CN102663759A - Remote-sensing image rapid-processing method - Google Patents
Remote-sensing image rapid-processing method Download PDFInfo
- Publication number
- CN102663759A CN102663759A CN201210119088XA CN201210119088A CN102663759A CN 102663759 A CN102663759 A CN 102663759A CN 201210119088X A CN201210119088X A CN 201210119088XA CN 201210119088 A CN201210119088 A CN 201210119088A CN 102663759 A CN102663759 A CN 102663759A
- Authority
- CN
- China
- Prior art keywords
- sensing image
- remote sensing
- gpu
- image
- memory
- 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
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a remote-sensing image rapid-processing method. The method comprises the following steps: firstly, judging whether to perform blocking on a remote-sensing image according to sizes of the remote-sensing image and a GPU (Graphics Processing Unit) video memory, and performing step two if the remote-sensing image is required to be subjected to blocking; secondly, dividing the remote-sensing image into a plurality of row blocks which can be processed through a computer memory and the GPU video memory, and respectively performing one-dimensional Fourier row transformation on each row block through the GPU; thirdly, dividing a remote-sensing image into a plurality of line blocks which can be processed through the computer memory and the GPU video memory after processed results are recombined into a whole remote-sensing image, and respectively performing one-dimensional Fourier line transformation on each row block through the GPU; and fourthly, recombining processed results, so as to accomplish the processing of the whole remote-sensing image. The method solves the problem that mass remote-sensing image rapid Fourier transformation is limited to computer configuration.
Description
Technical field
The present invention relates to the remote sensing communications field, refer to a kind of immediate processing method of remote sensing image especially.
Background technology
Fourier transform technique is a basic skills in the Flame Image Process, has important use in fields such as picture coding and compression, figure image intensifying, image restoration, image segmentation, picture noise removals and is worth.Along with the development of remote sensing technology, the remote sensing image data amount of obtaining is increasing, and the wave band number is also more and more.At present, have a lot of image processing methods can realize the fast Fourier transform (FFT) of image, but handle jumbo image often regular meeting the problem that internal memory, video memory overflow appears, can't accomplish processing to remote sensing image.
Summary of the invention
To the problem that prior art exists, the object of the present invention is to provide a kind ofly when handling little image, directly utilize the GPU technology to handle; To jumbo image, carry out the method for the piecemeal processing of image, to adapt to the remote sensing image fast processing of different sizes.
For realizing above-mentioned purpose, remote sensing image immediate processing method of the present invention may further comprise the steps:
1), judges whether remote sensing image is carried out piecemeal according to the size of remote sensing image and GPU video memory; If desired to remote sensing image block, then change step 2 over to);
2) remote sensing image is divided into some calculator memories and the manageable capable piece of GPU video memory, utilizes GPU that every capable piece is carried out one dimension Fourier line translation respectively;
3) the above-mentioned result who handles is reconfigured to behind the view picture remote sensing image, remote sensing image is divided into some calculator memories and the manageable row piece of GPU video memory, utilize GPU that every row piece is carried out one dimension Fourier rank transformation respectively;
4) will pass through the result that above-mentioned steps handles and reconfigure, promptly accomplish processing the view picture remote sensing image.
Further, step 1) is specially: if remote sensing image can be handled by GPU, then remote sensing image is all read in the GPU video memory, utilize the GPU parallel computation to carry out the two-dimension fourier transform of remote sensing image; If remote sensing image is excessive, can not be handled by GPU, then change step 2 over to).
Further, step 2) is specially: at first, remote sensing image is divided into the several rows piece that can be handled by calculator memory, and each the row piece that will cut apart reads in successively in the calculator memory; Secondly, be divided into the plurality of sub row piece that to be handled by GPU with reading in each row piece in the calculator memory, and each the son row piece that will cut apart reads in successively in the GPU video memory; Once more, utilize GPU that each the son row piece that reads in the GPU video memory is carried out the line translation of one dimension Fourier, deposit the result data after the line translation of one dimension Fourier in disk.
Further; Capable piece number according to memory size divided is through formula:
obtains; Wherein b by branch's piece number, p is each pixel shared byte number in internal memory, m; N is the size of image, and a is the free memory of system when handling image.
Further; The capable piece number of being divided according to GPU video memory size is according to formula:
obtains; Wherein B by branch's piece number, P is each pixel shared byte number in video memory, M; N is the size of image, and A is the available video memory of system when handling image.
