CN109801226A - Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm - Google Patents
Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm Download PDFInfo
- Publication number
- CN109801226A CN109801226A CN201811528890.8A CN201811528890A CN109801226A CN 109801226 A CN109801226 A CN 109801226A CN 201811528890 A CN201811528890 A CN 201811528890A CN 109801226 A CN109801226 A CN 109801226A
- Authority
- CN
- China
- Prior art keywords
- target
- algorithm
- image
- seawater
- sam
- 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 Analysis (AREA)
Abstract
The invention proposes a kind of based on SAM and weights the waterborne target detection method of auto-correlation CEM algorithm, solves the problems, such as that weighting auto-correlation CEM algorithm is bad for big Target scalar extraction effect.Method includes the following steps: firstly, carrying out SAM classification to high spectrum image progress dimensionality reduction and Endmember extraction, and using the end member of extraction to determine two macrotaxonomies: naval vessel class and seawater class；Secondly, subtracting naval vessel class pixel from all pixels as backdrop pels, the SAM-CEM algorithm by weighting autocorrelation matrix based on pure backdrop pels calculates detection result；Then, gray level image only comprising seawater and naval vessel is obtained by classification image, and carries out binaryzation and morphology processing, and the searching maximum white area of range is seawater region；Finally, by carrying out binaryzation to target acquisition image, the characteristics of using Ship Target in the seawater, removal not false-alarm targets in seawater region, so that it is determined that final Ship Target.
Description
Technical field
The present invention relates to a kind of waterborne target detection methods based on weighting auto-correlation CEM algorithm.
Background technique
Sea naval vessels target detection technique suffers from extremely wide application prospect in civil and military field.In civilian side
Face, naval vessel detection can be used for monitoring and management, the marine pollution control survey in bay and harbour naval vessel etc.[1]；In military neck
Domain, can be used for monitoring illegal ship such as steal into another country, pirate, the attack of terrorism, and to wartime surface monitoring, fight naval vessel detect
It examines[2].The research object of traditional algorithm of target detection is mainly with infrared image and synthetic aperture radar[3](Synthetic
ApertureRadar, SAR) based on image.With past hyperspectral remote sensing system and hyperspectral data processing technology over 20 years
It grows rapidly, the abundant spectral information that high-spectrum remote sensing is contained with it has big advantage in terms of target acquisition.
High-spectrum remote-sensing target acquisition algorithm can be divided into according to algorithm input: known target, known background, it is known that target, Unknown Background,
Unknown object, Unknown Background, unknown object, known background.
The case where for just knowing that Target scalar spectrum and Unknown Background, target spy mainly is carried out by three approach
Survey: first is simple match algorithm, such as minimum distance method, Spectral angle mapper scheduling algorithm；Second be using sample correlation matrix (or
Person's covariance matrix) property carry out target acquisition, as bound energy minimum method algorithm (CEM), adaptive cosine consistency are commented
Estimate device algorithm (ACE) etc.[4].Third is to carry out Endmember extraction using Decomposition of Mixed Pixels technology, obtains background information, is converted into
The case where known target, known background.
Bound energy min algorithm (constrained energyed minimization, CEM) is to just know that sense is emerging
The spectrum of interesting target and background it is unknown under conditions of algorithm that target is detected.Use the naval vessel end member of acquisition as mesh
Mark, then problem is converted into known target, the hyper-spectral target detection problem of Unknown Background, utilizes the property of sample correlation matrix
Carry out target acquisition.CEM algorithm is the thought using linearly constrained minimum variance beamformer, can extract specific direction
Signal and the interference in other directions of decaying.This method is well suited for the case where specific ingredient accounts for image variance ratio very little, can dash forward
Out certain terrestrial object information (target) and suppress other terrestrial object information (background), to reach the effect for separating certain atural object from image
Fruit, i.e. high spectrum image target acquisition.The algorithm is specific as follows:
In high spectrum imageIn, target d, what wherein n indicated every one dimensional image include pixel number, i.e. image
The product of height and width, the purpose of CEM are exactly to design a L dimensional linear filter[4]W=(ω1,ω1,...,ωL)T, so that such as
Filtering output energy is minimum under the conditions of lower:
wTD=1 (1)
When input is xiWhen, detect statistic yiFor by the output of filtering algorithm
Then, all observation samples are by the average output energy of filtering algorithm w
In formula,For the autocorrelation matrix of sample set.The design of filtering algorithm w can sum up in this way
For following minimum problems:
For constrained extremal problem, formula (4) are solved with Langrange multiplier method, then CEM operator (Harsanyi, 1993)
It is expressed as
CEM operator is acted on into each pixel in image, the distribution situation of target d in the picture, realization pair will be obtained
The detection of target d, final target acquisition function are as follows:
Due to only knowing target in CEM algorithm, and background is known nothing, the R in formula (4) is all pixels from phase
Matrix is closed, since R plays the role of inhibition for background and target, only target is generally all very during target acquisition
Small, so influence of the target for background energy can generally ignore, this is exactly CEM operator in Small object extraction timeliness
Fruit is significant, and the reason of for big Objective extraction poor effect.
