CN104268891A - Registration method for eight-channel imaging multispectral image - Google Patents

Abstract

The invention discloses a registration method for an eight-channel imaging multispectral image. The registration method comprises the following steps that firstly, edge detection and extraction are carried out on the multispectral image; secondly, registration and transformation are carried out on the multispectral image. According to the registration method for the eight-channel imaging multispectral image, due to the fact that an eight-channel imaging spectrometer is small in number of channels, and processing time is not limited, the method of extracting image feature points in an operating interference mode for matching is adopted, and the good registered composite image can be obtained.

Description

Eight passage imaging multi-spectrum image registration methods

Technical field

The present invention relates to a kind of eight passage imaging multi-spectrum image registration methods, adopt a kind of method operating intervention extraction image characteristic point and carry out mating, be applicable to the feature that 8 passage imaging spectrometers have the factors such as number of active lanes is less, the processing time is unrestricted.

Background technology

Multispectral imaging (Multispectral Imaging) is the abbreviation of multiband light spectral imaging technology, it is acquisition of information and the treatment technology of the unification of a kind of collection of illustrative plates, namely, while obtaining the space dimension information of target, the spectrum dimension information of this target is also obtained.According to the spectral band number of institute's obtaining information or the difference of spectral resolution size, current spectral imaging technology roughly can be divided into multispectral imaging (Multi-spectral Imaging), high light spectrum image-forming (Hyper-spectral Imaging) and Hyper spectral Imaging (Ultra-spectral Imaging).

Image registration (Alignment) refers to process or the method for two width images of the Same Scene taking from different time, different sensors or different visual angles or multiple image sequence being carried out to geometric position coupling, be a popular invention of Image Engineering application in recent years, be widely used at present in the numerous areas such as remote sensing image, medical image, three-dimensionalreconstruction, robot vision.

The registration of multispectral image is one of important step of whole spectrum imaging system data processing, and it is related to the accurate fusion problem of spectrum peacekeeping image dimension.In the data cube of multi-spectral imager, due to the motion of the target travel that exists in the rigging error of optical system and sampling process or instrument itself, cause the difference that there is the aspect such as geometric position and magnification between the multichannel image of Same Scene, therefore employing image registration link is necessary at the Data processing of multi-spectral imager, to make the image of data cube at space dimension accurate registration.

Up to now, image registration progressively defines theoretical system and the method for complete set.In conventional images method for registering, mainly comprise cross-correlation method, Fourier transform, point mapping, elastic model method, wavelet transformation registration method etc., these methods have respective features and application background.

Cross-correlation method is the method based on gradation of image information, and it is suitable for the registration between the image of same sensor acquisition.The calculated amount that the method requires is large, is not suitable for process non-linear deformation and local deformation problems, need to improve.

Fourier transform is after carrying out fast fourier transform to image, and application phase such as to be correlated with at the image registration of technical finesse rotation, Pan and Zoom mismatch.But Fourier transform method can not process the registration of the problems such as nonlinear deformation and different gray scale attributed graph picture.

It is the method for registering the most often adopted when not knowing the mapping mode of two width images that point maps.But the positional precision of unique point easily by the impact of the subjective judgement of people, can not obtain precise and stable registration result, therefore, some mapping method also often utilizes the feedback between each stage to find optimal transformation.

Elastic model method is mainly used in the registration between medical image at present.Method for registering based on wavelet transformation causes the great attention of people in recent years, is characterized in calculated amount when can reduce image registration greatly.

Accompanying drawing explanation

Fig. 1 is that Laplacian operator commonly uses template schematic diagram;

Fig. 2 is LOG operator 5 × 5 rank template schematic diagram.

Summary of the invention

In view of above-mentioned the deficiencies in the prior art part, the object of the present invention is to provide a kind of method for registering being applicable to eight passage imaging multispectral images, specifically take following technical scheme:

Step one, multispectral image rim detection and extraction;

Step 2, multi-spectral image registration convert.