Further; Step 3) is specially: at first; Remote sensing images through after the above-mentioned processing reconfigure to behind the view picture remote sensing image, remote sensing images are divided into the some row pieces that can be handled by calculator memory, and each row that will cut apart read in successively in the calculator memory soon; Secondly, be divided into the plurality of sub row piece that to be handled by GPU with reading in each row piece in the calculator memory, and each the sub-row piece that will cut apart reads in successively in the GPU video memory; Once more, utilize GPU that each the sub-row piece that reads in the GPU video memory is carried out one dimension Fourier rank transformation, deposit the result data behind the one dimension Fourier rank transformation in disk.
Further; Row piece number according to memory size divided is according to formula:
obtains; Wherein c is institute's apportion piece number, and p is each pixel shared byte number in internal memory, m; N is the size of image, and a is the free memory of system when handling image.
Further; The row piece number of being divided according to GPU video memory size is according to formula:
obtains; Wherein C is institute's apportion piece number, and P is each pixel shared byte number in video memory, M; N is the size of image, and A is the available video memory of system when handling image.
Further, the result data after the one-dimensional Fourier transform is being deposited in disk simultaneously, noting the address of each the piece remote sensing image after the processing.
Remote sensing image immediate processing method of the present invention; Solved the problem that magnanimity remote sensing image Fast Fourier Transform (FFT) is subject to computer configuration; According to image is big or small, calculator memory is big or small and GPU video memory size; Remote sensing images are carried out piecemeal handle, can avoid remote sensing image is significantly being carried out internal memory and video memory overflow problem in the processing procedure.Compare with existing treatment of remote method, have and handle the advantage quick, that applicability is strong, can be to significantly carrying out fast processing with small size remote sensing image.
Embodiment
Remote sensing image immediate processing method of the present invention adopts the piecemeal technology in order to reduce the write operation number of times, and this disposal route is specially: 1) according to the size of remote sensing image and GPU video memory, judge whether remote sensing image is carried out piecemeal.If remote sensing image can be handled by GPU, then remote sensing image is all read in the GPU video memory, utilize the GPU parallel computation to carry out the two-dimension fourier transform of remote sensing image; If remote sensing image is excessive, can not be handled by GPU, then change down the step over to.2) remote sensing image is divided into some calculator memories and the manageable capable piece of GPU video memory, utilizes GPU that every capable piece is carried out one dimension Fourier line translation respectively; For fear of remote sensing image internal memory taking place in reading in the calculator memory process overflows; At first need remote sensing image be divided into the several rows piece that can be handled by calculator memory, and each the row piece that will cut apart reads in the calculator memory successively and handles in order; Because GPU video memory capacity is less, general video memory capacity is all less than memory size, and GPU overflows if video memory greater than GPU video memory capacity, will take place when remote sensing image reads in the GPU video memory remote sensing image behind the piecemeal to the treatment of remote behind piecemeal the time.Therefore; After remote sensing image being pressed calculator memory size piecemeal; Also need to read in each row piece in the calculator memory and be divided into the plurality of sub row piece that to be handled by GPU; And each the son row piece that will cut apart reads in the GPU video memory successively; Utilize GPU that each the son row piece that reads in the GPU video memory is carried out the line translation of one dimension Fourier then, after the one dimension Fourier line translation of accomplishing all row pieces, deposit the result data after the line translation of one dimension Fourier in disk, promptly accomplish the row of view picture remote sensing image is handled.3) the above-mentioned result who handles is reconfigured to behind the view picture remote sensing image, remote sensing image is divided into some calculator memories and the manageable row piece of GPU video memory, utilize GPU that every row piece is carried out one dimension Fourier rank transformation respectively; In like manner; When remote sensing image is carried out one dimension Fourier rank transformation; Overflow with video memory for fear of internal memory takes place, also need carry out by the internal memory piecemeal big or small remote sensing image respectively, before carrying out piecemeal, at first will pass through remote sensing images after the one dimension Fourier line translation processing and reconfigure and be the view picture remote sensing image with video memory; Remote sensing images are divided into the some row pieces that can be handled by calculator memory, and each row that will cut apart read in successively in the calculator memory soon; Secondly, be divided into the plurality of sub row piece that to be handled by GPU with reading in each row piece in the calculator memory, and each the sub-row piece that will cut apart reads in successively in the GPU video memory; Once more, utilize GPU that each the sub-row piece that reads in the GPU video memory is carried out one dimension Fourier rank transformation, deposit the result data behind the one dimension Fourier rank transformation in disk.4) will pass through the result that one dimension Fourier rank transformation handles and reconfigure, and promptly accomplish the Fourier transform of view picture remote sensing image is handled.Result data after the one-dimensional Fourier transform is being deposited in disk simultaneously, can also note the address of each the piece remote sensing image after the processing, so that the reconfiguring fast of the remote sensing image of handling.