In order to allow area-of-interest as background participate in operation when it is small to the contribution of R, and with area-of-interest goal discrepancy
Different big pixel is big to the contribution of R, and Geng Xiurui proposes the CEM algorithm (weighted based on weighting autocorrelation matrix
Correlation matrix CEM, WCM-CEM) [8], when calculating object vector and each sample inner product, increase by one adds
Weight function g (x, y) is used to measuring vector x, and y similarity, with vector x, the similarity of y increases and reduces the value of g (x, y).Specifically
Algorithm characteristic is as follows:
In high spectrum imageIn, what wherein n indicated every one dimensional image include pixel number, target collection is divided intoTo need the L dimensional linear filter solved[4].It enables
In formula (7), P indicates the number of object vector, the pixel number for including that n indicates every one dimensional image in formula (8).
So filter vector w inhibits background signal as far as possible, will be attributed to following pole while enhancing signal of interest d
Value problem:
In formula, T is the sample autocorrelation matrix of all echo signals；E=wTTw is that echo signal is put down after wave filter
Export energy；R is the autocorrelation matrix of all samples；F=wTRw is that all sample means export energy.Due to just knowing that sense
Signal-of-interest, and background is known nothing, so using the average energy of all signals as background energy here.Then, formula (9)
It is equivalent to eigenvalue problem[9-10]
R-1Tw=λ w (10)
When only including a target d in target collection D, the autocorrelation matrix of echo signal is T=ddt, at this point, formula
(10) switch to
R-1ddtW=λ w (11)
It is available that CEM operator in formula (5) is substituted into (11) formula
The meaning of formula (12) is: CEM operatorFeature corresponding to unique features value with formula (11)
A vector only poor proportionality coefficient, i.e. CEM algorithm is special case of the formula (10) in only one target.
It is based on weighting auto-correlation target acquisition operator
In formula, R*For the improved weighting autocorrelation matrix about d, it is expressed as
In formula, f (0)=0, f (x) is monotonically increasing function (x >=0)；C is a constant；G (x, y) is measuring vector x, y phase
Like the function of degree, the value of g (x, y) with vector x, the increase of y similarity and reduce.
Although influence does not have in this way, reducing influence of the area-of-interest as background to R by auto-correlation function
It completely eliminates.
Summary of the invention
High spectrum image sea naval vessels target is effectively detected in order to realize, the invention proposes one kind to be based on SAM
With the waterborne target detection method of weighting auto-correlation CEM algorithm, solves weighting auto-correlation CEM algorithm and big Target scalar is mentioned
Take ineffective problem.
Technical scheme is as follows:
Firstly, carrying out dimensionality reduction and Endmember extraction to high spectrum image, and SAM classification is carried out to determine using the end member of extraction
Two macrotaxonomies: naval vessel class and seawater class；
Secondly, subtracting naval vessel class pixel from all pixels as backdrop pels, by weighting certainly based on pure backdrop pels
The SAM-CEM algorithm of correlation matrix calculates detection result；
Then, gray level image only comprising seawater and naval vessel is obtained by classification image, and carries out binaryzation and mathematics
Morphological scale-space, the searching maximum white area of range are seawater region；
Finally, the characteristics of using Ship Target in the seawater, removal does not exist by carrying out binaryzation to target acquisition image
False-alarm targets in seawater region, so that it is determined that final Ship Target.
The invention has the following advantages:
Dimensionality reduction, Endmember extraction are first passed through, converts known target, the target acquisition problem of Unknown Background for problem；For
The problem of echo signal participates in operation as background signal in auto-correlation CEM algorithm is weighted, improved SAM-CEM algorithm is proposed
Carry out target acquisition algorithm.Specifically: first being classified using the end member of extraction to high spectrum image, subtracted from all pixels
Naval vessel class pixel is as backdrop pels, the method for carrying out target acquisition by weighting autocorrelation matrix based on pure backdrop pels,
It finally accurately detected the parameters such as the number of sea naval vessels in high spectrum image, position and naval vessel area.Experimental result table
Bright, compared to weighting auto-correlation CEM algorithm, innovatory algorithm is embodied in: 1) false alarm rate becomes smaller, and separating effect is more preferably；2) target mean
Become larger, echo signal is enhanced, and the mean value of background signal becomes smaller, and background is inhibited；3) between target energy/background energy
Ratio be significantly improved.
