CN107301398B - A kind of identification method of image target of synthetic aperture radar realized based on GPU - Google Patents
A kind of identification method of image target of synthetic aperture radar realized based on GPU Download PDFInfo
- Publication number
- CN107301398B CN107301398B CN201710485297.9A CN201710485297A CN107301398B CN 107301398 B CN107301398 B CN 107301398B CN 201710485297 A CN201710485297 A CN 201710485297A CN 107301398 B CN107301398 B CN 107301398B
- Authority
- CN
- China
- Prior art keywords
- matrix
- gpu
- value
- sample
- parallel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
A kind of radar image target identification technology of the present invention field, and in particular to identification method of image target of synthetic aperture radar realized based on GPU.With the development of SAR imaging technique, the resolution ratio and data volume of SAR image increase sharply, so traditional PCA method efficiency based on CPU serial computing is too low and calculates cost prohibitive.The present invention utilizes the efficient computing capability of GPU general-purpose computations, carries out parallel parsing to PCA feature extracting method, and ask most the methods of value to carry out GPU the strong matrix multiplication of wherein concurrency, Jacobi feature decomposition and reduction and improve parallel.
Description
Technical field
The invention belongs to radar image target identification technology fields, and in particular to a kind of synthetic aperture realized based on GPU
Radar image target identification method.
Background technique
The identification of synthetic aperture radar (Synthetic Aperture Radar, hereinafter referred to as SAR) Image Automatic Target
(Automatic Target Recognition, ATR) technology refers in the case where no artificial assistance, right from large scene
Target is detected and is positioned and realized the judgement of situations such as model to target, attribute and equipment.SAR target identification institute at present
The method of use has very much, has more method in feature extraction and the design of classifier and has respective excellent scarce
Point.However compared to these plentiful and substantial theoretical results, the technology of SAR target identification is realized and real application research is but more slow
Slowly, practical SAR Target detection and identification system is also less or immature.The big reason for influencing this process is exactly SAR
The raising of image resolution ratio causes the data volume obtained to sharply increase, so that traditional CPU processing mode is difficult to handle in real time
Data.The existing research for SAR target identification method is directed to the nicety of grading for how improving target mostly, and rare to SAR
The research of Target Recognition Algorithms calculating speed.Therefore, it how under the premise of guaranteeing recognizer accuracy rate, improves algorithm and executes
Speed realizes the real-time of Target Recognition Algorithms, is the challenge that researcher faces.
In recent years, the high speed development of graphics processor (GPU, Graphic Process Unit), framework it is continuous complete
It is kind, so that it develops into the calculating instrument of the Floating-point Computation ability with superelevation from the initial rendering for being only applicable to figure, be
The slow problem of Target Recognition Algorithms speed provides powerful.The concurrency of GPU is determined by threads a large amount of in its piece, this
It is bigger that kind concurrency has GPU relative to traditional CPU serial arithmetic when calculating the data that data volume is big, closeness is high
Advantage.In addition, NVIDIA company continues to optimize the index of GPU various aspects, guarantee computing capability in advance under, reduction function as far as possible
Consumption is to reduce calculating cost.
Summary of the invention
The present invention provides a kind of identification method of image target of synthetic aperture radar realized based on GPU, can effectively solve
The certainly problem of Target Recognition Algorithms real-time difference.Principal Component Analysis (Principle Component Analysis, referred to as
It PCA is) a kind of more commonly used SAR image feature extracting method, it corresponds to the thought of high information based on big variance, by original number
According to projection to new coordinate space, with less several generalized variables reflected sample raw information as much as possible.Its realization
Steps are as follows:
Each sample process is constituted a matrix P for original n training sample at one-dimensional data by the first stepm*n
={ X1,X2,X3,…,Xn, and calculate equalization matrix P.
Second step calculates the covariance matrix for finding out the matrix P after mean value:
Third step, Jacobi iterative method carry out Eigenvalues Decomposition to covariance matrix Q:
Q=UTΛU
The covariance matrix for being k for an order, wherein covariance matrix has k nonzero eigenvalue, and U is orthogonal matrix, often
One column are also that individual features are worth corresponding feature vector.At this moment the preceding d maximum eigenvalue institute of sorted characteristic value is usually taken
Corresponding vector, as PCA projector space, to reduce the dimension of feature.Select big characteristic value because characteristic value it is bigger that
The feature mapped to its corresponding feature vector direction is more obvious.