Method of the present invention is compared with traditional image FFT conversion process and also to be had treatment of remote advantage fast except can adapting to the advantage to treatment of remote significantly.The algorithm of traditional image FFT conversion is to do line translation earlier; Then the result is made the ranks transposition; Then the result behind the transposition is done line translation again; This step is the rank transformation to original image in fact, and the purpose of transposition is in order to reuse the program of line translation, and the result of last gained also will make a transposition to meet the order that ranks are arranged.Because when row is deposited into row, the data in the delegation are become discrete to each row by continuous storage, when writing disk like this with regard to needs jump carry out; Suppose the wide W of being of original image, height is H, and delegation's transposition is become row; The write operation number of times that needs to carry out is W time; Need write W * H time altogether to the entire image transposition, reduce the operating efficiency of writing greatly, cause remote sensing image carry out FFT when handling speed very slow.And method of the present invention at first is divided into the several rows piece with image, and every has m capable, and result behind the one dimensional fourier transform is divided into the plurality of sub piece, and every comprises the n row.Like this; The number of times of one dimensional fourier transform write operation is
when carrying out rank transformation; Directly corresponding row piece is read in internal memory, carry out one dimensional fourier transform and get final product, needn't carry out the transposition of image; Reduce the processing time greatly, improved treatment effeciency.
The present invention draws according to following formula according to the block count b of internal memory:
The byte number s that reads in each piece image of internal memory is:
Wherein b by branch's piece number, p is each pixel shared byte number in internal memory, m, n are the size of image, the free memory of a system when handling image.Suppose that computing machine is 4 CPU; Internal memory through test 51% uses, and when physical memory is 4G, has 4080713728 bytes; Available physical memory is 1877MB; One 16000 * 12000 image size be 0.54GB then the piece number that divides of image be exactly one, read in 576000000 bytes of byte number of each the piece image in the internal memory, be exactly 549.32MB.One 20000 * 12000 image size be 1.12GB then the piece number that divides of image be exactly two, read in 600000000 bytes of byte number of each the piece image in the internal memory, be exactly 572.20MB.
Because the video memory capacity is less, the piece that therefore also needs will be read in the internal memory is divided into littler sub-piece, could utilize GPU efficiently to handle.When the size of sub-piece can have influence on speed and the reading and writing data of data computing in GPU and the interactive speed of I/O.The sub-piece that divides is too little, and intermediate result is temporary to disk the time, and the number of times of file read-write is just a lot, has increased frequent mutual between the I/O; The sub-piece that divides is too big, will cause video memory to overflow.The calculating of sub-block size is similar with the block size calculating in being read into internal memory, just seldom does huge legendary turtle at this and states.
In order to verify the effect of the inventive method, below list corresponding experimental data and explain:
With HJ-1A Satellite CCD image is example; Handle 16000 * 12000,20000 * 20000,30000 * 20000,40000 * 30000,40000 * 40000,80000 * 80000 respectively; Because the size of actual image also need resample to image and obtain data to be tested less than the size of image in the table.
The processing of adopting two kinds of methods to carry out jumbo image in the experiment is compared.
First method: traditional FFTW method.Main thought is that the intermediate result of Flame Image Process is stored with fixed cache, according to the max cap. of buffer memory, handles the multirow data at every turn, and result is formed one, in buffer memory, carries out transposition earlier, and is written in the disk.
Second method: method of the present invention.
Three kinds of algorithm direct transforms of table 2 (unit: s) operation time
Can find out from table 2, when handling jumbo image, utilize remote sensing image block disposal route of the present invention, can realize the Fourier transform processing of image faster, this method has been given full play to the performance of GPU parallel processing.No matter be normal image, still be jumbo image, can realize the fast processing of image, for utilizing fourier transform technique to carry out big treatment of picture from now on technical support is provided.