In short, the present invention has preferably shown echo signal especially and inhibited background signal, avoids weighting auto-correlation CEM and calculate
Echo signal participates in operation to the influence of detection result as background signal in method, efficiently solves original algorithm to big target
The ineffective problem of Objects extraction, also achieves satisfactory result in the target acquisition to sea naval vessels.
Detailed description of the invention
Fig. 1 is waterborne target detection method flow chart of the invention.
Fig. 2 is the example of MNF transformation, wherein the 1st ingredient of (a) MNF, (b) the 21st ingredient of MNF, (c) MNF transform characteristics
Value.
Fig. 3 is the example of PPI changing image.
Fig. 4 is the example for selecting end member by N-dimensional view tool.
Fig. 5 is the example of area-of-interest.
Fig. 6 is the example of the end member spectral profile extracted.
Fig. 7 is using the example for extracting endmember spectra angle classification results.
Fig. 8 is the detection result for directly using CEM algorithm.
Fig. 9 is the detection result for weighting autocorrelative CEM algorithm.
Figure 10 is the detection result after innovatory algorithm of the present invention.
Figure 11 is seawater background extracting result.
Figure 12 is the detection result gray level image after innovatory algorithm of the present invention.
Figure 13 is the binarization result after innovatory algorithm of the present invention.
Figure 14 is the Ship Target and coordinate finally extracted.
Figure 15 is the ROC curve that auto-correlation weights CEM algorithm and the improved SAM-CEM algorithm of the present invention, wherein (a)
Auto-correlation weights CEM algorithm, (b) the improved SAM-CEM algorithm of the present invention.
Specific embodiment
Below in conjunction with drawings and examples, the invention will be further described.
The present invention is first to MNF dimensionality reduction[5-6]Image afterwards carries out Endmember extraction using PPI method[7], determining naval vessel end member
Afterwards, it is converted into known target, the target acquisition problem of Unknown Background；The end member extracted is recycled to divide high spectrum image
Class；Target acquisition is finally carried out by improved SAM-CEM algorithm using the naval vessel end member extracted.The present invention is known to synthesis
On the basis of three kinds of target, Unknown Background detection approach, by the improvement to weighting auto-correlation CEM algorithm, weighting is avoided certainly
Using echo signal as background signal participation operation and to the influence of detection result in related CEM algorithm [8][4], and solve original
There is the problem that algorithm is bad to big Target scalar extraction effect, is achieved in the target acquisition to sea naval vessels satisfactory
As a result.
Algorithm improvement is summarized as follows:
In formula (9), since background signal energy F is that all pixels in image pass through the average output after filter
Energy, also including interesting target signal energy, that is to say, that based on signal d both conducts in weighting auto-correlation CEM algorithm
Echo signal participates in operation, while also participating in operation as a part of background signal.When echo signal be distributed in the picture it is dilute
When few, it can be ignored the image of F, CEM operator significant effect in terms of Small object extraction, and work as interesting target
When distributed more widely general in the picture, its influence to F cannot be had ignored, so that CEM operator is in big Target scalar extraction effect
It is bad.
In order to make influence of the echo signal d to R small as far as possible, first image is pre-processed, steps are as follows:
1) first spectrum picture is carried out through MNF dimension-reduction treatment
2) spectrum picture is converted using pure pixel index (PPI), obtains naval vessel, seawater, day using N-dimensional view tool
The end member of the atural objects such as sky, island, building.
3) Spectral angle mapper classification is carried out using SAM using the end member obtained, such Ship Target is substantially classified from background
Out, but other targets in addition to crossing naval vessel will be divided into target context, and target classification acquisition Ship Target is denoted asP indicates the pixel number for being divided into naval vessel in formula.
In formula (9), R is by all pixels in imageThe weights such as autocorrelation matrix be weighted and averaged,
And R plays the role of inhibiting background in weighting autocorrelative CEM algorithm.In order to allow influence of the echo signal d to R as far as possible
It is small even without the R in formula (9) being changed to B, it is all samplesSubtract P object vectorAfterwards
Autocorrelation matrix, such formula (9) then becomes
In formula, T is the sample autocorrelation matrix of all echo signals；E=wTTw is that echo signal is put down after wave filter
Export energy；B is the autocorrelation matrix that all samples subtract n-P vector composition after P object vector；
The average output energy of n-P vector after P object vector is subtracted for all samples.
Equally, eigenvalue problem is then equivalent to for the evaluation problem of formula (15).
B-1Tw=λ w (16)
In this way when in formula (11) R replaced by B, then new operator is
In formula,Wherein f (0)=0, f (x) is monotonic increase letter
Number (x >=0)；C is a constant；G (x, y) is measuring vector x, the function of y similarity.Here the angle between spectral vector is selected
For weighted sample matrix, due in improved autocorrelation matrix B*In, it has subtracted and has been believed using the target after SAM preliminary classification
Number d, therefore, echo signal d will reduce to B*Influence.
Concrete processing procedure of the invention is as shown in Figure 1.
1. dimension-reduction treatment
Minimal noise separation (Minimum Noise Fraction, MNF) transformation for determine in image data dimension
It counts, the noise in isolated data.MNF is substantially the principal component transform of stacked processing twice, as a result as shown in Figure 2.For the first time
Transformation is based on the estimation to noise covariance matrix, for separating and readjusting the noise in data, so that noise has list
Position variance, and wave band is uncorrelated, i.e. noise whitening.The noise data of second step whitening carries out the principal component transform of standard.It is logical
The analysis to characteristic value and high-spectral data is crossed, determines spectral Dimensions less than 20, to 10 dimension spectroscopic datas and 20 dimension spectrum numbers
On the basis of analysis, 20 dimension spectroscopic data of discovery is more preferable in Spectral angle mapper classification and CEM target acquisition effect, so finally will be former
417 dimensions of beginning high spectrum image are reduced to 20 dimensions.
2. Endmember extraction
" end member " is defined as being used to represent the Utopian clear data of category feature in high-spectral data, extracts pure spectrum number
It is stated to be spectrum Endmember extraction, is the premise that high-spectral data solution mixes the analysis of other high-spectral datas.Pure pixel index
(pure pixel index, PPI) algorithm (Boardman et al, 1995) is that one kind finds wave spectrum most in Hyperspectral imaging
The method of Pure pixel, be using end member be monomorphous of the high spectrum image formed in feature space end member the characteristics of with
And the property of the vector projection of monomorphous carries out Endmember extraction[11-12].The characteristics of according to monomorphous, monomorphous is arbitrarily straight in space
Projection on line must be straight line, and the endpoint of line segment must be the projection on monomorphous vertex.Pure pixel index is used to record each picture
Member is projected to line segment endpoint number.It is projected by generating a large amount of random straights, a width PPI image is generated, such as Fig. 3 institute
Show, the DN value of each pixel indicates that the pixel is marked as the total degree of Pure pixel.Obviously, pure picture corresponding to some pixel
First index is bigger, illustrates that a possibility that pixel is end member is higher.
The present embodiment carries out Endmember extraction (such as Fig. 4) by the N-dimensional visualization tool of Envi software.Pass through PPI first
Image meets the area-of-interest (such as Fig. 5) of threshold value to obtain, and then selection has multiple point sets in the interface n-D Controls
Group or cluster corner region are area-of-interest, select 6 area-of-interests for the end member for the 6 class atural objects to be extracted altogether.
Because high spectrum image can be divided into 6 class atural objects by the observation to image[14-15], be respectively as follows: seawater, sky, naval vessel,
Island, building 1, building 2.6 end members are generated using 6 area-of-interests of acquisition, 6 colors interested are respectively red
Color, green, blue, yellow, cyan, carmetta (such as Fig. 6).By area-of-interest using the end member selected as mean value wave spectrum, such as
Shown in table one, the end member on naval vessel is denoted as A1.
One Endmember extraction result of table
It is as shown in Figure 7 that Spectral angle mapper classification (SAM) result is carried out using the extracted end member of table one.
3. target acquisition
It is directly as shown in Figure 8 using the detection result of CEM algorithm.Due to only knowing target in CEM algorithm, and to background one
Ignorant, the R in formula (4) is the autocorrelation matrix of all pixels, since R plays background and target in the work of inhibition
With only target generally all very little during target acquisition, so influence of the target for background energy can generally neglect
Slightly disregard, this significant effect that is exactly CEM operator when Small object extracts, and the reason of for big Objective extraction poor effect.