Sample data is projected to the Projection Character obtained in step 3 spatially, so that it may obtain principal component by the 4th step
Feature:
YPCA=TP
PCA algorithm implementation flow chart is provided according to above four step, as shown in Fig. 1.
Next parallel parsing is carried out to each realization process of PCA algorithm, to finally realize that the GPU of algorithm is counted parallel
It calculates, reaches speed-increasing effect.
The first step averages to input sample matrix by rows, and then mean value is individually subtracted in each column vector in original matrix P
Vector.During calculating every a line mean value, parallel computation can be carried out with matrix behavior basic unit, per thread is realized
The calculating of a line.
Second step, the calculating of covariance matrix, i.e. matrix multiple.Each element of covariance matrix Q is by matrix P
Certain a line and a certain column count inner product obtain, and the calculating process of each element is relatively independent, are that a kind of typically suitable GPU is parallel
The operation of realization.
Since matrix multiplication is there are good concurrency, it can realize that each thread is corresponded in calculating matrix Q by CUDA
One value of corresponding position, allows each thread to read the i-th row and P of PTJth column, thus the position calculating matrix Q be (i, j)
The value of element.Calculating complete matrix Q needs kernel function access global storage to read the data for participating in calculating, wherein reading
Take P matrix PT.width secondary, PTMatrix P.height times.In order to allow the thread quantity of each block to reach the whole of warp size
Several times, the size that we define block is 16*16, so the size of grid is (n/16, n/16).
Third step, the feature decomposition of covariance matrix Q.By decomposing Jacobi iterative algorithm calculating process, it can be found that
There are the concurrencys between data for it, and be related to largely can Parallel Implementation matrix operation.
The detailed process of Jacobi iterative method are as follows: set n rank matrix A as the real symmetric matrix to feature decomposition, change each time
Dai Zhongjun chooses the element in matrix A on an off-diagonal of maximum absolute value, is set as apq, the coordinate position on A be (p,
q).Specific rotation transformation formula is as follows:
A0=A
Wherein spin matrix JKIt is as follows:
The method is also named bilateral rotary process, as described above, each round JKAfter bilateral transformation, the off-diagonal position of matrix A
Element quadratic sum is reducedElement quadratic sum increases on diagonal lineBy several wheel iteration, original matrix A is become
Diagonal matrix, the element on diagonal line is matrix exgenvalue.
Above-mentioned spin matrix JKIn play turning effort be then 2 × 2 submatrix R, R form it is as follows:
Need to find a kind of suitable dispatching algorithm now, 2 × 2 submatrixs all to matrix are carried out and only once revolved
Turn, such process is known as a wheel iteration.In every wheel iteration, the bilateral rotation that n/2 group is independent of each other can be at most carried out simultaneously.
As it can be seen that Jacobi Rotation Algorithm has good data parallelism, number can be realized by the common single cycle of competitive sports
According to rational management.
In addition, in Jacobi iterative algorithm during GPU Parallel Implementation, relate to what a large amount of matrix multiplications and asking most were worth
It calculates.It wherein the parallel parsing of matrix multiplication and realizes by the agency of above, next introduces parallel reduction method and seek matrix most
The GPU of value realizes process.
The method of CPU serial computing matrix or array maximum value is calculated by all elements in traversal array, this
The time complexity of sample processing is N.And here, we realize the calculating of maximum value using the method for reduction on GPU, will be very big
Raising program operational efficiency.The maximum value of adjacent 2 elements is found out first, is then asked on the basis of last calculated result again
The maximum value of 2 adjacent maximas out, that is, be equivalent to the maximum value for having found out 4 adjacent elements, according to above-mentioned rule according to
The secondary element maximum value for finding out adjacent 8,16,32 etc..
As can be seen that this method based on GPU reduction maximizing only needs the time of logN that can both complete N on CPU
The calculating of time is greatly saved and calculates the time, and N is bigger, and acceleration effect will be better.The tool of CUDA parallel reduction is given below
Body realizes step.
