CN100483147C - High spectrum sub-pixel target detection method and device - Google Patents

Abstract

The present invention discloses a high spectrum sub-pixel object detecting method. The method includes that: a reverse model of the objective spectrum and the pixel spectrum two-dimensional matrix of the image needing to be detected is created, the regression coefficient vector of the reverse model is acquired; the Mahalanobis distance of each pixel is acquired according to the regression coefficient vector; the pixel corresponding to the regression coefficient of the Mahalanobis distance being greater than the threshold is determined as the sub-pixel objective point. The present invention also provides a high spectrum sub-pixel object detecting device, the sub-pixel object detecting method and the device thereof using the present invention need the prior information of the background terminal spectrum; are not sensitive to the complexity of the background spectrum; and have higher object detecting accuracy and faster operation speed.

Description

High spectrum sub-pixel target detection method and device

Technical field

The present invention relates to high-spectrum remote-sensing Detection Techniques field, relate in particular to a kind of high spectrum sub-pixel target detection method and device.

Background technology

High-spectrum remote-sensing is one of most important technological breakthrough in the earth observing system of twentieth century end, it has overcome traditional single band, the multispectral remote sensing limitation at aspects such as wave band number, wavelength band, meticulous information representations, provide sensor information with narrower wave band interval, more wave band quantity, can from spectral space, be segmented and be differentiated, obtained widespread use in fields such as resources remote sensing, environmental remote sensing, ecological Remote Sensing to atural object.The topmost characteristics of high-spectrum remote sensing data are to be one with traditional image dimension and spectrum dimension information fusion, when obtaining the spatial surface image, obtain the continuous spectrum information of each atural object, this spectral information can reflect the nuance of Weak target and background under the complex background, therefore can find that space characteristics be difficult to the terrain object that maybe can't survey by abundant spectral information.Because the high-spectrum remote-sensing spatial resolution is limited at present, single pixel is difficult to include only a kind of atural object composition, therefore its spectrum is the aliasing of multiple object spectrum often, is called sub-pixel target detection for this type of volume size less than the target detection of pixel spatial resolution.Aforesaid spatial resolution promptly is meant the ground object minimum dimension that can distinguish on the image.High spectrum target detection technology is widely used in fields such as military target detection, mineral prospecting, vegetation distribution assessment, environmental pollution detection.

At present common sub-pixel target detection method mainly comprises: the detection method that the complex spectrum least square separates, based on the detection method of orthogonal subspaces projection (OSP, Orthogonal Subpixel Projection) with based on the detection method of Generalized Likelihood Ratio etc.

Wherein, the detection method that the complex spectrum least square separates is: because the arbitrary pixel spectrum in the testing image all can be regarded the weighted sum of the spectrum of multiple base substance as, weights are the shared blending ratio of each component in the corresponding pixel, if the pure spectrum of each pixel component can obtain from known spectra database in the image atural object, then can utilize component spectra that Constrained Least Square Methods calculates the target correspondence shared ratio in mixed pixel spectrum, thereby survey target less than the ground pixel.This method can be carried out quantitative test to pixel spectrum, but prerequisite is the prior imformation that needs each pixel component spectra, therefore is difficult to be applied to the target detection field of background the unknown.

Based on the detection method of OSP, mainly be with the orthogonal subspaces of pixel spectrum vector projection, thereby eliminate background information, outstanding target optical spectrum information to background characteristics.Exist a kind of unsupervised orthogonal subspaces projecting method that inferior pixel target is detected in the prior art, this method at first utilizes a kind of method of iteration to find the end member spectrum of one group of background atural object, it also is the pure spectrum of each component in the atural object, then utilize these background atural object end member spectrum structures to detect operator, dynamically inferior pixel target is detected.This method can construct OSP and detect operator under background atural object end member condition of unknown, ground unrest there is certain inhibiting effect, but, cause detection accuracy not high, and calculate consuming time longer because it is not high to calculate resulting background atural object end member spectrum degree of accuracy.

Also there is a kind of detection method in the prior art based on Generalized Likelihood Ratio, mainly be that hypothesis background spectrum information satisfies certain multidimensional distribution earlier, construct certain detection operator then and judge in the pixel whether contain potential target optical spectrum by the method for test of hypothesis.This method can provide theoretic false alarm rate, thereby regulates threshold value adaptively, but this method can only detect the singular point in the background, and needs certain image information, and the quality of training sample is bigger to the influence of result of detection.Because this detection method needs the prior imformation of background atural object, and comparatively responsive to the background complexity, can not carry out in the small target detection field of change of background complexity.

In sum, sub-pixel target detection method of the prior art mostly needs the prior imformation of background atural object end member spectrum, and is comparatively responsive to the complexity of background spectrum, and the accuracy of target detection is lower, speed is slower.

Summary of the invention

In view of this, fundamental purpose of the present invention is to provide a kind of high spectrum sub-pixel target detection method and device, need background end member spectrum prior imformation to solve sub-pixel target detection method of the prior art, the low and slow defective of target detection accuracy.

For achieving the above object, technical scheme of the present invention is achieved in that

A kind of high spectrum sub-pixel target detection method of the present invention may further comprise the steps:

Set up the inversion model of target optical spectrum and testing image pixel spectrum two-dimensional matrix;

Obtain the regression coefficient vector of described inversion model;

Obtain the mahalanobis distance of each pixel according to described regression coefficient vector;

Judge that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

Wherein, the described inversion model of setting up target optical spectrum and testing image pixel spectrum two-dimensional matrix specifically comprises:

The three-dimensional high-spectral data of the testing image pixel that obtains is expressed as the two-dimensional matrix of high spectrum reflection rate: R _{M * n}=[p ₁, p ₂... p _{Y * i+j}... p _{X * y}], 0＜i≤x, 0＜j≤y,

Or R _{M * n}=[p ₁, p ₂... P _{I+x * j}... p _{X * y}], 0＜i≤x, 0＜j≤y,

Wherein, R _{M * n}The two-dimensional matrix of expression pixel spectrum, [p ₁, p ₂... p _{Y * i+j}... p _{X * y}] and [p ₁, p ₂... P _{I+x * j}... * p _{X * y}] expression pixel the spectrum vector, m represents the wave band number, n presentation video pixel sum, x presentation video pixel line number, y presentation video pixel columns, n=x * y;

Set up the linear relationship of described target optical spectrum and described testing image pixel spectrum two-dimensional matrix:

S _t＝R _m×n×c _PLS+E _m×1，

Wherein, S _tExpression target optical spectrum vector, R _{M * n}The two-dimensional matrix of expression pixel spectrum, c _PLSThe expression regression coefficient, E _{M * 1}The expression noise vector.

Wherein, describedly three-dimensional high-spectral data is expressed as two-dimensional matrix and sets up between the linear relationship, also comprise: described target optical spectrum and described testing image pixel spectrum two-dimensional matrix are carried out pre-service.

Wherein, described pre-service comprises the orthonormal transformation processing or adds scatter correction and handle.

Wherein, described orthonormal transformation is handled and is specifically comprised:

$p_{\mathrm{hk},\mathrm{SNV}}=\frac{p_{\mathrm{hk}}-{\stackrel{\‾}{p}}_h}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{(p_{\mathrm{hk}}-{\stackrel{\‾}{p}}_h)}^2}}{(m-1)}^{1/2}$

Wherein, p _{Hk, SNV}Expression is handled in the testing image of back the reflectance value of h pixel at k wave band, p through orthogonal transformation _hH pixel is at the mean value of each wave band place reflectivity in the expression testing image, and m represents the wave band number, and m-1 represents degree of freedom;

Described additional scatter correction is handled and is specifically comprised:

Calculate the averaged spectrum vector: $\stackrel{\‾}{p}=\frac{1}{n}\underset{h=1}{\overset{n}{\mathrm{\Σ}}}{p}_h;$

Each pixel spectrum is carried out linear regression: p _h=m _hP+b _h

Add scatter correction: $p_{h\left(\mathrm{MSC}\right)}=\frac{(p_h-b_h)}{m_h};$

Wherein, p represents the averaged spectrum vector, Expression is to the summation of all pixel spectrum vectors, p _hThe spectrum vector of representing h pixel, m _h, b _hRepresent h pixel spectrum vector p respectively _hWith the slope and the intercept of the linear regression of all pixel averaged spectrum, p _{H (MSC)}Pixel spectrum vector behind the additional scatter correction of expression process.

Wherein, obtain the regression coefficient vector of described inversion model, specifically comprise by the offset minimum binary alternative manner:

A, obtain the initial weight vector according to the target optical spectrum vector: $w_n={\mathrm{RS}}_t^T,$ Wherein, S _tExpression target optical spectrum vector, w _nExpression initial weight vector, R represents the two-dimensional matrix of pixel spectrum;

B, according to described initial weight vector calculation score vector: t _n=R ^Tw _n, wherein, t _nThe expression score vector, R represents the two-dimensional matrix of pixel spectrum, w _nExpression initial weight vector;

C, calculate the load vector of described target optical spectrum according to described score vector: q _n=S _tt _n, wherein, q _nThe load vector of representing described target optical spectrum, S _tExpression target optical spectrum vector, t _nThe expression score vector;

D, calculate the load vector of described pixel spectrum two-dimensional matrix according to described score vector: p _n=Rt _n, wherein, p _nRepresent the load vector of described pixel spectrum two-dimensional matrix, R represents the two-dimensional matrix of pixel spectrum, t _nThe expression score vector;

E, obtain described regression coefficient vector: c _PLS=W (P ^TW) ^-1Q ^T, wherein, Q={q ₁, q ₂... q _n, P={p ₁, p ₂... p _n, W={w ₁, w ₂... .w _n;

F, calculating residual sum of squares (RSS): $\mathrm{SS}\left(n\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{(S_t-R_0{c}_{\mathrm{PLS}})}^2,$ Wherein, SS (n) represents residual sum of squares (RSS), and m represents wave band number, R ₀Represent original pixel spectrum two-dimensional matrix, S _tExpression target optical spectrum vector, the setting threshold values is G, if SS (n-1)-SS (n)≤G then gets the corresponding c of this SS (n) _PLSBe regression coefficient vector; Otherwise, make S _t=S _t-t _nq _n, R=R-t _np _n, return step a and repeat aforesaid operations, up to SS (n-1)-SS (n)≤G, obtain regression coefficient vector c then _PLS

The present invention also provides a kind of high spectrum sub-pixel target detection device, comprising:

The modelling unit is used to set up the inversion model of target optical spectrum and testing image pixel spectrum two-dimensional matrix;

The regression coefficient vector acquiring unit is used to obtain the regression coefficient vector of described inversion model;

The mahalanobis distance acquiring unit is used for obtaining according to described regression coefficient vector the mahalanobis distance of each pixel;

Identifying unit is used to judge that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

Wherein, described modelling unit comprises:

Matrix generates subelement, is used for the three-dimensional high-spectral data of testing image pixel is expressed as the two-dimensional matrix of high spectrum reflection rate;

Linear relationship is set up subelement, is used to set up the linear relationship of described target optical spectrum and described testing image pixel spectrum two-dimensional matrix.

Wherein, described modelling unit also comprises: the pre-service subelement is used for the testing image pixel spectrum two-dimensional matrix of described target optical spectrum and the generation of described matrix generation subelement is carried out pre-service and offer linear relationship setting up subelement.

Wherein, described device also comprises: data capture unit is used to obtain the three-dimensional high-spectral data of testing image pixel and offers the modelling unit.

Compared with prior art, the present invention has the following advantages:

The used inversion model of the present invention does not need the prior imformation of background end member spectrum, only needs the target optical spectrum information of priori, and is therefore insensitive to the complexity of background spectrum; Detection accuracy only depends on the levels of precision of target optical spectrum, thus target detection accuracy height; Because the iterative calculation method in the offset minimum binary only needs a small amount of matrix inversion operation, and adopted speed Ma Shi singular value detection method faster again, made detection method of the present invention have arithmetic speed faster in the test section.

Description of drawings

Fig. 1 is the process flow diagram of a kind of high spectrum sub-pixel target detection method of the present invention;

Fig. 2 is a synoptic diagram of the original high spectrum image of the embodiment of the invention;

Fig. 3 is another synoptic diagram of the original high spectrum image of the embodiment of the invention;

Fig. 4 is the target optical spectrum synoptic diagram of the embodiment of the invention;

Fig. 5 is the pixel point synoptic diagram behind the target optical spectrum of sneaking into of the embodiment of the invention;

Fig. 6 is the regression coefficient schematic three dimensional views of the embodiment of the invention;

Fig. 7 is the mahalanobis distance synoptic diagram of the embodiment of the invention;

Fig. 8 is the sub-pixel target detection result schematic diagram of the embodiment of the invention;

Fig. 9 is the composition structural representation of a kind of high spectrum sub-pixel target detection device of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments technical scheme of the present invention is done further and to be elaborated:

As shown in Figure 1, Fig. 1 is the process flow diagram of a kind of high spectrum sub-pixel target detection method of the present invention, mainly may further comprise the steps:

Step 101 is set up the inversion model of target optical spectrum and testing image pixel spectrum two-dimensional matrix.

At first, the three-dimensional high-spectral data that gets access to the testing image pixel is expressed as the two-dimensional matrix of high spectrum reflection rate.The three-dimensional high-spectral data of this testing image pixel is to be obtained by imaging spectrometer, imaging spectrometer can obtain a width of cloth bearing member, image at each wave band, and be high spectrum image by the cube that the image overlay of the continuous wave band of some row forms, this high spectrum image comprises the two dimensions of information of image space and the one-dimension information of spectral band direction, therefore is three-dimensional high-spectral data.Obtain the three-dimensional high-spectral data of testing image pixel, and the two-dimensional matrix that is expressed as of high-spectral data that should three-dimensional:

R _m×n＝[p ₁，p ₂...p _y×i+j...p _x×y]，0<i≤x，0<j≤y (1)

Or R _{M * n}=[p ₁, p ₂... p _{I+x * j}... p _{X * y}], 0＜i≤x, 0＜j≤y (2)

Wherein, (1) formula is the representation that the testing image pixel launches by row, and (2) formula is the representation that the testing image pixel launches by row, R _{M * n}The two-dimensional matrix of expression pixel spectrum, x represents high spectrum image pixel line number, and y represents high spectrum image pixel columns, and m represents the wave band number, n presentation video pixel sum, n=x * y.[p ₁, p ₂... p _{Y * i+j}... p _{X * y}] and [p ₁, p ₂... p _{I+x * j}... p _{X * y}] the spectrum vector of expression high spectrum image pixel, at the R that launches by row _{M * n}In, p _{Y * i+j}The capable j of i is listed as the spectrum vector of corresponding pixel in the expression testing image; In like manner, at the R that launches by row _{N * m}In, p _{I+x * j}Represent that then the capable j of i in the testing image is listed as the spectrum vector of corresponding pixel.The spectrum vector of each pixel comprises the reflectance value of this pixel at each wave band place, and for example: the spectrum vector of supposing h pixel is p _h, p then _h=[p _H1, p _H2... p _Hk... p _Hm] ^T, [p _H1, p _H2... p _Hk... p _Hm] ^TRepresentative [p _H1, p _H2... p _Hk... p _Hm] transposed matrix, p wherein _HkRepresent the reflectance value of this h pixel at k wave band place.Because it promptly is the reflectance value of this pixel point at this wave band place that imaging spectrometer gets access to the value of image picture elements point at each wave band place, so this reflectance value is a known quantity.

Then, target optical spectrum and testing image pixel spectrum two-dimensional matrix are carried out pre-service, to proofread and correct the spectral error that causes because of atmospheric scattering.This pre-service can be orthonormal transformation and handles, and also can be additional scatter correction and handles.Certainly, preprocess method of the present invention is not limited only to above-mentioned two kinds of disposal routes, and other any disposal routes that can proofread and correct the spectral error that causes because of atmospheric scattering also should belong to protection scope of the present invention.

The formula that orthonormal transformation is handled is as follows:

$p_{\mathrm{hk},\mathrm{SNV}}=\frac{p_{\mathrm{hk}}-{\stackrel{\&OverBar;}{p}}_h}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{(p_{\mathrm{hk}}-{\stackrel{\&OverBar;}{p}}_h)}^2}}{(m-1)}^{1/2}---\left(3\right)$

Wherein, p _{Hk, SNF}Expression is handled in the testing image of back the reflectance value of h pixel at k wave band, p through orthogonal transformation _hH pixel is at the mean value of each wave band place reflectivity in the expression testing image, and m represents the wave band number, and m-1 represents degree of freedom.

The process that additional scatter correction is handled comprises:

At first, calculate the averaged spectrum vector:

$\stackrel{\&OverBar;}{p}=\frac{1}{n}\underset{h=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_h---\left(4\right)$

Then, each pixel spectrum is carried out linear regression:

p _h＝m _hp+b _h (5)

Add scatter correction again:

$p_{h\left(\mathrm{MSC}\right)}=\frac{(p_h-b_h)}{m_h}---\left(6\right)$

Above-mentioned (4) in (6) formula, p represents the averaged spectrum vector,

Expression is to the summation of all pixel spectrum vectors, p _hThe spectrum vector of representing h pixel, m _h, b _hRepresent h pixel spectrum vector p respectively _hWith the slope and the intercept of the linear regression of all pixel averaged spectrum, p _{H (MSC)}Pixel spectrum vector behind the additional scatter correction of expression process.

Respectively each pixel spectrum in target optical spectrum and the testing image is carried out after the pre-service more accurate spectral information after the atmospheric scattering error that then can be eliminated.

At last, set up the linear relationship of target optical spectrum and testing image pixel spectrum two-dimensional matrix, this linear relationship is expressed as follows:

S _t＝R _m×n×c _PLS+E _m×1 (7)

Wherein, S _tExpression target optical spectrum vector, R _{M * n}The two-dimensional matrix of expression pixel spectrum, c _PLSThe expression regression coefficient vector, c _PLSThe size of value can reflect the percentage contribution of target optical spectrum to each pixel spectrum, c _PLSBe worth greatly more, show that then target optical spectrum is big more to the contribution of pixel spectrum, E _{M * 1}The expression noise vector is the one dimension noise matrix.

Step 102 is obtained the regression coefficient vector of this inversion model.

Because R _{M * n}Order less than wave band number, R _{M * n}In each pixel spectrum have high correlation, adopt common least square method that regression coefficient is carried out c _PLSIt is right to need during estimation

Invert,

Represent R _{M * n}Transposed matrix, and work as R _{M * n}In the variable height correlation time, determinant

Almost approach zero, right

Inverting to produce serious round-off error, therefore adopts common least square method to ask for the regression coefficient vector of this inversion model, specifically comprises:

A, according to target optical spectrum vector calculation initial weight vector:

$w_n={\mathrm{RS}}_t^T---\left(8\right)$

Wherein, S _tExpression target optical spectrum vector, S _tBe known quantity,

Represent S _tTransposed matrix, w _nExpression initial weight vector, R represents the two-dimensional matrix of pixel spectrum.

B, according to initial weight vector calculation score vector:

t _n＝R ^Tw _n (9)

Wherein, t _nThe expression score vector, R represents the two-dimensional matrix of pixel spectrum, R ^TRepresent the transposed matrix of R, w _nExpression initial weight vector.

C, calculate the load vector of target optical spectrum according to score vector:

q _n＝S _tt _n (10)

Wherein, S _tExpression target optical spectrum vector, q _nExpression load vector, t _nThe expression score vector.

D, calculate the load vector of pixel spectrum two-dimensional matrix according to score vector:

p _n＝Rt _n (11)

Wherein, p _nThe load vector of expression pixel spectrum two-dimensional matrix, R represents the two-dimensional matrix of pixel spectrum, t _nThe expression score vector.

E, calculating regression coefficient vector:

c _PLS＝W(P ^TW) ^-1Q ^T (12)

Wherein, Q={q ₁, q ₂... q _n, P={p ₁, p ₂... p _n, W={w ₁, w ₂... w _n, (P ^TW) ^-1Representative is to (P ^TW) carry out inversion operation.

F, calculating residual sum of squares (RSS):

$\mathrm{SS}\left(n\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(S_t-R_0{c}_{\mathrm{PLS}})}^2---\left(13\right)$

Wherein, SS (n) represents residual sum of squares (RSS), and m represents wave band number, R ₀Represent original pixel spectrum two-dimensional matrix,, get the corresponding c of this SS (n) if SS (n-1)-SS (n)≤G then shows convergence _PLSBe regression coefficient vector; Otherwise, make S _t=S _t-t _nq _n, R=R-t _np _n, return step a and repeat aforesaid operations, up to SS (n-1)-SS (n)≤G, obtain regression coefficient vector c then _PLSIn actual applications, the G value can be chosen according to actual needs.

Step 103 is obtained the mahalanobis distance of each pixel according to the regression coefficient vector that obtains.

According to the regression coefficient vector that obtains and utilize following formula to calculate the mahalanobis distance of each pixel, formula is as follows:

$d_h\left(c_h\right)=\sqrt{\frac{{(c_h-\stackrel{\&OverBar;}{c})}^2}{\mathrm{std}{\left(c\right)}^2}},h\&le;n---\left(14\right)$

Wherein, c _hThe regression coefficient value of corresponding h pixel in the expression regression coefficient vector, n represents the pixel sum, the standard deviation of std (c) expression regression coefficient.d _h(c _h) mahalanobis distance of regression coefficient of h pixel correspondence of expression.Aforementioned std (c) can try to achieve by following formula:

$\mathrm{std}\left(c\right)=\sqrt{\frac{1}{n-1}\underset{h=1}{\overset{n}{\mathrm{\&Sigma;}}}{(c_h-\stackrel{\&OverBar;}{c})}^2}---\left(15\right)$

Wherein, the standard deviation of std (c) expression regression coefficient, c _hThe regression coefficient value of representing h pixel correspondence, c represents the mean value of all pixel regression coefficients, $\stackrel{\&OverBar;}{c}=\underset{h=1}{\overset{n}{\mathrm{\&Sigma;}}}{c}_h.$

Step 104 judges that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

After calculating the pairing mahalanobis distance of each pixel regression coefficient, detection threshold can obtain fast according to following mode: can think that promptly regression coefficient must be the background pixel less than 0 pixel, then must there be the pixel point of a mahalanobis distance maximum in those known background pixels, the pairing mahalanobis distance of desirable this pixel is the segmentation threshold of background and target, i.e. the mahalanobis distance detection threshold.

Further elaborate below in conjunction with the high spectrum sub-pixel target detection method of specific embodiment the invention described above.The used high-spectrum remote sensing data of this example derives from airborne imaging spectrometer, airborne imaging spectrometer is for adopting the imaging spectrometer of push-scanning image mode, obtain the spatial image information at 224 wavelength places 0.4 micron-2.45 microns wavelength coverage, the wavelength interval is 10 nanometers, when aircraft during in 20 km high-altitude flights, image spatial resolution can reach 20 * 20 meters.The high spectrum image that present embodiment uses as shown in Figure 2, size is 614 * 512 pixels, each pixel spectrum comprises 224 wavelength, wavelength coverage from 369.85 nanometers to 2506.81 nanometers.Concrete detection process is as follows:

A, get that the image of 64 * 64 pixels is the testing image of present embodiment in the white box shown in Fig. 2, Fig. 3 then be the enlarged diagram of this testing image, the total n=64 * 64=4096 of testing image pixel.Reject 35 lower bad wave bands of signal to noise ratio (S/N ratio) such as 1-6,33,107-114 in 224 wave bands, 153-168,222-224 wave band, all the other 190 wave bands are the use wave band of present embodiment, also are that wave band is counted m=190.The higher roof spectrum of reflected energy value of getting arrow indication in the black circles of Fig. 2 lower right corner is the target optical spectrum of present embodiment, and the spectrogram of this target optical spectrum as shown in Figure 4.Coordinate is (10,32), (10,42), (10,52), (32 in extraction 64 * 64 images, 32), (32,42), (32,52), (42,32), (42,42), nine pixel points of (42,52), sneak into 5% target optical spectrum respectively, sneak into behind the target optical spectrum pixel point as shown in Figure 5.If the pixel in the image is launched by row, then the spectrum vector that the pixel point of (10,32) is corresponding is p _{64 * 10+32}, other pixel points are also in like manner described no longer one by one at this.

The testing image pixel point that it is pointed out that the embodiment of the invention is chosen arbitrarily, and the quantity of choosing pixel point also is arbitrarily, chooses nine pixel points herein simultaneously and is in order to show that detection method of the present invention can survey a plurality of pixel points simultaneously.After aforesaid nine pixel points are sneaked into target optical spectrum, the information that has promptly had target optical spectrum in these nine pixel points, then can will exist these nine points of target optical spectrum all to detect by detection method of the present invention, this also be the final purpose of the embodiment of the invention.

B, the three-dimensional high-spectral data of nine pixel points being obtained is expressed as the high spectrum reflection rate matrix of two dimension, method for expressing is not going to repeat as hereinbefore.

C, target optical spectrum and pixel spectrum two-dimensional matrix are carried out pre-service respectively, disposal route as hereinbefore, this also no longer stating more.Then, set up inversion model between target optical spectrum and the pixel spectrum.

D, utilize the regression coefficient vector c of single dependent variable offset minimum binary method to above-mentioned inversion model _PLSFind the solution, setting threshold values in this offset minimum binary method is 10 ^-3The three-dimensional plot of asking for regression coefficient as shown in Figure 6, the X among the figure, Y-axis are represented horizontal stroke, the ordinate of testing image respectively, the Z axle is represented the value of regression coefficient.As can be seen from Figure 6, the pixel point that is mixed with 5% target optical spectrum differs greatly with the regression coefficient value of not sneaking into the background pixel point of target optical spectrum, and the size of visible regression coefficient can reflect the percentage contribution of target optical spectrum to pixel spectrum.

E, calculate the mahalanobis distance of each pixel point regression coefficient.Result of calculation as shown in Figure 7, the transverse axis among the figure is represented the pixel number, the longitudinal axis is represented the mahalanobis distance value.Can check threshold values by above-mentioned selection of threshold method is 5.1402, and as can be seen from Figure 7, the mahalanobis distance of selected nine pixel point regression coefficients is all greater than setting threshold.Therefore, testing result as shown in Figure 8, selected nine pixel points all are judged as inferior pixel objective point, testing result conforms to actual conditions.

The present invention also provides a kind of high spectrum sub-pixel target detection device, and as shown in Figure 9, this device comprises: data capture unit 100, modelling unit 200, regression coefficient vector acquiring unit 300, mahalanobis distance acquiring unit 400 and identifying unit 500.Wherein, data capture unit 100 is used to obtain the three-dimensional high-spectral data of testing image pixel.Modelling unit 200 connects data capture unit 100, and the pixel high-spectral data that is used for obtaining according to data capture unit 100 is set up the inversion model of target optical spectrum and testing image pixel spectrum.Regression coefficient vector acquiring unit 300, link model are set up unit 200, are used to obtain the regression coefficient vector of inversion model.Mahalanobis distance acquiring unit 400 connects regression coefficient vector acquiring unit 300, is used for obtaining according to the regression coefficient vector that obtains the mahalanobis distance of each pixel.Identifying unit 500 connects mahalanobis distance acquiring unit 400, is used to judge that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

Wherein, modelling unit 200 also comprises: matrix generates subelement 210, pre-service subelement 220 and linear relationship and sets up subelement 230.Matrix generates subelement 210, is used for the three-dimensional high-spectral data of testing image pixel is expressed as the two-dimensional matrix of high spectrum reflection rate.Pre-service subelement 220, connection matrix generates subelement 210, is used for the testing image pixel spectrum two-dimensional matrix of target optical spectrum and 210 generations of matrix generation subelement is carried out pre-service and offer linear relationship setting up subelement 230.Linear relationship is set up subelement 230, connects pre-service subelement 220, is used to set up the linear relationship of target optical spectrum and pixel spectrum two-dimensional matrix to be measured.

In sum, the method for a kind of high spectrum sub-pixel target detection of the present invention and device, used inversion model does not need the prior imformation of background end member spectrum, only needs the target optical spectrum information of priori, and is therefore insensitive to the complexity of background spectrum; Detection accuracy only depends on the levels of precision of target optical spectrum, thus target detection accuracy height; Because the iterative calculation method in the offset minimum binary only needs a small amount of matrix inversion operation, and adopted speed Ma Shi singular value detection method faster again, made detection method of the present invention have arithmetic speed faster in the test section.

The above is preferred embodiment of the present invention only, is not to be used to limit protection scope of the present invention.

Claims (10)

1, a kind of high spectrum sub-pixel target detection method is characterized in that, may further comprise the steps:

Set up the inversion model of target optical spectrum and testing image pixel spectrum two-dimensional matrix;

Obtain the regression coefficient vector of described inversion model;

Obtain the mahalanobis distance of each pixel according to described regression coefficient vector;

Judge that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

2, high spectrum sub-pixel target detection method according to claim 1 is characterized in that the described inversion model of setting up target optical spectrum and testing image pixel spectrum two-dimensional matrix specifically comprises:

The three-dimensional high-spectral data of the testing image pixel that obtains is expressed as the two-dimensional matrix of high spectrum reflection rate: R _{M * n}=[p ₁, p ₂... p _{Y * i+j}... p _{X * y}], 0＜i≤x, 0＜j≤y,

Or R _{M * n}=[p ₁, p ₂... p _{I+x * j}... p _{X * y}], 0＜i≤x, 0＜j≤y,

Wherein, R _{M * n}The two-dimensional matrix of expression pixel spectrum, [p ₁, p ₂... p _{Y * i+j}... p _{X * y}] and [p ₁, p ₂... p _{I+x * j}... p _{X * y}] expression pixel the spectrum vector, m represents the wave band number, n presentation video pixel sum, x presentation video pixel line number, y presentation video pixel columns, n=x * y;

Set up the linear relationship of described target optical spectrum and described testing image pixel spectrum two-dimensional matrix:

S _t＝R _m×n×c _PLS+E _m×1，

Wherein, S _tExpression target optical spectrum vector, R _{M * n}The two-dimensional matrix of expression pixel spectrum, c _PLSThe expression regression coefficient, E _{M * 1}The expression noise vector.

3, as high spectrum sub-pixel target detection method as described in the claim 2, it is characterized in that, describedly three-dimensional high-spectral data is expressed as two-dimensional matrix and sets up between the linear relationship, also comprise: described target optical spectrum and described testing image pixel spectrum two-dimensional matrix are carried out pre-service.

As high spectrum sub-pixel target detection method as described in the claim 3, it is characterized in that 4, described pre-service comprises that orthonormal transformation handles or add scatter correction and handle.

As high spectrum sub-pixel target detection method as described in the claim 4, it is characterized in that 5, described orthonormal transformation is handled and specifically comprised:

$p_{\mathrm{hk},\mathrm{SNV}}=\frac{p_{\mathrm{hk}}-{\stackrel{\&OverBar;}{p}}_h}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{(p_{\mathrm{hk}}-{\stackrel{\&OverBar;}{p}}_h)}^2}}{(m-1)}^{1/2}$

Wherein, p _{Hk, SNF}Expression is handled in the testing image of back the reflectance value of h pixel at k wave band, p through orthogonal transformation _hH pixel is at the mean value of each wave band place reflectivity in the expression testing image, and m represents the wave band number, and m-1 represents degree of freedom;

Described additional scatter correction is handled and is specifically comprised:

Calculate the averaged spectrum vector: $\stackrel{\&OverBar;}{p}=\frac{1}{n}\underset{h=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_h;$

Each pixel spectrum is carried out linear regression: p _h=m _hP+b _h

Add scatter correction: $p_{h\left(\mathrm{MSC}\right)}=\frac{(p_h-b_h)}{m_h};$

Wherein, p represents the averaged spectrum vector,

Expression is to the summation of all pixel spectrum vectors, p _hThe spectrum vector of representing h pixel, m _h, b _hRepresent h pixel spectrum vector p respectively _hWith the slope and the intercept of the linear regression of all pixel averaged spectrum, p _{H (MSC)}Pixel spectrum vector behind the additional scatter correction of expression process.

6, high spectrum sub-pixel target detection method according to claim 1 is characterized in that, obtains the regression coefficient vector of described inversion model by the offset minimum binary alternative manner, specifically comprises:

A, obtain the initial weight vector according to the target optical spectrum vector: $w_n={\mathrm{RS}}_t^T,$ Wherein, S _tExpression target optical spectrum vector, w _nExpression initial weight vector, R represents the two-dimensional matrix of pixel spectrum;

B, according to described initial weight vector calculation score vector: t _n=R ^Tw _n, wherein, t _nThe expression score vector, R represents the two-dimensional matrix of pixel spectrum, w _nExpression initial weight vector;

C, calculate the load vector of described target optical spectrum according to described score vector: q _n=S _tt _n, wherein, q _nThe load vector of representing described target optical spectrum, S _tExpression target optical spectrum vector, t _nThe expression score vector;

D, calculate the load vector of described pixel spectrum two-dimensional matrix according to described score vector: p _n=Rt _n, wherein, p _nRepresent the load vector of described pixel spectrum two-dimensional matrix, R represents the two-dimensional matrix of pixel spectrum, t _nThe expression score vector;

E, obtain described regression coefficient vector: c _PLS=W (P ^TW) ^-1Q ^T, wherein, Q={q ₁, q ₂... q _n, P={p ₁, p ₂... p _n, W={w ₁, w ₂... .w _n;

F, calculating residual sum of squares (RSS): $\mathrm{SS}\left(n\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(S_i-R_0{c}_{\mathrm{PLS}})}^2,$ Wherein, SS (n) represents residual sum of squares (RSS), and m represents wave band number, R ₀Represent original pixel spectrum two-dimensional matrix, S _tThe expression target optical spectrum is to heavy, and the setting threshold values is G, if SS (n-1)-SS (n)≤G then gets the corresponding c of this SS (n) _PLSBe regression coefficient vector; Otherwise, make S _t=S _t-t _nq _n, R=R-t _np _n' return step a repetition aforesaid operations, up to SS (n-1)-SS (n)≤G, obtain regression coefficient vector c then _PLS

7, a kind of high spectrum sub-pixel target detection device is characterized in that, comprising:

The modelling unit is used to set up the inversion model of target optical spectrum and testing image pixel spectrum two-dimensional matrix;

The regression coefficient vector acquiring unit is used to obtain the regression coefficient vector of described inversion model;

The mahalanobis distance acquiring unit is used for obtaining according to described regression coefficient vector the mahalanobis distance of each pixel;

Identifying unit is used to judge that mahalanobis distance is inferior pixel objective point greater than the pairing pixel of the regression coefficient of threshold values.

8, as high spectrum sub-pixel target detection device as described in the claim 7, it is characterized in that described modelling unit comprises:

Matrix generates subelement, is used for the three-dimensional high-spectral data of testing image pixel is expressed as the two-dimensional matrix of high spectrum reflection rate;

Linear relationship is set up subelement, is used to set up the linear relationship of described target optical spectrum and described testing image pixel spectrum two-dimensional matrix.

9, as high spectrum sub-pixel target detection device as described in the claim 8, it is characterized in that, described modelling unit also comprises: the pre-service subelement is used for the testing image pixel spectrum two-dimensional matrix of described target optical spectrum and the generation of described matrix generation subelement is carried out pre-service and offer linear relationship setting up subelement.

10, as high spectrum sub-pixel target detection device as described in each in the claim 7 to 9, it is characterized in that described device also comprises: data capture unit is used to obtain the three-dimensional high-spectral data of testing image pixel and offers the modelling unit.

Images