Preferably, above-mentioned steps is specially:

Step one, multispectral image rim detection and extraction, LOG wave filter is adopted to carry out rim detection, image is first carried out suitable level and smooth, with restraint speckle, and then carry out asking micro-, the basis of Laplacian operator adds the LOG operator of Gauss conversion and realization, for two dimensional image signal, first comes smoothing with following Gauss function:

$G(x,y,\mathrm{\σ})=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

G (x, y, σ) is a function with circular symmetry, and its level and smooth effect controls by σ, owing to carrying out linear smoothing to image, is mathematically carry out convolution, make g (x, y) be level and smooth after image, obtain:

g(x，y)＝G(x，y，σ)*f(x，y)

Wherein f (x, y) is level and smooth front image,

Because marginal point is the place that in image, gray-value variation is violent, the sudden change of this image intensity will produce a peak in first order derivative, or be equivalent to a generation zero cross point in second derivative, and be nonlinear along the second derivative of gradient direction, calculate comparatively complicated, so substitute with Laplacian operator, namely use:

${\▿}^{2}g(x,y)={\▿}^2(G(x,y)*f(x,y))=\left({\▿}^{2}G(x,y,\mathrm{\σ})\right)*f(x,y)$

Zero cross point as marginal point, in formula for LOG wave filter,

${\▿}^{2}G(x,y,\mathrm{\σ})=\frac{{\∂}^{2}G}{{\∂x}^2}+\frac{{\∂}^{2}G}{{\∂y}^2}=\frac{1}{{\mathrm{\π\σ}}^4}(\frac{x^2+y^2}{2{\mathrm{\σ}}^2}-1)\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

Get σ >=1;

Step 2, multi-spectral image registration convert, and transformation for mula is

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=k\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}\mathrm{\Δx}\\ \mathrm{\Δy}\end{array}\right]$

In formula: (x, y) be the point of piece image and benchmark image, (x ', y ') after conversion in corresponding second width image, and k, θ and Δ x and Δ y are the scale factor of the first width figure and the second width figure, twiddle factor and coordinate translation amount respectively.

Preferably, above-mentioned LOG operator realizes by template, adopts 5 × 5 templates as shown in Table

0	0	-1	0	0
					0	-1	-2	-1	0
-1	-2	16	-2	-1
					0	-1	-2	-1	0
0	0	-1	0	0

Compared to prior art, eight passage imaging multi-spectrum image registration methods provided by the invention have the factors such as number of active lanes is less, the processing time is unrestricted for 8 passage imaging spectrometers, adopt a kind of method operating intervention extraction image characteristic point and carry out mating, good registration composite diagram can be obtained.

Embodiment

The invention provides a kind of eight passage imaging multi-spectrum image registration methods, for making object of the present invention, technical scheme and effect clearly, clearly, developing simultaneously referring to accompanying drawing, the present invention is described in more detail for embodiment.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Eight passage imaging spectrometers have the factors such as number of active lanes is less, the processing time is unrestricted, and therefore image registration of the present invention adopts a kind of manual intervention extraction image characteristic point to carry out the method for mating, and the method can classify as point mapping.

Point mapping needs to extract the edge details of each width image so that extraction to unique point or characteristic matching vector before carrying out registration.Method for detecting image edge at present existing ripe method can utilize, and can be divided into two large classes substantially at present, i.e. first order differential operator and Second Order Differential Operator by the edge detection operator that people adopt.First order differential operator mainly contains gradient operator, Roberts cross operator, Prewitt operator and Sobel operator; Second Order Differential Operator mainly refers to Laplacian (Laplce) operator.These operators are all in response to gray level change, or average gray level change.

If added Gauss (Gauss) conversion before second-order differential Laplacian operator, then form famous LOG operator and Gauss-Laplace operator.Above-mentioned edge detection operator respectively has its feature.

The edge detection operator that the present invention adopts is LOG operator, and it on the basis of Laplacian operator, adds Gauss conversion and realizes, and is therefore necessary first to introduce Laplacian operator.