Assuming that we need to calculate the maximum value in the array that one possesses 1526 elements, per thread alignment processing one
A data, then it is 6 that block quantity, which is arranged, and the number of threads of each block is 256, and grid and block are one-dimensional.It adopts
Optimized with shared memory and accessed, firstly, per thread reads a data, to complete to deposit from global memory to shared
The data copy of reservoir.Then simultaneously operating is done by synchronous function _ _ syncthreads (), to guarantee copy function whole
Complete, the data in such same block just can safety by other thread accesses.It connects down and is realized by for () circulation
The half thread that reduction procedure, i.e. each round circulation only use last circulation ask the maximum of tid and tid+s thread
Value, is as a result stored in the tid thread, i.e. reduction procedure.S can be understood as span herein, for each thread block
For block, have:
S=1, only tid=0 when recycling for the first time, the calculating that 2,4,6,8 ..., 254 threads compare in execution, i.e. s_
Compared with the element that data [tid] and span thereafter are 1 and maximizing, as a result it is stored in s_data [tid].
S=2, only tid=0 when second of circulation, the calculating that 4,8,16 ..., 252 threads compare in execution, i.e. s_
Compared with the element that data [tid] and span thereafter are 2 and maximizing, as a result it is stored in s_data [tid].
And so on.
When last time s=128, only thread tid=0 really performs calculating, i.e. s_data [tid] and span thereafter
Element for 128 compares and maximizing, is as a result stored in s_data [0].The numerical value in s_data [0] is exactly this at this time
The maximum value of 256 data in block.Attached drawing 2 has given in parallel reduction algorithm thread dispatch situation in single thread block.
4th step, sample data project on feature space, can GPU Parallel Implementation by matrix multiplication.
The invention has the benefit that the present invention utilizes the efficient computing capability of GPU general-purpose computations, to PCA feature extraction
Method carries out parallel parsing, and the strong matrix multiplication of wherein concurrency, Jacobi feature decomposition and reduction are sought most the methods of value
GPU is carried out to improve parallel.
Detailed description of the invention
Fig. 1 is PCA algorithm flow chart;
Fig. 2 is parallel reduction thread scheduling figure;
Fig. 3 is PCA feature extraction speed-up ratio with the increased situation of change of sample size.
Specific embodiment
With reference to the accompanying drawing, the technical schemes of the invention are described in detail
In order to verify effective acceleration of the GPU to SAR Target Recognition Algorithms, using in uniform hardware environment and development platform
Operation time as index, compares, to guarantee that variable is single.GPU model NVIDIA GeForce GTX 750Ti,
Possess 512 CUDA processor cores, 1G video memory.CPU model Intel Core i7-4790,3.60GHz, 16.0G memory.
Experimental implementation system is 64 Windows7.CUDA version is CUDA7.5.Programmed environment is Visual Studio2010, programming
Language is CUDA C language.
In addition, experimental data uses MSTAR image data, MSTAR is simply introduced now.
MSTAR (Moving and Stationary Target Acquisition Recognition) project initiation in
1994, it was a SAR ATR project by the multiple research institution's joint studyes in the U.S..Wherein, the laboratory U.S. Sandia
It is responsible for providing the original SAR data of 0.3~1m of X-band resolution ratio.The U.S. is responsible in the laboratory Wright establishing to be studied for model
Various landform back scattering directional diagrams and for sort research 18 kinds of surface cars obtain database, can to each car
The sample of 72 different perspectivess and different directions is provided.And the responsible offer special analysis such as laboratory MIT Lincoln, extract and
Sorting algorithm.Present MSTAR data have become the standard database of examination SAR target identification and sorting algorithm.Major part exists
The SAR target identification and sorting algorithm delivered on authoritative magazine and meeting are tested and are assessed using MSTAR data.
The present invention is the acceleration situation in order to improve SAR image target identification parallel by GPU.PCA feature extraction algorithm
Input be all data set P and data dimension d, and data dimension is fixed as d=64*64.This experiment will gradually change sample
This number increases to 6400 from 400, not only can parser in the acceleration situation at the end CPU and the end GPU, equally can analyze
Algorithm is when the end GPU is calculated with the increased time change situation of sample number.In order to which expanding data library is increased with reaching in experiment
The requirement of sample number, by the artificial fog-level to photo increase by 100% and 200%.
