CN103247029B - A kind of high spectrum image geometrical registration method generated for spliced detector - Google Patents

Abstract

The invention discloses a kind of high spectrum image geometrical registration method generated for spliced detector, it is applied to ' Pin '-shaped row detector, the image geometry misalignment that its arrangement mode causes.First its method finds the long and narrow ground object target on high spectrum image, the edge extracting that centroid position carries out Target scalar is calculated by asking, respectively linear fit is done to the edge of Target scalar in dislocation two edge image, relatively the matching of dislocation two edge image is biased, thus obtains the sub-picture dot number of dislocation image.According to result, resampling is carried out to dislocation image and realize geometry correction.Can detected with high accuracy dislocation value and do not introduce other errors by geometric dislocation method for registering of the present invention, be a key breakthrough in this field, method is reliable and with practical value.The remote sensing images geometrical registration that simultaneously can be detector of the same type provides reference frame.

Description

A kind of high spectrum image geometrical registration method generated for spliced detector

Technical field

The present invention relates to image processing field, specifically refer to a kind of geometrical registration method of the high spectrum image generated for spliced detector, it is for the geometric dislocation phenomenon in correcting image.

Background technology

In order to obtain comparatively Large visual angle cartographic feature, need to increase detector Spatial Dimension.By index demands such as instrument ground resolution, fabric widths, and the restriction of infrared mercury-cadmium tellurid detector design technology, the hyperspectral imager newly developed is in short-wave infrared (1000nm-2500nm) scope, focal plane arrangement employing four pieces of detector array apportion two row stagger the design proposal of splicing, and the detector array imaging adopting two light path systems to carry out line by line, the picture that four pieces of detector arrays are formed the most at last splices, and sees Fig. 1.But be similar to polynary arrangement due to detector array focal plane, and laterally having certain dislocation, show the inconsistent phenomenon for adjacent detector arrays scanning strip on image.Therefore need after imaging to carry out the correction of many field of view (fov) registration, to guarantee that the geometric positioning accuracy of image can meet design and devdlop requirement.

Traditional geometrical registration method is roughly divided into based on frequency field, is correlated with based on gray scale and the method for feature based.Method based on frequency field utilizes Fourier transform by two image conversions to be matched to frequency domain, then utilizes cross-power spectrum directly to calculate translation vector between two width images, but many atural object multi information of image makes the error that calculates larger; The method relevant based on gray scale is searching two width image lap, and utilize the similarity of gray level in the color system corresponding to them for the registration position of criterion searching image, this needs the overlapping region between two width images with larger area; Relatively more conventional method is the method for feature based, and specific algorithm, for extract feature set from image subject to registration, utilizes the corresponding relation of feature to mate.But the overlapping region between the different detector array of the hyperspectral imager of this research is less than 10 pixels, and feature difficulty is extracted, and registration accuracy is low, is not suitable for use and is correlated with based on gray scale or the method for feature based, very large matching error can be caused.Therefore considering that the linear fit bias difference by calculating dislocation image neutral line atural object obtains the sub-pixed mapping number of dislocation, according to trying to achieve result, geometrical registration being completed to the method that dislocation image carries out resampling.

Summary of the invention

Based on above-mentioned prior art Problems existing, the object of the invention is the method for registering proposing a kind of geometric space image offset, to revise the non-matching problem of image space caused due to optical-mechanical system and detector array restriction.

First this matching process finds special objective atural object, then carry out statistic according to centroid method and calculate atural object boundary information, treat registering images after carrying out edge extracting and carry out edge linear fit, thus calculate the bias difference of two fitting a straight lines, be the sub-pixed mapping value of skew, finally according to side-play amount, cubic spline resampling carried out to realize correcting to image.

Its concrete steps are:

1) high spectrum image is screened.Carry out artificial visual interpretation screening to the high spectrum image that hyperspectral imager collects, select these long and narrow ground object targets of river, bridge, highway, wheatland edge or shore line, selected image should be the image covering the splicing of different detector.

2) edge extracting is carried out to ground object target.Concrete grammar is:

A) the record fixing stitching position of detector and the point of interface of long and narrow Target scalar are cut-point (x ₀, y ₀), with cut-point (x ₀, y ₀) to select respectively to misplace the ordinate value y of image the right and left long and narrow atural object for benchmark _j, j=0,1,2...m, select ordinate point number >8, to ensure the precision of linear fit.

B) abscissa value x ' is calculated.Suppose that stitching image both sides gray-scale value average is respectively for step a) in each y _j, at borderline region and the differentiation of non-critical, transition pixel can be there is and gray-scale value between with between.Defining the edge detected is that pixel is split into two gray-scale values the part that area is consistent, finds frontier point and namely asks pixel gray-scale value to be time abscissa positions x ', see Fig. 2.If sampled point is square pixel, move horizontally (y from left to right _j, j=0,1,2...m, x _i, i=0,1,2...n), select gray-scale value closest to 1/2 pixel, if its gray-scale value is DN _ij, calculate its weights W _ijbe distributed as:

$W_{\mathrm{ij}}=\frac{{\mathrm{DN}}_{\mathrm{ij}}-\stackrel{\‾}{N}}{\stackrel{\‾}{M}-\stackrel{\‾}{N}}---\left(6\right)$

x’=x _i-(1-W _ij)（7）

In formula: represent the gray-scale value average of border the right and left respectively; x _irepresent y _jcorresponding abscissa value; DN _ijfor x ' represents the abscissa value of the frontier point that edge extracting goes out.

3) simple linear regression analysis is carried out to edge feature, solve expression formula.When after the one group of coordinate points obtaining dislocation image both sides, set up Linear Regression Model in One Unknown and solve linear representation, specific algorithm is: first set up Linear Regression Model in One Unknown expression formula:

y _t＝a+b*x _t+e _t(8)

In formula: a and b is undetermined parameter; x _tand y _trepresent each transverse and longitudinal coordinate points respectively; T=1,2 ..., n is the subscript of each group of observation data; e _tfor stochastic variable.

Note with be respectively the match value of parameter a and b, then Linear Regression Model in One Unknown is:

$\hat{y}=\hat{a}+\hat{b}*x---\left(9\right)$

Wherein for the valuation of y, for the difference of actual observed value and valuation, the least square fitting principles and requirements error e of parameter a and b _tquadratic sum reach minimum, that is:

$E=\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{e^2}_t=\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{(y_t-{\hat{y}}_t)}^2=\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{(y_t-a-{\mathrm{bx}}_t)}^2$

Obtain minimum value

According to the necessary condition of extreme value, have:

$\frac{\∂E}{\∂a}=-2\underset{t=1}{\overset{n}{\mathrm{\Σ}}}(y_t-a-{\mathrm{bx}}_t)=0---\left(10\right)$

$\frac{\∂E}{\∂b}=-2\underset{t=1}{\overset{n}{\mathrm{\Σ}}}(y_t-a-{\mathrm{bx}}_t)x_t=0---\left(11\right)$

Solving equations, asks parameter a, the match value of b for:

$\hat{a}=\stackrel{\‾}{y}-\hat{b}*\stackrel{\‾}{x}---\left(12\right)$

$\hat{b}=\frac{\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{x}_t{y}_t-\frac{1}{n}\left(\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{x}_t\right)\left(\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{y}_t\right)}{\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{x^2}_t-\frac{1}{n}{\left(\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{x}_t\right)}^2}---\left(13\right)$

In formula be respectively transverse and longitudinal coordinate x _tand y _taverage, t=1,2 ..., n.

4) according to fitting result computational geometry dislocation sub-pixed mapping number.Obtain the marginal information of the atural object of image subject to registration according to step 3) respectively, consider that the atural object of dislocation two edge image is same atural object, therefore the linear equation slope that obtains of matching is identical, and two equations are respectively y ₁=a ₁+ b*x and y ₂=a ₂+ b*x, wherein a ₁, a ₂represent linear equation to be respectively biased, b represents gain, i.e. slope.The sub-pixed mapping number then misplaced is Δ y=a ₁-a ₂.

5) to dislocation image resampling to realize geometrical registration.According to 4) middle result, Δ y represents the relative right image magnitude of misalignment of left image, carries out cubic spline interpolation, finally obtain the image after geometrical registration to right image.

The present invention has following beneficial effect:

1) propose a kind of new geometrical registration method, especially for the image to be matched that overlapping region is little, significantly improve registration accuracy.

2) applicability of method is higher, is not only applicable to high spectrum image, and the image for other types is also suitable for, with practical value.

3) this method is first by edge extracting computed image side-play amount, then according to side-play amount resampling image, allows can ignore in orientation owing to calculating the error introduced in precision.

Accompanying drawing explanation

Fig. 1 is hyperspectral imager short-wave infrared detector array schematic diagram.

Fig. 2 is for asking frontier point horizontal ordinate schematic diagram.

The image original subject to registration of Fig. 3 emulation.

Fig. 4 invention process flow diagram.

Embodiment

According to the present invention, geometrical registration is carried out to the image that a width space dimension misplaces.Consideration emulating image can investigate the registration accuracy of the method more accurately, and the detector array mode according to Fig. 1 emulates the effect image after sampling.Wherein designed image size is single band 800*1024, wherein about space dimension 512 place location drawing picture, there is dislocation, and the gray-scale value on boundary both sides is respectively 255 and 0, and design sketch is shown in Fig. 3.

Details are as follows:

1) determine that in image, misalignment position is (511,562), arrange the ordinate y of different points to be calculated in the region, left and right of image, the ordinate value that wherein left region is determined is 506,513,520,527,534,538,543,546,550; Right area ordinate is 570,574,578,581,585,588,595,602,613.

2) calculate the x respective value of boundary corresponding to each y value above according to formula (1) and (2), the results are shown in following table:

Left region abscissa value x	Left region ordinate value y	Right region abscissa value x	Right region ordinate value y
				720.2549	506	812.028	570
730.2549	513	817.349	574
				740.2549	520	823.1765	578
750.2549	527	827.349	581
				760.2549	534	833.1765	585
766.1137	538	837.349	588
				773.1765	543	847.349	595
777.349	546	857.349	602
				783.1765	550	873.1765	613

3) according to the marginal point coordinate that step 2 detects, do simple linear regression analysis respectively, obtaining expression formula is respectively:

y1=1.42518*x-0.77718

y2=1.42518*x-0.57934

Calculate Δ y=0.77718-0.57934=0.19784, and the dislocation value of the original image of simulation is 0.2 pixel, its relative error calculated is

$E=\frac{0.2-0.19784}{0.2}*100\%=1.08\%$

4) the image offset amount calculated according to step 3) cuts out rear resampling to right image, to obtain the image after geometrical registration.

Claims (1)

1. a geometrical registration method for the high spectrum image generated for spliced detector, is characterized in that the following steps comprised:

1) screen high spectrum image, from the high spectrum image that detector generates, select these long and narrow ground object targets of river, bridge, highway, wheatland edge or shore line, selected image should be the image covering the splicing of different detector;

2) carry out edge extracting to ground object target, concrete grammar is:

A) the record fixing stitching position of detector and the point of interface of long and narrow Target scalar are cut-point (x ₀, y ₀), with cut-point (x ₀, y ₀) to select respectively to misplace the ordinate value y of image the right and left long and narrow atural object for benchmark _j, j=0,1,2...m, select ordinate point number >8, to ensure the precision of linear fit;

B) calculate abscissa value x ', suppose that stitching image both sides gray-scale value average is respectively for step a) in each y _j, at borderline region and the differentiation of non-critical, transition pixel can be there is and gray-scale value between with between, defining the edge detected is that pixel is split into two gray-scale values the part that area is consistent, finds frontier point and namely asks pixel gray-scale value to be time abscissa positions x ', if sampled point is square pixel, move horizontally x from left to right _i, i=0,1,2...n, select gray-scale value closest pixel, if its gray-scale value is DN _ij, calculate its weights W _ijbe distributed as:

$W_{ij}=\frac{{\mathrm{DN}}_{ij}-\stackrel{\‾}{N}}{\stackrel{\‾}{M}-\stackrel{\‾}{N}}---\left(1\right)$

x’＝x _i-(1-W _ij)(2)

In formula: represent the gray-scale value average of border the right and left respectively; x _irepresent y _jcorresponding abscissa value; X ' represents the abscissa value of the frontier point that edge extracting goes out;

3) simple linear regression analysis is carried out to edge feature, solves expression formula, when after the one group of transverse and longitudinal coordinate points obtaining dislocation image both sides, set up Linear Regression Model in One Unknown and solve linear representation:

y _i＝a+b*x _i+e _i(3)

In formula: a and b is undetermined parameter; x _iand y _irepresent each transverse and longitudinal coordinate points respectively; I=1,2 ..., n is the subscript of each group of observation data; e _ifor stochastic variable;

According to least square fitting principles and requirements error e _iquadratic sum reach minimum, solving equation group, asks parameter a, the match value of b for:

$\hat{a}=\stackrel{\‾}{y}-\hat{b}*\stackrel{\‾}{x}---\left(4\right)$

$\hat{b}=\frac{{\mathrm{\Σ}}_{i=1}^n{x}_i{y}_i-\frac{1}{n}\left({\mathrm{\Σ}}_{i=1}^n{x}_i\right)\left({\mathrm{\Σ}}_{i=1}^n{y}_i\right)}{{\mathrm{\Σ}}_{i=1}^n{x_i}^2-\frac{1}{n}{\left({\mathrm{\Σ}}_{i=1}^n{x}_i\right)}^2}---\left(5\right)$

In formula be respectively transverse and longitudinal coordinate x _iand y _iaverage, i=1,2 ..., n;

4) according to fitting result computational geometry dislocation sub-pixed mapping number, according to step 3) obtain the marginal information of the atural object of image subject to registration respectively, consider that the atural object of dislocation two edge image is same atural object, therefore the linear equation slope that obtains of matching is identical, and two equations are respectively y ₁=a ₁+ b*x and y ₂=a ₂+ b*x, wherein a ₁, a ₂represent linear equation to be respectively biased, b represents gain, i.e. slope, then the sub-pixed mapping number misplaced is △ y=a ₁-a ₂;

5) to dislocation image resampling to realize geometrical registration, according to step 4) in result, Δ y represents the relative right image magnitude of misalignment of left image, carries out cubic spline interpolation, finally obtain the image after geometrical registration to right image.