Claims (9)
1. remote sensing image immediate processing method wherein, may further comprise the steps:
1), judges whether remote sensing image is carried out piecemeal according to the size of remote sensing image and GPU video memory; If desired to remote sensing image block, then change step 2 over to);
2) remote sensing image is divided into some calculator memories and the manageable capable piece of GPU video memory, utilizes GPU that every capable piece is carried out one dimension Fourier line translation respectively;
3) the above-mentioned result who handles is reconfigured to behind the view picture remote sensing image, remote sensing image is divided into some calculator memories and the manageable row piece of GPU video memory, utilize GPU that every row piece is carried out one dimension Fourier rank transformation respectively;
4) will pass through the result that above-mentioned steps handles and reconfigure, promptly accomplish processing the view picture remote sensing image.
2. remote sensing image immediate processing method as claimed in claim 1; Wherein, Step 1) is specially: if remote sensing image can be handled by GPU, then remote sensing image is all read in the GPU video memory, utilize the GPU parallel computation to carry out the two-dimension fourier transform of remote sensing image; If remote sensing image is excessive, can not be handled by GPU, then change step 2 over to).
3. remote sensing image immediate processing method as claimed in claim 1, wherein, step 2) be specially: at first, remote sensing image is divided into the several rows piece that can be handled by calculator memory, and each the row piece that will cut apart reads in successively in the calculator memory; Secondly, be divided into the plurality of sub row piece that to be handled by GPU with reading in each row piece in the calculator memory, and each the son row piece that will cut apart reads in successively in the GPU video memory; Once more, utilize GPU that each the son row piece that reads in the GPU video memory is carried out the line translation of one dimension Fourier, deposit the result data after the line translation of one dimension Fourier in disk.
4. remote sensing image immediate processing method as claimed in claim 3; Wherein, Capable piece number according to memory size divided is through formula:
obtains; Wherein b by branch's piece number, p is each pixel shared byte number in internal memory, m; N is the size of image, and a is the free memory of system when handling image.
5. remote sensing image immediate processing method as claimed in claim 4; Wherein, The capable piece number of being divided according to GPU video memory size is according to formula:
obtains; Wherein B by branch's piece number, P is each pixel shared byte number in video memory, M; N is the size of image, and A is the available video memory of system when handling image.
6. remote sensing image immediate processing method as claimed in claim 1; Wherein, Step 3) is specially: at first; Remote sensing images through after the above-mentioned processing reconfigure to behind the view picture remote sensing image, remote sensing images are divided into the some row pieces that can be handled by calculator memory, and each row that will cut apart read in successively in the calculator memory soon; Secondly, be divided into the plurality of sub row piece that to be handled by GPU with reading in each row piece in the calculator memory, and each the sub-row piece that will cut apart reads in successively in the GPU video memory; Once more, utilize GPU that each the sub-row piece that reads in the GPU video memory is carried out one dimension Fourier rank transformation, deposit the result data behind the one dimension Fourier rank transformation in disk.
7. remote sensing image immediate processing method as claimed in claim 6; Wherein, Row piece number according to memory size divided is according to formula:
obtains; Wherein c is institute's apportion piece number, and p is each pixel shared byte number in internal memory, m; N is the size of image, and a is the free memory of system when handling image.
8. remote sensing image immediate processing method as claimed in claim 7; Wherein, The row piece number of being divided according to GPU video memory size is according to formula:
obtains; Wherein C is institute's apportion piece number, and P is each pixel shared byte number in video memory, M; N is the size of image, and A is the available video memory of system when handling image.
9. like claim 3 or 6 described remote sensing image immediate processing methods, wherein, the result data after the one-dimensional Fourier transform is being deposited in disk simultaneously, note the address of each the piece remote sensing image after the processing.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210119088XA CN102663759A (en) | 2012-04-20 | 2012-04-20 | Remote-sensing image rapid-processing method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210119088XA CN102663759A (en) | 2012-04-20 | 2012-04-20 | Remote-sensing image rapid-processing method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN102663759A true CN102663759A (en) | 2012-09-12 |
Family
ID=46773237
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201210119088XA Pending CN102663759A (en) | 2012-04-20 | 2012-04-20 | Remote-sensing image rapid-processing method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN102663759A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110390679A (en) * | 2019-07-03 | 2019-10-29 | 上海联影智能医疗科技有限公司 | Image processing method, computer equipment and readable storage medium storing program for executing |
CN110796652A (en) * | 2019-10-30 | 2020-02-14 | 上海联影智能医疗科技有限公司 | Image processing method, computer device, and storage medium |
CN110941399A (en) * | 2019-12-05 | 2020-03-31 | 北京金山云网络技术有限公司 | Data processing method and device and electronic equipment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100088356A1 (en) * | 2008-10-03 | 2010-04-08 | Microsoft Corporation | Fast computation of general fourier transforms on graphics processing units |
CN101783021A (en) * | 2010-02-02 | 2010-07-21 | 深圳市安健科技有限公司 | Method for speeding up DR image processing by using operation of GPU |
-
2012
- 2012-04-20 CN CN201210119088XA patent/CN102663759A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100088356A1 (en) * | 2008-10-03 | 2010-04-08 | Microsoft Corporation | Fast computation of general fourier transforms on graphics processing units |
CN101783021A (en) * | 2010-02-02 | 2010-07-21 | 深圳市安健科技有限公司 | Method for speeding up DR image processing by using operation of GPU |
Non-Patent Citations (2)
Title |
---|
DANIEL KAUKER ET AL.: "《Memory Saving Discrete Fourier Transform on GPUs》", 《2010 10TH IEEE INTERNATIONAL CONFERENCE ON COMPUTER AND INFORMATION TECHNOLOGY》 * |
邓劲: "《图形处理器上的快速傅里叶变换》", 《现代电子技术》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110390679A (en) * | 2019-07-03 | 2019-10-29 | 上海联影智能医疗科技有限公司 | Image processing method, computer equipment and readable storage medium storing program for executing |
CN110390679B (en) * | 2019-07-03 | 2022-04-26 | 上海联影智能医疗科技有限公司 | Image processing method, computer device, and readable storage medium |
CN110796652A (en) * | 2019-10-30 | 2020-02-14 | 上海联影智能医疗科技有限公司 | Image processing method, computer device, and storage medium |
CN110941399A (en) * | 2019-12-05 | 2020-03-31 | 北京金山云网络技术有限公司 | Data processing method and device and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111199273B (en) | Convolution calculation method, device, equipment and storage medium | |
CN109461119B (en) | Image filling method and device in convolutional neural networks FPGA acceleration | |
US20130027416A1 (en) | Gather method and apparatus for media processing accelerators | |
US11797855B2 (en) | System and method of accelerating execution of a neural network | |
CN103198451B (en) | A kind of GPU realizes the method for fast wavelet transform by piecemeal | |
CN108920540B (en) | Spark-based parallel raster data processing method | |
CN112668708B (en) | Convolution operation device for improving data utilization rate | |
CN110390382B (en) | Convolutional neural network hardware accelerator with novel feature map caching module | |
CN105739951A (en) | GPU-based L1 minimization problem fast solving method | |
CN102222316A (en) | Double-buffer ping-bang parallel-structure image processing optimization method based on DMA (direct memory access) | |
CN105488753B (en) | A kind of pair of image carries out the method and device of two-dimension fourier transform or inverse transformation | |
US20180204313A1 (en) | 2d discrete fourier transform with simultaneous edge artifact removal for real-time applications | |
Asaduzzaman et al. | A time-efficient image processing algorithm for multicore/manycore parallel computing | |
CN102663759A (en) | Remote-sensing image rapid-processing method | |
CN104239231B (en) | A kind of method and device for accelerating L2 cache preheating | |
CN109447239B (en) | Embedded convolutional neural network acceleration method based on ARM | |
CN109615067A (en) | A kind of data dispatching method and device of convolutional neural networks | |
CN112348182A (en) | Neural network maxout layer computing device | |
CN113222129A (en) | Convolution operation processing unit and system based on multi-level cache cyclic utilization | |
CN112446005A (en) | Computational optimization | |
CN104954749B (en) | A kind of information recording method | |
CN109416743B (en) | Three-dimensional convolution device for identifying human actions | |
US20220188380A1 (en) | Data processing method and apparatus applied to graphics processing unit, and electronic device | |
CN108920097A (en) | A kind of three-dimensional data processing method based on Laden Balance | |
CN114925780A (en) | Optimization and acceleration method of lightweight CNN classifier based on FPGA |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C12 | Rejection of a patent application after its publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20120912 |