Table 1 gives PCA feature extraction experimental result, and it is real parallel by GPU to give PCA algorithm by calculating speed-up ratio
Existing acceleration effect.
1 PCA feature extraction experimental result of table
Attached drawing 3, which gives PCA feature extraction speed-up ratio line chart, to be caused due to the limitation of GPU computing resource with sample
Several increases, the increased trend of speed-up ratio slow down.
The identification method of image target of synthetic aperture radar realized based on GPU of the invention can effectively utilize GPU's
Powerful calculating ability, and the feature that combination algorithm computation complexity is high, greatly improve algorithm execution speed.
Claims (1)
1. a kind of identification method of image target of synthetic aperture radar realized based on GPU, this method are mentioned using Principal Component Analysis
SAR image feature is taken, sample is obtained, which is characterized in that uses the algorithm based on GPU, improves the following recognition methods to sample
Processing speed:
S1, each sample process of extraction is formed into a matrix P for original n training sample at one-dimensional datam*n=
{X1,X2,X3,…,Xn, equalization matrix P is obtained, is averaged to input sample matrix by rows, it is then each in original matrix P
Mean vector is individually subtracted in column vector;During calculating every a line mean value, using the parallel algorithm based on GPU, with matrix
The basic thread unit of behavior carries out parallel computation;
S2, calculating find out the covariance matrix of the matrix P after mean value:
Each element of covariance matrix Q is obtained by certain a line and a certain column count inner product of matrix P, the meter of each element
Calculation process is mutually indepedent, and using the parallel computation of GPU, per thread corresponds to the value of a position in calculating matrix Q;
S3, Eigenvalues Decomposition is carried out to covariance matrix Q using Jacobi iterative method:
Q=UTΛUΛ
The covariance matrix for being k for an order, wherein covariance matrix has k nonzero eigenvalue, and U is orthogonal matrix, Mei Yilie
It is also that individual features are worth corresponding feature vector, vector corresponding to the preceding d maximum eigenvalue of sorted characteristic value is taken to make
For PCA projector space;In Jacobi iterative process, each round rotation transformation can only change one 2 × 2 in original matrix square
Battle array;It is primary to carry out the rotation transformation that n/2 group is independent of each other simultaneously using the parallel algorithm based on GPU, i.e., once realize original square
The transformation of all elements, the n/2 group rotation transformation are independent of each other in battle array, specifically include matrix multiplication and most value is asked to transport
It calculates, wherein ask most value operation that can realize GPU parallel computation by reducing method, it will be several wait ask the data being most worth to read respectively
In a thread block, this is parallel for thread block grade, Thread-Level Parallelism is then realized in per thread block, finally by each thread block institute
Most value is asked to make once relatively, most value finally can be obtained;
S4, sample data is projected on the Projection Character space T obtained in step S3, obtains the feature of principal component:
YPCA=TP.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710485297.9A CN107301398B (en) | 2017-06-23 | 2017-06-23 | A kind of identification method of image target of synthetic aperture radar realized based on GPU |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710485297.9A CN107301398B (en) | 2017-06-23 | 2017-06-23 | A kind of identification method of image target of synthetic aperture radar realized based on GPU |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107301398A CN107301398A (en) | 2017-10-27 |
CN107301398B true CN107301398B (en) | 2019-04-30 |
Family
ID=60136330
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710485297.9A Active CN107301398B (en) | 2017-06-23 | 2017-06-23 | A kind of identification method of image target of synthetic aperture radar realized based on GPU |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107301398B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109444886B (en) * | 2018-12-30 | 2020-06-26 | 成都汇蓉国科微系统技术有限公司 | Parallel computing processing method, system and medium for bistatic SAR real-time imaging |