Laplacian operator is a kind of second derivative operator, and to 1 continuous function f (x, y), it is defined as follows in the Laplacian value of position (x, y):

${\mathrm{\Δ}}^{2}f=\frac{{\∂}^{2}f}{{\∂x}^2}+\frac{{\∂}^{2}f}{{\∂y}^2}$

Derive:

$\frac{{\∂}^{2}f}{{\∂x}^2}=\frac{\∂\left(G_x\right)}{\∂x}=\frac{\∂\left(f\right[i,j+1]-f[i,j\left]\right)}{\∂x}=f[i,j+2]-2f[i,j+1]+f[i,j]$

$\frac{{\∂}^{2}f}{{\∂y}^2}=\frac{\∂\left(G_y\right)}{\∂y}=\frac{\∂\left(f\right[i+1,j]-f[i,j\left]\right)}{\∂y}=f[i+1,j+1]-2f[i,j+1]+f[i-1,j+1]$

Above approximate expression is centered by [i, j+1], replaces with j-1:

$\frac{{\∂}^{2}f}{{\∂x}^2}=f[i,j+1]-2f[i,j]+f[i,j-1]$

$\frac{{\∂}^{2}f}{{\∂y}^2}=f[i+1,j]-2f[i,j]+f[i-1,j]$

In the handling procedure of digital picture, the Laplacian value of computing function also can realize by various template.The coefficient being corresponding center pixel to the basic demand of template should be positive, and the coefficient of corresponding center adjacent pixel should be negative, and they and should be zero.Two kinds of conventional Laplacian operator templates as shown in Figure 1.

Laplacian operator is a kind of second derivative operator, so quite responsive to the noise in image.It often produces the wide edge of double image element in addition, and can not provide the information of edge direction.Due to above reason, Laplacian operator is seldom directly used in Edge detected, and determines after being mainly used in known edge pixel that this pixel is in the dark space of image or area pellucida.

For Laplacian operator, after second differential, the boundary pixel extracted is not on original border.Because second differential operator does not have directivity, also cannot revise location distortion.Therefore, locating distortion is Laplacian operator Problems existing.

First image registration of the present invention adopt LOG wave filter to carry out edge extracting.

In the various method of Edge extraction, because the larger change of gray scale always corresponds to some larger derivatives, so the edge detection operator proposed the earliest is gradient operator and Laplacian operator.Gradient method computing is simple, software simulating is convenient, but produce wider response in its region near border, the result of gained usually needs refinement in addition, this not only have impact on the positioning precision on border, and the quality on border can be affected, and Laplacian operator is high frequency sensitivity, so larger by the impact of high frequency noise.In order to effectively suppress the impact of high frequency noise, a kind of improving one's methods first is carried out suitable level and smooth, and with restraint speckle, and then carry out asking micro-, namely LOG filter skirt detects.

For two dimensional image signal, first come smoothing with following Gauss function:

$G(x,y,\mathrm{\σ})=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

G (x, y, σ) is a function with circular symmetry, its level and smooth effect controls by σ, owing to carrying out linear smoothing to image, is mathematically carry out convolution, make g (x, y) be level and smooth after image, obtain: g (x, y)=G (x, y, σ) * f (x, y), wherein f (x, y) is level and smooth front image.

Because marginal point is the place that in image, gray-value variation is violent, the sudden change of this image intensity will produce a peak in first order derivative, or be equivalent to a generation zero cross point in second derivative, and be nonlinear along the second derivative of gradient direction, calculate comparatively complicated, so substitute with Laplacian operator, namely use:

${\▿}^{2}g(x,y)={\▿}^2(G(x,y)*f(x,y))=\left({\▿}^{2}G(x,y,\mathrm{\σ})\right)*f(x,y)$

Zero cross point as marginal point, in formula for LOG wave filter.

${\▿}^{2}G(x,y,\mathrm{\σ})=\frac{{\∂}^{2}G}{{\∂x}^2}+\frac{{\∂}^{2}G}{{\∂y}^2}=\frac{1}{{\mathrm{\π\σ}}^4}(\frac{x^2+y^2}{2{\mathrm{\σ}}^2}-1)\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

Above formula is exactly LOG edge detection operator.LOG operator is Mexico's straw hat shape, being similar to ganglia retinae receptive field spatial organization, can regard as and be made up of focus of excitation district and an inhibition surrounding zone, usually occur with the template form of 62 σ × 62 σ sizes, when σ gets different values, then the image border under available operators detection different scale.Usually, we get σ >=1.

The feature of LOG wave filter: the Gauss function part G in this wave filter can image blur, effectively eliminates all yardsticks and changes much smaller than the image intensity of Gauss distribution space constant σ.Why selecting Gauss function to carry out blurred picture because of it is all level and smooth, localization in spatial domain and frequency domain inside, and the possibility therefore introducing any change do not occurred in original image is minimum.

After the edge extracting completing image, namely can carry out image conversion according to unique point, thus realize registration.Piece image is aimed at another piece image, often need carry out a series of conversion to piece image.

Mainly there is rotational differential and translational difference between each width image, but enlargement factor difference is very little, therefore selects following methods to carry out registration transformation, transformation for mula is as follows:

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=k\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}\mathrm{\Δx}\\ \mathrm{\Δy}\end{array}\right]$

In formula: (x, y) be the point of piece image and benchmark image, (x ', y ') after conversion in corresponding second width image, and k, θ and Δ x and Δ y are the scale factor of the first width figure and the second width figure, twiddle factor and coordinate translation amount respectively; These parameters are obtained by concrete operations.

Eight passage imaging multi-spectrum image registration methods provided by the invention have the factors such as number of active lanes is less, the processing time is unrestricted for 8 passage imaging spectrometers, adopt a kind of method operating intervention extraction image characteristic point and carry out mating, good registration composite diagram can be obtained.

Be understandable that, for those of ordinary skills, can be equal to according to technical scheme of the present invention and inventive concept thereof and replace or change, and all these change or replace the protection domain that all should belong to the claim appended by the present invention.

Claims (3)

1. eight passage imaging multi-spectrum image registration methods, is characterized in that comprising the following steps:

Step one, multispectral image rim detection and extraction;

Step 2, multi-spectral image registration convert.

2. eight passage imaging multi-spectrum image registration methods, is characterized in that comprising the following steps:

Step one, multispectral image rim detection and extraction, LOG wave filter is adopted to carry out rim detection, image is first carried out suitable level and smooth, with restraint speckle, and then carry out asking micro-, the basis of Laplacian operator adds the LOG operator of Gauss conversion and realization, for two dimensional image signal, first comes smoothing with following Gauss function:

$G(x,y,\mathrm{\σ})=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

G (x, y, σ) is a function with circular symmetry, and its level and smooth effect controls by σ, owing to carrying out linear smoothing to image, is mathematically carry out convolution, make g (x, y) be level and smooth after image, obtain:

g(x，y)＝G(x，y，σ)*f(x，y)

Wherein f (x, y) is level and smooth front image,

Because marginal point is the place that in image, gray-value variation is violent, the sudden change of this image intensity will produce a peak in first order derivative, or be equivalent to a generation zero cross point in second derivative, and be nonlinear along the second derivative of gradient direction, calculate comparatively complicated, so substitute with Laplacian operator, namely use:

${\▿}^{2}g(x,y)={\▿}^2(G(x,y)*f(x,y))=({\▿}^{2}G(x,y,\mathrm{\σ}))*f(x,y)$

Zero cross point as marginal point, in formula for LOG wave filter,

${\▿}^{2}G(x,y,\mathrm{\σ})=\frac{{\∂}^{2}G}{\∂x^2}+\frac{{\∂}^{2}G}{\∂y^2}=\frac{1}{\mathrm{\π}{\mathrm{\σ}}^4}(\frac{x^2+y^2}{2{\mathrm{\σ}}^2}-1)\exp (-\frac{1}{2{\mathrm{\σ}}^2}(x^2+y^2))$

Get σ >=1;

Step 2, multi-spectral image registration convert, and transformation for mula is

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=k\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]+\left[\begin{array}{c}\mathrm{\Δx}\\ \mathrm{\Δy}\end{array}\right]$

In formula: (x, y) be the point of piece image and benchmark image, (x ', y ') after conversion in corresponding second width image, and k, θ and Δ x and Δ y are the scale factor of the first width figure and the second width figure, twiddle factor and coordinate translation amount respectively.

3. eight passage imaging multi-spectrum image registration methods as claimed in claim 2, is characterized in that: described LOG operator realizes by template, adopt 5 × 5 templates as shown in Table

0 0 -1 0 0 0 -1 -2 -1 0 -1 -2 16 -2 -1 0 -1 -2 -1 0 0 0 -1 0 0