Based on the CEM algorithm [8] of weighting autocorrelation matrix, when calculating object vector and each sample inner product, increase by one
A weighting function g (x, y) is used to measuring vector x, and y similarity, with vector x, the similarity of y increases and reduces the value of g (x, y).
In this way, being the reduction of influence of the area-of-interest as background to R by auto-correlation function, but influence not completely eliminate,
CEM algorithm detection result using weighting autocorrelation matrix is as shown in Figure 9.
In order to thoroughly reduce area-of-interest as background to R in formula (13)*Influence, the invention proposes improved
Based on autocorrelative SAM-CEM algorithm is weighted, the end member for first passing through extraction carries out SAM and classifies to determine two important classification:
Naval vessel class and seawater class；Naval vessel class pixel is subtracted from all pixels again as pure backdrop pels.Although may include in classification
Part non-targeted pixel, such as the red pixel in Fig. 7 in addition to crossing naval vessel, but compared to for naval vessel these accidentally divide
Pixel proportion it is smaller.So by classifying and subtracting the pretreatment works such as target, so that interesting target is for formula
(17) R in*Contribution nearly close to zero, i.e., area-of-interest is as background to R*Influence substantially completely eliminate, innovatory algorithm
CEM detection result afterwards is as shown in Figure 10.
By being analyzed above it will be seen that target is bigger, then the goal pels subtracted are more, and target is for R*Tribute
It offers smaller.Efficiently solve the problems, such as that original autocorrelative CEM algorithm of weighting is bad to big Target scalar extraction effect.
4. target separates
Image after using innovatory algorithm to carry out target acquisition makes Ship Target more obvious, but the reality of Ship Target
Border coordinate and naval vessel number are calculated completely not yet.In order to obtain naval vessel real coordinate position and number, to target acquisition
Image afterwards carries out binaryzation, as shown in Table 3, selects 60% threshold value of thresholding, 0.348 effect best, binarization result such as Figure 13
It is shown.But the image after binaryzation not only includes naval vessel information, and island is also identified as to comprising part the false-alarm of Ship Target
Target, thus need weed out be not naval vessel search coverage.By analysis it will be seen that naval vessel is all in the packet of seawater
Among enclosing, so all Ship Targets are found as long as finding seawater region, to reject island region to statistics naval vessel
Several influences.
Target separating step is as follows:
1) Spectral angle mapper classification (SAM) charting is carried out using the naval vessel end member extracted, the pixel that would be classified as seawater is set to
255, other are set to 0.
2) morphology processings such as corroded and expanded to seawater region above[8], find the maximum white of range
Region is seawater region, as shown in figure 11.
3) the CEM detection result gray level image (such as Figure 12) after innovatory algorithm, to the gray level image select threshold value 0.348 into
Row binaryzation, as a result as shown in figure 13.
4) Ship Target is searched in seawater region, i.e. it is Ship Target that searching Fig. 8 and Figure 13, which is 255 pixel,.It is logical
Crossing and calculating the target number for obtaining naval vessel is 7, and the centre coordinate of Ship Target and pixel number be as shown in table four and Figure 14,
Final realize separates Ship Target from seawater background.
Two Ship Target Detection statistical information of table
| Naval vessel number | Naval vessel centre coordinate | Pixel number |
| 1 | (192.59,245.82) | 1008 |
| 2 | (521.51,255.18) | 64 |
| 3 | (965.18,278.13) | 96 |
| 4 | (741.32,290.9) | 99 |
| 5 | (518.91,295.23) | 582 |
| 6 | (836.62,313.53) | 468 |
| 7 | (749.78,336.27) | 815 |
The present invention is according to dimensionality reduction, Endmember extraction[13], target acquisition and target separate 4 coastal high-spectral datas of step into
Row processing, has finally extracted the Ship Target for including in image, and calculated the number on naval vessel and the seat at naval vessel center
Cursor position.The target acquisition of CEM algorithm, weighting auto-correlation CEM algorithm, innovatory algorithm that Fig. 8-11 is shown is as a result, from Figure 11 phase
Become more blue than the background of Figure 10, illustrates that improved SAM-CEM target acquisition algorithm makes background obtain smaller inhibition.Below
It improves index by performance of the quantitative method to innovatory algorithm to analyze, mainly from ROC curve[4], false-alarm number, mean value side
Difference and target background energy ratio several aspects embody, as shown in Figure 15 and table three, table four, table five.
The detection parameters of 3 two kinds of algorithms of table compare
The false-alarm number of 4 two kinds of algorithms of table compares
5 two kinds of algorithm backgrounds of table and target component compare
AUC is the area below curve in table three, and S.E. is the area of standard deviation, and C.I. is confidence interval.From Figure 15
ROC curve and table three as can be seen that compared to weighting auto-correlation CEM algorithm, the ROC curve of innovatory algorithm rises to from 0.78
0.83, confidence level rises to 0.86 from interval from 0.81, illustrates that the reliability of innovatory algorithm is improved.
From the point of view of table four, for weighting auto-correlation CEM algorithm, when thresholding 50%, false alarm rate 100%；Thresholding
When 60%, naval vessel is all detected, but is non-Ship Target enormous amount；When thresholding 70%, naval vessel is all detected, but part warship
Ship target missing is serious；When thresholding greater than 70%, false alarm rate is increased rapidly, so the threshold value of weighting auto-correlation CEM algorithm
It can not determine.And for improved SAM-CEM algorithm, when thresholding 50%, Ship Target is all detected, but can not be by warship
Ship target is separated；When 60% thresholding, separating effect is best；When thresholding greater than 70%, false alarm rate is increased rapidly, so
60% threshold value 0.34815 of final choice.By the analysis to false-alarm number, illustrate innovatory algorithm in 60% thresholding, separation is imitated
Fruit is best, and weights auto-correlation CEM algorithm and be difficult to find optimum thresholding.
In order to calculate the energy of signal and background, the gray value of final detection image is normalized between 0-255.From
Table five can be seen that improved algorithm, and target mean becomes 206.17 from 204.73, illustrates that echo signal is enhanced, and carry on the back
The mean value of scape signal becomes 13.19 from 13.20, illustrates that background is inhibited.By to target energy after innovatory algorithm and back
The ratio between scape energy becomes 580352/6579079=0.0882 from 576437/6566665=0.0878, illustrates mesh after innovatory algorithm
Mark is enhanced and background is inhibited.
In summary it analyzes, innovatory algorithm target acquisition performance is significantly improved, and has shown echo signal especially and has inhibited
Background signal, so as to avoid echo signal in weighting auto-correlation CEM algorithm as background signal participation operation and to detection
As a result influence.
Bibliography
[1]Tello M,Martinez C L,Mallorqui J J.A Novel Algorithm for Ship
Detection in SAR Imagery Based on The Wavelet Transform[J].IEEE Geoscience
and Remote Sensing Letter,2005,2(2):201-205.
[2]Buck H,Sharghi E,Bromley K,et al.Ship Detection&Classification
from Overhead Imagery[C].Proc.Of SPIE,2007 6696(66961C).
[3]Gao G,KUANG G Y,ZHANG Q,LI D R.Fast Detecting and Locating Groups
of Targets in High-Resolution SAR Images[J].Pattern Recognition,2007,40:1378-
1384.
[4] Zhang Bing, Gao Lianru classification hyperspectral imagery and the Beijing target acquisition: Science Press, 2011.
[5] Zhang Baohua, Huang Wenqian, slight damage of the such as Li Jiangbo based on high light spectrum image-forming technology and MNF detection apple
[J] spectroscopy and spectrum analysis, 2014,34 (5): 1367-1372.
[6] dimension reduction method [J] spectroscopy and spectrum analysis in a kind of spectrum analysis of Li Qingbo, Jia Zhaohui, 2013,33
(3): 780-783.
[7] Wang Jie, poplar the Liao Dynasty, two kinds of TM image Endmember extraction algorithms combined based on space with spectrum of the such as Shen Jinxiang
[J] spectroscopy and spectrum analysis, 2011,31 (5): 1286-1289.
[8] target acquisition of Geng Xiurui high-spectrum remote sensing and sorting technique research [D], Chinese Academy of Sciences's remote sensing application
Research institute Ph.D. Dissertation
[9] Li Shanshan, Zhang Bing, even such as is based on the smallest hyper-spectral target detection algorithm research Acta Optica of variance to height
[J], 2010,30 (7): 2116-2122.
[10] Li Na is sought, Fang Yonghua, small target deteection algorithm [J] light based on Endmember extraction in Li Xin high spectrum image
Learn journal, 2007,27 (7): 7.
Hyperspectral imaging Endmember extraction algorithm [J] of the spectrum minimal information entropy such as [11] Yang Keming, Liu Shiwen, Wang Linwei
Spectroscopy and spectrum analysis, 2014,34 (8): 2229-2233.`
A kind of improved N-FINDR EO-1 hyperion Endmember extraction algorithm [J] electronics of the such as [12] Zhao Junhui, Qi Bin, Wang Yulei
With information journal .2012,34 (2): 499-502.
[13] Zhang Chunmei, arid biogeographic zone meadow Spectral Characteristics Analysis [J] of the Zhang Jianming based on Hyperspectral imaging, spectroscopy with
Spectrum analysis .2012,32 (2): 445-448
[14] a kind of high spectrum image unsupervised classification based on Space Consistency drop member of Yue Jiang, Zhang Yili, Xu Hangwei
[J] spectroscopy and spectrum analysis, 2012,32 (7): 1860-1864.
[15] Xu Weidong, Yin's ball, Kuang Dingbo COMPARISON OF DIFFERENT SPECTRAL MATCH MODELS [J] is infrared with millimeter wave journal,
2005,24 (8): 296-300.
Claims (3)
1. the waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm, which comprises the following steps:
1) dimension-reduction treatment is carried out to high spectrum image using MNF transformation；
2) Endmember extraction is carried out using PPI method be converted into the mesh to known target, Unknown Background after determining naval vessel end member
Mark detection problem；
3) SAM classification: naval vessel class and seawater class is carried out to high spectrum image by the end member extracted；It is subtracted from all pixels again
Naval vessel class pixel is as pure backdrop pels；Target spy is carried out by weighting the method for autocorrelation matrix based on pure backdrop pels
It surveys；
4) binaryzation is carried out to the image that target acquisition obtains, and Ship Target is separated from seawater background.
2. the waterborne target detection method according to claim 1 based on SAM and weighting auto-correlation CEM algorithm, feature
It is, step 4) specifically includes:
A gray level image only comprising seawater and naval vessel is obtained by the determining classification image of step 3), and carries out binaryzation sum number
Morphological scale-space is learned, the searching maximum white area of range is seawater region；
Binaryzation is carried out by the image obtained to target acquisition, removes the not false-alarm targets in seawater region, so that it is determined that
Final Ship Target.
3. the waterborne target detection method according to claim 2 based on SAM and weighting auto-correlation CEM algorithm, feature
It is, in the step 3) method by weighting autocorrelation matrix based on pure backdrop pels, the function for solving extreme value is changed to:
In formula, T is the sample autocorrelation matrix of all echo signals；W is filter vector；E=wTTw is echo signal by filtering
Energy is averagely exported after device；B is the autocorrelation matrix that all samples subtract n-P vector composition after P object vector；The average output energy of n-P vector after P object vector is subtracted for all samples；
Correspondingly, new target acquisition operator is
In formula,Wherein f (0)=0, f (x) is monotonically increasing function (x
≥0)；C is a constant；G (x, y) is measuring vector x, the function of y similarity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811528890.8A CN109801226A (en) | 2018-12-13 | 2018-12-13 | Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811528890.8A CN109801226A (en) | 2018-12-13 | 2018-12-13 | Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109801226A true CN109801226A (en) | 2019-05-24 |
Family
ID=66556707
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811528890.8A Pending CN109801226A (en) | 2018-12-13 | 2018-12-13 | Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109801226A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112016511A (en) * | 2020-09-08 | 2020-12-01 | 重庆市地理信息和遥感应用中心 | Remote sensing image blue top room detection method based on large-scale depth convolution neural network |
| CN114663654A (en) * | 2022-05-26 | 2022-06-24 | 西安石油大学 | Improved YOLOv4 network model and small target detection method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101806898A (en) * | 2010-03-19 | 2010-08-18 | 武汉大学 | Hyperspectral remote sensing image target detecting method based on variable end members |
| CN102609703A (en) * | 2012-03-05 | 2012-07-25 | 中国科学院对地观测与数字地球科学中心 | Method and device for detecting target ground object in hyperspectral image |
| CN105115906A (en) * | 2015-08-13 | 2015-12-02 | 成都理工大学 | Sericitization alteration information extracting method based on own features of image |
| KR20180064154A (en) * | 2016-12-05 | 2018-06-14 | 주식회사 포스코 | Training guide supplying system for virtual compact endless cast and rolling mill based on trainee's control |
| CN108195771A (en) * | 2017-12-18 | 2018-06-22 | 河海大学 | A kind of ocean target in hyperspectral remotely sensed image target identification method |
| CN108593575A (en) * | 2018-04-18 | 2018-09-28 | 成都理工大学 | A kind of multi-spectrum remote sensing alteration information metallogenic prognosis method based on wave spectrum hourglass tool |
-
2018
- 2018-12-13 CN CN201811528890.8A patent/CN109801226A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101806898A (en) * | 2010-03-19 | 2010-08-18 | 武汉大学 | Hyperspectral remote sensing image target detecting method based on variable end members |
| CN102609703A (en) * | 2012-03-05 | 2012-07-25 | 中国科学院对地观测与数字地球科学中心 | Method and device for detecting target ground object in hyperspectral image |
| CN105115906A (en) * | 2015-08-13 | 2015-12-02 | 成都理工大学 | Sericitization alteration information extracting method based on own features of image |
| KR20180064154A (en) * | 2016-12-05 | 2018-06-14 | 주식회사 포스코 | Training guide supplying system for virtual compact endless cast and rolling mill based on trainee's control |
| CN108195771A (en) * | 2017-12-18 | 2018-06-22 | 河海大学 | A kind of ocean target in hyperspectral remotely sensed image target identification method |
| CN108593575A (en) * | 2018-04-18 | 2018-09-28 | 成都理工大学 | A kind of multi-spectrum remote sensing alteration information metallogenic prognosis method based on wave spectrum hourglass tool |
Non-Patent Citations (2)
| Title |
|---|
| WANG ZHENGZHOU ET AL.: "Surface Ship Target Detection in Hyperspectral Images Based on Improved Variance Minimum Algorithm", 《8TH INTERNATIONAL CONFERENCE ON DIGITAL IMAGE PROCESSING (ICDIP)》 * |
| 耿修瑞: "高光谱遥感图像目标探测与分类技术研究", 《中国博士学位论文全文数据库》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112016511A (en) * | 2020-09-08 | 2020-12-01 | 重庆市地理信息和遥感应用中心 | Remote sensing image blue top room detection method based on large-scale depth convolution neural network |
| CN114663654A (en) * | 2022-05-26 | 2022-06-24 | 西安石油大学 | Improved YOLOv4 network model and small target detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | A novel region-merging approach for coastline extraction from sentinel-1A IW mode SAR imagery | |
| CN107563303B (en) | Robust ship target detection method based on deep learning | |
| Abdelmounaime et al. | New Brodatz-based image databases for grayscale color and multiband texture analysis | |
| CN108389188B (en) | Sparse hyperspectral abnormal target detection method | |
| CN102214298B (en) | Method for detecting and identifying airport target by using remote sensing image based on selective visual attention mechanism | |
| CN102096824B (en) | Multi-spectral image ship detection method based on selective visual attention mechanism | |
| Wang et al. | A robust infrared dim target detection method based on template filtering and saliency extraction | |
| CN109801226A (en) | Waterborne target detection method based on SAM and weighting auto-correlation CEM algorithm | |
| Sun et al. | Constrained-target band selection with subspace partition for hyperspectral target detection | |
| CN112434642B (en) | Sea-land segmentation method suitable for processing large-scene optical remote sensing data | |
| CN111259756A (en) | Pedestrian re-identification method based on local high-frequency features and mixed metric learning | |
| CN108364011A (en) | PolSAR image multi-stage characteristics extract and unsupervised segmentation method | |
| Fang et al. | Detection of building shadow in remote sensing imagery of urban areas with fine spatial resolution based on saturation and near-infrared information | |
| CN112163606B (en) | Infrared small target detection method based on block contrast weighting | |
| CN112418156B (en) | Unsupervised classification-based optical remote sensing image sea surface ship target detection method | |
| CN115393577A (en) | Infrared marine small target anti-interference detection method | |
| Chen et al. | An edge detection method for hyperspectral image classification based on mean shift | |
| Fu | Texture feature extraction and recognition of underwater target image considering incomplete tree wavelet decomposition | |
| Yan et al. | Integration of spatial-spectral information for hyperspectral image classification | |
| Wu et al. | Spectra-difference based anomaly-detection for infrared hyperspectral dim-moving-point-target detection | |
| Singh et al. | The H/A/Alpha polarimetric decomposition theorem and complex wishart distribution for snow cover monitoring | |
| CN109446899A (en) | A kind of cloud object detection method based on four spectral coverage remote sensing images | |
| CN109949320A (en) | The super pixel dividing method of high spectrum image based on entropy and mutual information | |
| Raja et al. | Evaluation of fusion approaches for face recognition using light field cameras | |
| Zhang et al. | Region of interest detection based on salient feature clustering for remote sensing images |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190524 |