CN111289975B (en) * | 2020-01-21 | 2022-04-22 | 博微太赫兹信息科技有限公司 | Rapid imaging processing system for multi-GPU parallel computing |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101937082A (en) * | 2009-07-02 | 2011-01-05 | 北京理工大学 | GPU (Graphic Processing Unit) many-core platform based parallel imaging method of synthetic aperture radar |
WO2016101279A1 (en) * | 2014-12-26 | 2016-06-30 | 中国海洋大学 | Quick detecting method for synthetic aperture radar image of ship target |
CN106338730A (en) * | 2016-08-11 | 2017-01-18 | 电子科技大学 | Asynchronous parallel synthetic aperture radar backward projection imaging method |
CN106597439A (en) * | 2016-12-12 | 2017-04-26 | 电子科技大学 | Synthetic aperture radar target identification method based on incremental learning |
-
2017
- 2017-06-23 CN CN201710485297.9A patent/CN107301398B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101937082A (en) * | 2009-07-02 | 2011-01-05 | 北京理工大学 | GPU (Graphic Processing Unit) many-core platform based parallel imaging method of synthetic aperture radar |
WO2016101279A1 (en) * | 2014-12-26 | 2016-06-30 | 中国海洋大学 | Quick detecting method for synthetic aperture radar image of ship target |
CN106170819A (en) * | 2014-12-26 | 2016-11-30 | 中国海洋大学 | A kind of diameter radar image Ship Target method for quick |
CN106338730A (en) * | 2016-08-11 | 2017-01-18 | 电子科技大学 | Asynchronous parallel synthetic aperture radar backward projection imaging method |
CN106597439A (en) * | 2016-12-12 | 2017-04-26 | 电子科技大学 | Synthetic aperture radar target identification method based on incremental learning |
Non-Patent Citations (4)
Title |
---|
一种合成孔径雷达图像特征提取与目标识别的新方法;宦若虹;《电子与信息学报》;20080520;全文 * |
合成孔径雷达图像目标识别问题研究;杨文;《航天返回与遥感》;20060904;全文 * |
合成孔径雷达目标识别理论与关键技术研究;崔宗勇;《中国博士学位论文全文数据库 信息科技辑》;20160315;全文 * |
基于PCA特征的快速SAR图像目标识别方法;王世晞;《国防科技大学学报》;20080615;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN107301398A (en) | 2017-10-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Bogatskiy et al. | Lorentz group equivariant neural network for particle physics | |
US8860715B2 (en) | Method and system for evaluation using probabilistic boosting trees | |
Cheng et al. | Minimal kinematic constraints and mT2 | |
Yu et al. | A global optimization approach to robust multi-model fitting | |
Bongsoo Choy et al. | Enriching object detection with 2d-3d registration and continuous viewpoint estimation | |
CN106096506B (en) | Based on the SAR target identification method for differentiating doubledictionary between subclass class | |
EP3742343A1 (en) | Autonomous vehicle simulation using machine learning | |
CN108957450A (en) | A kind of millimetre-wave radar GPU real time three-dimensional imaging method | |
Halyo et al. | GPU Enhancement of the Trigger to Extend Physics Reach at the LHC | |
Halyo et al. | First evaluation of the CPU, GPGPU and MIC architectures for real time particle tracking based on Hough transform at the LHC | |
CN102073982A (en) | Method for realizing acceleration of anisotropic diffusion filtration of overlarge synthetic aperture radar (SAR) image by graphic processing unit (GPU) | |
CN107301398B (en) | A kind of identification method of image target of synthetic aperture radar realized based on GPU | |
Llerena et al. | Gaussian bounding boxes and probabilistic intersection-over-union for object detection | |
Bornhauser et al. | Determination of the CMSSM parameters using neural networks | |
Lantz et al. | Speeding up particle track reconstruction using a parallel Kalman filter algorithm | |
Jiménez et al. | Fast box-counting algorithm on GPU | |
Rashwan et al. | Matrixnets: A new scale and aspect ratio aware architecture for object detection | |
Cholachgudda et al. | Performance Analysis of Deep Neural Networks for Tomato Leaf Disease Classification with Server-Based Computing | |
Ramalho et al. | Efficient independent component analysis on a GPU | |
CN104123119A (en) | Dynamic vision measurement feature point center quick positioning method based on GPU | |
Wang et al. | Three-dimensional nonlinear force-free field reconstruction of solar active region 11158 by direct boundary integral equation | |
Cárdenas-Montes et al. | Concurrent CPU-GPU code optimization: The two-point angular correlation function as case study | |
Wenzel et al. | Accelerating navigation in the VecGeom geometry modeller | |
Kofler et al. | Kd-tree based n-body simulations with volume-mass heuristic on the GPU | |
Araújo et al. | Comparing the use of sum and difference histograms and gray levels occurrence matrix for texture descriptors |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |