce substantial amounts of caching and branch prediction mechanism as CPU so that GPU is not suitable for more
Branch operations, thus in the step such as affine transformation and Ransac using intelTBB frameworks CPU multi-threading parallel process.This
Text will take the SIFT algorithms part progress CUDA parallelization processing for accounting for 80%, and remainder is carried out many using intelTBB frameworks
Thread process, makes full use of hardware resource.
6 this patents use MapReduce technological means, because MapReduce is Google answering in proposition in 2004
The parallel computational model of data processing is carried out for large-scale cluster, is also the core calculations pattern of current cloud computing, is utilized
MapReduce improves the treatment effeciency of the large-scale parallel registration of mass remote sensing image.
Brief description of the drawings
Fig. 1 is the flow chart of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation；
Fig. 2 is that the 16m multispectral images and 8m of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation are more
Spectrum and multispectral and 0.8m panchromatic images the process chart of 2m panchromatic images, No. 2 3.2m of high score；
Fig. 3 be the domestic remote sensing satellite rapid registering method based on parallel computation of the invention 8m is multispectral and 2m panchromatic images
Process chart
Fig. 4 is the multi-source remote sensing satellite multispectral image of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation
Flow chart
Fig. 5 is the algorithmic match affine transformation calibration stream of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation
Cheng Tu
Fig. 6 is the MapReduce task distribution flows of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation
Figure
Fig. 7 is the MapReduce task distributing structures of the domestic remote sensing satellite rapid registering method of the invention based on parallel computation
Figure.
Embodiment
The present invention is described further below in conjunction with accompanying drawing：
In figure：The non-multi-source remote sensing satellite multispectral image method of 1- processing, 2- processing multi-source remote sensing satellite multispectral image methods,
The non-optical multispectral and panchromatic image method of 3- processing, 4- processing optics is multispectral and panchromatic image method
Embodiment：
The present embodiment includes：As shown in figure 1, the domestic remote sensing satellite rapid registering method based on parallel computation, the registration side
Method includes, when individually handling non-multi-source remote sensing satellite multispectral image, and the non-multi-source remote sensing satellite multispectral image is through remote
Journey invocation of procedure agreement RPC processing forms image before the multispectral registration of non-multi-source remote sensing satellite, and the non-multi-source remote sensing satellite is more
Image combines through task distribution and after converting calibration process and refers to image output 1 before spectrum registration；
Environment series has 30m's multispectral on visible image, and registering flow refers to No. 1 16m of high score handling process.
As shown in Fig. 2 high score series includes No. 1 16m multispectral image of high score, its RPC parameter combination DEM digital elevation mould
Type, can make raw video possess thick matching precision and eliminate terrain error, with panchromatic image have into whole ratio due to multispectral
Breadth and close shooting time, its merge can RPC processing after perform, next only need to fusion evaluation carry out phase
Close registration process.
Its day for painting series paint that No. 1 10m is multispectral and 2.5m it is panchromatic have with No. 1 multispectral panchromatic image of high score it is similar
Property, so registration task can be completed using similar handling process.
As shown in figure 4, when individually processing multi-source remote sensing satellite multispectral image, the multispectral shadow of multi-source remote sensing satellite
As forming image, the multi-source remote sensing satellite before the multispectral registration of multi-source remote sensing satellite through remote procedure call protocol RPC processing
Image combines after secondary registration process through task distribution and after converting calibration process and refers to image output 2 before multispectral registration；
In multi-source remote sensing satellite image registration process, because the image using world Fig. 9 m is as referring to image, its phase one
As with current gap 1-3, some areas feature changes than more serious, and have certain difference in resolution ratio, pass through spy
Levy the mode of Point matching have certain probability can not complete registration or registration accuracy it is relatively low.
As shown in Fig. 2 when handling non-optical multispectral and panchromatic image, the multispectral image and panchromatic image are each
Image is formed before image before non-optical multispectral and panchromatic fusion, the non-optical Multi-spectral image fusion through RPC processing and panchromatic is melted
Form the non-optical multispectral and panchromatic preceding image of registration before closing after image fused processing, it is described non-optical multispectral and panchromatic match somebody with somebody
Image combines through task distribution and after converting calibration process and refers to image output 3 before accurate；
No. 1 8m of high score is multispectral and 2m panchromatic images, and No. 2 3.2m of high score are multispectral and 0.8m panchromatic images, and its RPC parameter is combined
DEM digital elevation models, can make raw video possess thick matching precision and eliminate terrain error, due to multispectral and panchromatic shadow
As having into the breadth and close shooting time of whole ratio, it, which is merged, to be performed after RPC processing, next only needed to melting
Group photo is as carrying out correlation registration processing.
As shown in figure 3, when handling multispectral optics and panchromatic image, the optics multispectral image and panchromatic image are each
Image, the optics Multi-spectral image fusion before the multispectral and panchromatic fusion of optics are formed after matched processing again after the RPC that hangs oneself processing
Formed before preceding image and panchromatic fusion after the fused processing of image it is multispectral and it is panchromatic registration before image, the optics it is multispectral and
Image combines through task distribution and after converting calibration process and refers to image output 4 before panchromatic registration.
Because the optical image of resource series includes number 02C of resource and resource three, China-Brazil Earth Resources Satellite is for I
State's first generation mode transmission earth resources satellite, due to multispectral image and panchromatic image breadth into it is not whole ratio, shoot inclination angle
Different, shooting time has larger difference, so being needed before fusion treatment is carried out first to the progress of multispectral and panchromatic image
Match somebody with somebody, and handled in the method for B-spline elastic registrating for resource series satellite, to solve different phase different inclination angles
The atural object deformation problems caused.
It is distant due to including series 16m multispectral images, environment series 30m multispectral images, multi-source using multispectral image
Feeling satellite 9m multispectral images and multispectral and panchromatic image includes high score series 8m multispectral and 2m panchromatic images, high score series
3.2m is multispectral and 0.8m panchromatic images, multispectral optics and panchromatic image, day paint that serial 10m is multispectral and 2.5m panchromatic images
Deng the technological means of the processing of satellite image, due to according to different domestic satellite images, using multinomial Rigid Registration and sample
Bar function non-rigid registration is combined so that domestic satellite data improves matching somebody with somebody for multi-source different resolution in registration process
Quasi- precision, in combination with concurrent processing framework MapReduce, CPU multithreading and the SIFT of many stream handle parallel computations of GPU
+ Ransac feature matching methods, with solving the influence of weather and atural object deformation according to the mode of SIFT feature, and flow
Cheng Hua, it is full-automatic be rapidly completed the processing of each scape satellite image with realize work as day data on the day of the requirement reached the standard grade of processing.
The matching treatment includes：
Image before multispectral and panchromatic matching is formed, the matched processing of image before image and panchromatic matching before the multispectral matching
And handled in the method for B-spline elastic registrating for resource series satellite.
The secondary registration process step includes：
Image and multispectral reference image, which are distributed and converted through task, after the multi-source remote sensing satellite multispectral image registration calibrates
Combined after processing and refer to image output；
If conversion calibration process is exported successfully, effective image feature data storage is retained after cloud removing in Grid square
In storehouse；
If converting calibration process output failure, enter failure queue, the then effective shadow of reservation stored in grid data bank
The calibration secondary carry out task distribution of output data of matching affine transformation and conversion calibration as characteristic and in failure queue
Exported after processing.
In multi-source remote sensing satellite image registration process, because the image using world Fig. 9 m is as image is referred to, at that time
Mutually typically with current gap 1-3, some areas feature changes than more serious, and have certain difference in resolution ratio, it is logical
Crossing the mode of Feature Points Matching has that certain probability can not complete registration or registration accuracy is relatively low, and multi- source Remote Sensing Data data is by matching somebody with somebody
After quasi- task, successfully image needs by cloud removing registration, retains effective image feature data storage in Grid square
In storehouse, if failure queue will be entered after registration failure, reference is used as using the grid feature data in grid data bank
Image, carries out secondary registration, to complete registration task.
The task distribution and conversion calibration process include concurrent processing framework MapReduce tasks and distributed and algorithmic match
Affine transformation calibrates SIFT＆Ransac.
The RPC processing is that possess thick matching precision after RPC combinations digital complex demodulation is handled and eliminate landform to miss
Difference.
Domestic remote sensing satellite data mainly include the image numbers such as high score series, resource series, environment are serial, day is painted at this stage
According to because different data sources has resolution ratio and the difference of multiband shooting time, not causing in processing highway route design not
Together, high score series includes No. 1 16m multispectral image of high score, and No. 1 8m of high score is multispectral and 2m panchromatic images, and No. 2 3.2m of high score are more
Spectrum and 0.8m panchromatic images.High score series of satellites image, its RPC parameter combination DEM digital elevation model, can make original shadow
As possessing thick matching precision and eliminating terrain error, with panchromatic image there is into the breadth of whole ratio and close bat due to multispectral
The time is taken the photograph, it, which is merged, to be performed after RPC processing, next only need to carry out correlation registration processing to fusion evaluation.
It is distant that the multispectral image includes high score series 16m multispectral images, environment series 30m multispectral images, multi-source
Feel satellite multispectral image.
It is multispectral with 2m panchromatic images, high score series 3.2m light more that the multispectral and panchromatic image includes high score series 8m
Spectrum and 0.8m panchromatic images, multispectral optics and panchromatic image, day paint that serial 10m is multispectral and 2.5m panchromatic images.
As shown in figure 5, the algorithmic match affine transformation calibration includes：Piecemeal will be carried out with registering image read in caching,
Certain limit is extended out according to the piecemeal image overlap area with registration and reads in caching, in caching the registering piecemeal image of band through stream at
Internal memory is read in after reason successively, the piecemeal image with registration in internal memory is synchronously transmitted when kernel function is run, after output
Through unifiedly calculating, equipment framework CUDA texture features are built piecemeal image with registration to metric space and characteristic point is calculated simultaneously
The result for accelerating and matching distance matrix reads in buffering, and the result data that distance matrix is matched is exported to main frame end memory,
The result of the distance matrix matching carries out excluding Exceptional point and carries out affine transformation or batten through multiple threads function
Functional transformation, the image after the data correction after conversion is exported to complete registration.
It is unified meter with reference to the SIFT+Ransac feature matching methods of many stream handles of GPU and CPU multithreads computings
Calculation equipment framework Compute Unified Device Architecture, CUDA are the production of NVIDIA companies for video card
Its equipment framework is described as single instrction multithreading Single Instruction by the general-purpose computations system of exploitation, NVIDIA
Multiple Thread, SIMT, this equipment possess the internal storage structure system and a variety of links mechanism of complexity, are program
There is provided maximized broadband to support, CUDA particular architectures cause it to be advantageous to handle view data, can greatly carry
Hi-vision data-handling efficiency, is essentially pointwise computing calculating when SIFT feature is extracted with distance matrix matching, is conducive to
GPU computings, because GPU does not introduce substantial amounts of caching and branch prediction mechanism as CPU so that GPU is not suitable for more
Branch operations, thus in the step such as affine transformation and Ransac using intelTBB frameworks CPU multi-threading parallel process.This
Text will take the SIFT algorithms part progress CUDA parallelization processing for accounting for 80%, and remainder is carried out many using intelTBB frameworks
Thread process, makes full use of hardware resource.
Threading building module Thread Building Blocks, TBB, are the multiple programming exploitations of Intel Company's exploitation
Instrument.It is a set of C++ ATLs, with following characteristic：1）The abstract of higher degree can be used to task；2）Attach times
Business scheduler program, can efficiently handle load balance across multiple logically and physically kernels；3）With the thread that can be used directly
Safety container；4）The parallel method that parallel_for and parallel_reduce etc. can be used general；5）There is provided without lock
Concurrency programming support；6）C++ is realized, without any extension or using grand., can be effectively and rapidly based on more than TBB characteristics
Extensive code is transformed, is suitable for the secondary acceleration lifting of history codes, has applied well in the project imitative
Penetrate the processing task of the steps such as conversion and Ransac.
The characteristics of SIFT Scale-Invariant Feature Transform methods be rotation to characteristic point and
Yardstick has rotational invariance.This consistency, which is used for remote sensing images, has very high stability, accuracy and identification capability,
A feature is accurately found control point of the same name, be that motion correction provides the foundation.Due to SIFT characteristic, moreover it is possible to reduce
The difference of the atural object of the satellite image of different phases, the influence of cloud and mist weather, are extensive automatically process there is provided more good
Control point of the same name detection solution.RANSAC is " RANdom SAmple Consensus（Random sampling is consistent）"
Abbreviation.It from one group of observation data set comprising " point not in the know ", can estimate the parameter of mathematical modeling by iterative manner.
RANSAC methods are introduced in SIFT, Wrong control point can be rejected, increase registering control points accuracy, root-mean-square error is reduced.
SIFT＆Ransac algorithms are mainly comprised the following steps：1）Metric space extremum detection；2）Local feature point location；3）Direction assignment；4）
Feature Descriptor is generated；5）Feature Points Matching based on distance matrix；6）Ransac removes control errors point；7）Affine transformation mould
Type is calculated.
Domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that
The reference image is world Fig. 9 m image；
Remote sensing image concurrent processing framework based on MapReduce
MapReduce is the parallel computation mould for being applied to large-scale cluster progress data processing that Google was proposed in 2004
Type, is also the core calculations pattern of current cloud computing, and this project is intended to improve the big rule of mass remote sensing image using MapReduce
The treatment effeciency of the parallel registration of mould.
As shown in Figure 6,7, Image registration：By calling HTTP interface transmission algorithm model treatment to ask, job-client:
Remote sensing image processing service is provided in the way of WEB service, the required parameter of user is packaged at a remote sensing image data
Reason task, and task is sent to job-server, job-server in an asynchronous manner:By the task from job-client
Task queue is stored in, and utilizes mysql by queue persistence.Then task is sent in sequence in idle job-worker and located
Reason, waits after job-worker completion tasks, the task is removed from queue, and task result is pushed to client
End, job-worker:Operated in multiple servers, be mutually independent in the way of multi-process.Job-worker, which is received, to be come
Request is handled from job-server task, former data are read from storage, then process image data as desired, and will place
The result of reason is uploaded under the path specified in storage, and SIFT＆Ransac algorithm models use mapreduce parallel computation frames
Frame, realizes many granularity parallel computations of single task.MapReduce processing data processes are broadly divided into 2 stages：The map stages and
The reduce stages.The map stages are first carried out, then perform the reduce stages：
, it is necessary to carry out " burst " to input, each map task handles one " burst " before map functions are formally performed.
After burst, many machines just can carry out map simultaneously and work.Map function pairs data carry out " pretreatment ", defeated
Go out desired " concern ".Map to every record output with<key,value>Pair form output.
Before the reduce stages are entered, also by data related in each map（Key identical data）Sum up in the point that one
Rise, be sent to a reducer.Here the multiple map multiple reducer of output " mixedly " correspondence situation is related to,
This process is called " shuffling ".
Next the reduce stages are entered.Identical key map outputs can reach same reducer.Reducer pairs
The multiple value of key identicals carry out " reduce operations ", and last key a string of value pass through the effect of reduce functions
Afterwards, a value is become.
Operation principle：Invention is a kind of to be based on MapReduce concurrent processing frameworks, with reference to CPU multithreadings and GPU
The SIFT+Ransac feature matching methods of many stream handle parallel computations, and according to different domestic satellite images, using many
Formula Rigid Registration and spline function non-rigid registration are combined so that domestic satellite data improving in registration process is more
The registration accuracy of source different resolution, the influence of weather and atural object deformation is solved according to the mode of SIFT feature, and is flowed
Cheng Hua, it is full-automatic be rapidly completed the processing of each scape satellite image with realize work as day data on the day of the requirement reached the standard grade of processing, this
Invention solve number of control points not enough, control determine skewness, by cloud, mist, uneven illumination etc. influence it is larger, with Quasi velosity
It is too slow, it is not inconsistent text and is handled on the day of day data and the requirement of public service, the handling process side of domestic satellite remote sensing date are provided
Method is less, such as the problem of each wave band of China-Brazil Earth Resources Satellite series is in the presence of in terms of rotation, translation, scaling, with stream
Cheng Hua, it is full-automatic with being rapidly completed the handling of each scape satellite image, the coordinate of unified multi-source satellite image and Online release
Figure browsed, different phase different inclination angles are caused atural object deformation problems, multi-source remote sensing satellite image registration accuracy are high, abundant profit
With hardware resource, the advantageous effects of the treatment effeciency for the large-scale parallel registration for improving mass remote sensing image.
Using technical scheme, or those skilled in the art is under the inspiration of technical solution of the present invention, design
Go out similar technical scheme, and reach above-mentioned technique effect, be to fall into protection scope of the present invention.
Claims (9)
1. the domestic remote sensing satellite rapid registering method based on parallel computation, it is characterised in that：The method for registering includes：
When individually handling non-multi-source remote sensing satellite multispectral image, the non-multi-source remote sensing satellite multispectral image is through long-range mistake
Journey invocation protocol RPC processing forms image before the multispectral registration of non-multi-source remote sensing satellite, and the non-multi-source remote sensing satellite is multispectral
Image combines through task distribution and after converting calibration process and refers to image output before registering；
When individually processing multi-source remote sensing satellite multispectral image, the multi-source remote sensing satellite multispectral image is adjusted through remote process
Shadow before image before the multispectral registration of multi-source remote sensing satellite, the multispectral registration of multi-source remote sensing satellite is formed with agreement RPC processing
Image output is referred to as being combined after task distribution and conversion calibration process after secondary registration process；
When handling non-optical multispectral and panchromatic image, the multispectral image and panchromatic image RPC processing of respectively hanging oneself are formed
Image is fused before image and panchromatic fusion before image before non-optical multispectral and panchromatic fusion, the non-optical Multi-spectral image fusion
Image before non-optical multispectral and panchromatic registration is formed after processing, the described non-optical multispectral and panchromatic preceding image of registration is through task
Combined after distribution and conversion calibration process and refer to image output；
When handling multispectral optics and panchromatic image, the optics multispectral image and panchromatic image are respectively hung oneself after RPC processing
Image before the multispectral and panchromatic fusion of optics is formed after matched processing again, image and panchromatic is melted before the optics Multi-spectral image fusion
Image before image before multispectral and panchromatic registration, the multispectral and panchromatic registration of the optics is formed before closing after the fused processing of image
Combined through task distribution and after converting calibration process and refer to image output.
2. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
Stating matching treatment includes：
Image before multispectral and panchromatic matching is formed, the matched processing of image before image and panchromatic matching before the multispectral matching
And handled in the method for B-spline elastic registrating for resource series satellite.
3. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
Stating secondary registration process step includes：
Image and multispectral reference image, which are distributed and converted through task, after the multi-source remote sensing satellite multispectral image registration calibrates
Combined after processing and refer to image output；
If conversion calibration process is exported successfully, effective image feature data storage is retained after cloud removing in Grid square
In storehouse；
If converting calibration process output failure, enter failure queue, the then effective shadow of reservation stored in grid data bank
The calibration secondary carry out task distribution of output data of matching affine transformation and conversion calibration as characteristic and in failure queue
Exported after processing.
4. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
Stating task distribution and conversion calibration process includes the distribution of concurrent processing framework MapReduce tasks and algorithmic match affine transformation school
Quasi- SIFT＆Ransac.
5. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
It is to possess thick matching precision after RPC combinations digital complex demodulation is handled and eliminate terrain error to state RPC processing.
6. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
Stating multispectral image includes high score series 16m multispectral images, environment series 30m multispectral images, multi-source remote sensing satellite light more
Compose image.
7. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
State multispectral and panchromatic image include high score series 8m it is multispectral with 2m panchromatic images, high score series 3.2m is multispectral and 0.8m is complete
Color image, optics is multispectral and panchromatic image, day paint that serial 10m is multispectral and 2.5m panchromatic images.
8. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
Stating the calibration of algorithmic match affine transformation includes：Piecemeal will be carried out with registering image and read in caching, according to the piecemeal image with registration
Overlapping region extends out the piecemeal image with registration in certain limit reading caching, caching and reads in internal memory successively after stream process,
Piecemeal image with registration in internal memory is synchronously transmitted when kernel function is run, and the piecemeal image of the band registration after output is through system
One computing device framework CUDA texture features are built to metric space and characteristic point is calculated and accelerated and matched distance matrix
Result read in buffering, the result data that distance matrix is matched is exported to main frame end memory, the knot of the distance matrix matching
Fruit carries out excluding Exceptional point and progress affine transformation through multiple threads function or spline function is converted, by the number after conversion
Export to complete registration according to the image after correction.
9. the domestic remote sensing satellite rapid registering method according to claim 1 based on parallel computation, it is characterised in that institute
State with reference to the image that image is world Fig. 9 m.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710247014.7A CN107025664A (en) | 2017-04-17 | 2017-04-17 | Domestic remote sensing satellite rapid registering method based on parallel computation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710247014.7A CN107025664A (en) | 2017-04-17 | 2017-04-17 | Domestic remote sensing satellite rapid registering method based on parallel computation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107025664A true CN107025664A (en) | 2017-08-08 |
Family
ID=59527534
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710247014.7A Pending CN107025664A (en) | 2017-04-17 | 2017-04-17 | Domestic remote sensing satellite rapid registering method based on parallel computation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107025664A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107798731A (en) * | 2017-11-10 | 2018-03-13 | 泰瑞数创科技(北京)有限公司 | A kind of method based on satellite image automatic modeling |
| CN108804220A (en) * | 2018-01-31 | 2018-11-13 | 中国地质大学（武汉） | A method of the satellite task planning algorithm research based on parallel computation |
| CN110276712A (en) * | 2018-12-29 | 2019-09-24 | 中国科学院软件研究所 | Spaceborne image processing apparatus and spaceborne image procossing and decision-making platform |
-
2017
- 2017-04-17 CN CN201710247014.7A patent/CN107025664A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107798731A (en) * | 2017-11-10 | 2018-03-13 | 泰瑞数创科技(北京)有限公司 | A kind of method based on satellite image automatic modeling |
| CN108804220A (en) * | 2018-01-31 | 2018-11-13 | 中国地质大学（武汉） | A method of the satellite task planning algorithm research based on parallel computation |
| CN110276712A (en) * | 2018-12-29 | 2019-09-24 | 中国科学院软件研究所 | Spaceborne image processing apparatus and spaceborne image procossing and decision-making platform |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10692244B2 (en) | Learning based camera pose estimation from images of an environment | |
| CN112862874B (en) | Point cloud data matching method and device, electronic equipment and computer storage medium | |
| US10762620B2 (en) | Deep-learning method for separating reflection and transmission images visible at a semi-reflective surface in a computer image of a real-world scene | |
| CN107679537B (en) | A kind of texture-free spatial target posture algorithm for estimating based on profile point ORB characteristic matching | |
| US9942535B2 (en) | Method for 3D scene structure modeling and camera registration from single image | |
| CN110009705A (en) | Image is created using the mapping for indicating variety classes pixel | |
| CN110176054A (en) | For training the generation of the composograph of neural network model | |
| CN106803267A (en) | Indoor scene three-dimensional rebuilding method based on Kinect | |
| CN112328730A (en) | Map data updating method, related device, equipment and storage medium | |
| CN109936745A (en) | For improving the method and system of the decompression of original video data | |
| CN107025664A (en) | Domestic remote sensing satellite rapid registering method based on parallel computation | |
| US20210287096A1 (en) | Microtraining for iterative few-shot refinement of a neural network | |
| CN105488844B (en) | The display methods of magnanimity model real-time shadow in a kind of three-dimensional scenic | |
| CN108932742A (en) | A kind of extensive infrared terrain scene real-time rendering method based on remote sensing image classification | |
| CN103279974A (en) | High-accuracy high-resolution satellite imaging simulation engine and implementation method | |
| CN113095316B (en) | Image rotation target detection method based on multilevel fusion and angular point offset | |
| CN105913378B (en) | Take the parallel joining method of remote sensing image of level task dependence into account | |
| Lv et al. | Real-time image processing for augmented reality on mobile devices | |
| Adreani et al. | Digital twin framework for smart city solutions | |
| US20220067883A1 (en) | Dynamic image smoothing based on network conditions | |
| CN110930482B (en) | Method and device for determining bone parameters of human hand, electronic equipment and storage medium | |
| CN116912645A (en) | Three-dimensional target detection method and device integrating texture and geometric features | |
| Li et al. | Monocular 3-D Object Detection Based on Depth-Guided Local Convolution for Smart Payment in D2D Systems | |
| CN117132704A (en) | Three-dimensional reconstruction method of dynamic structured light, system and computing equipment thereof | |
| CN106096085A (en) | A kind of radiant energy dfensity analogy method of tower-type solar thermal power generating system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170808 |