CN102567735B - Method for automatically picking up control point sections of remote sensing images - Google Patents

Abstract

The invention discloses a method for automatically picking up control point sections of remote sensing images, which relates to the digital remote sensing image processing technology. As for a big remote sensing image, textural features (sum probability distribution variance of autocorrelation coefficient and gray level co-occurrence matrix) of the images are used, image sheets which meet requirements of a control point (ground control point (GCP), ground control point) section and have big textures are screened and judged according to provincial characteristics of the image sheets, so that a process of automatically picking up the control point sections is achieved. The method applies image information expressed by the textural features of the images, and effective and automatic picking-up of the control point sections of the remote sensing images is achieved through an automatic algorithm.

Description

The method of a kind of automatic extraction remote sensing images reference mark section

Technical field

The present invention relates to technical field of remote sensing image processing, is the method for a kind of automatic extraction remote sensing images reference mark section, be based on image coefficient of autocorrelation and gray level co-occurrence matrixes with the probability distribution variance.

Background technology

Along with the continuous increase at the rail remote sensing satellite, the remote sensing images quantity of obtaining also grows with each passing day, and utilizes the reference mark section of remote sensing images to carry out images match, and the image information storage becomes the important foundation of remote sensing image processing.Because the reference mark section need comprise significantly terrestrial object information, generally adopts artificial or semi-automatic mode to extract.

When carrying out the extraction of remote sensing images reference mark section, consider the effect of reference mark section, mainly be to provide reference picture for processing such as images match, need to comprise significantly terrestrial object information in its image information, road for example, road junction, the parking lot, the lake, airport etc. often need image to have bigger textural characteristics.The extracting method of the reference mark of the remote sensing images that generally adopt section at present, majority is based on artificial or semi-automatic, and efficient is not high and workload is huge.And when the amount of images increase was very fast, the extraction of can not in time cutting into slices was so real-time is poor.

Summary of the invention

The method that the purpose of this invention is to provide the section of a kind of automatic extraction remote sensing images reference mark, section artificial extraction in remote sensing images reference mark is wasted time and energy and the problem of real-time difference to solve.

For achieving the above object, technical solution of the present invention is:

The method of a kind of automatic extraction remote sensing images reference mark section, it comprises:

Step 1: extract the candidate image section automatically, and carry out normalized;

Step 2: the textural characteristics value of calculated candidate image slices, comprise coefficient of autocorrelation and based on the gradation of image co-occurrence matrix with the probability variance;

Step 3: according to the image feature value size that obtains, criterion be in five coefficient of autocorrelation any one greater than 0.25, and on this basis, require the gradation of image co-occurrence matrix with the probability variance greater than 1, judge whether have than the large texture feature, whether meet image reference mark slicing characteristics, determine satisfactory image slices.

The method of described automatic extraction remote sensing images reference mark section, its described step 1 is that traversal is extracted image slices from original remote sensing images, cuts into slices as the candidate; For piece image section f, the normalized formula is:

f(i，j)＝(f(i，j)-μ)/σ

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (symbol/expression removes for i, gradation of image value j), and μ is a gradation of image value average, and computing formula is:

$\mathrm{\μ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), the symbol ∑ represent the summation, symbol * expression product;

σ is a gradation of image value standard deviation, and computing formula is:

$\mathrm{\σ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{(f(i,j)-\mathrm{\μ})}^2}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], (i, j) coordinate is that (the symbol ∑ is represented summation for i, gradation of image value j), and symbol * expression product, symbol subscript 2 are represented to count square to f in the presentation video.

The method of described automatic extraction remote sensing images reference mark section, the eigenwert of its described step 2 calculated candidate image slices is:

A) five of the calculated candidate image slices coefficient of autocorrelation, for piece image section f, its coefficient of autocorrelation computing formula is:

$\mathrm{\ρ}(\mathrm{\Δx},\mathrm{\Δy})=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)\×f(i+\mathrm{\Δx},j+\mathrm{\Δy})}{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{f}^2(i,j)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) coordinate is (i, gradation of image value j), Δ x in the presentation video, Δ y is an image displacement in the x and y direction, and the symbol ∑ is represented summation, and symbol * expression product, symbol subscript 2 are represented to count square; Wherein the displacement l of Cai Yonging is 10, chooses 4 direction θ ₁, θ ₂, θ ₃, θ ₄, value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve the displacement Δ x on the four direction respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates four coefficient of autocorrelation ρ ₁, ρ ₂, ρ ₃, ρ ₄, then four coefficient of autocorrelation are averaged, obtain the 5th coefficient of autocorrelation ρ ₅

B) the calculated candidate image slices based on gray level co-occurrence matrixes with the probability variance, at first ask for the gray level co-occurrence matrixes of image f, computing formula is:

Wherein, (i, j) be image f at x, the coordinate figure of y both direction, coordinate range be [0 ..., N-1], f (i, j) coordinate is (i, gradation of image value j) in the presentation video, Δ x, Δ y be image at x, the displacement on the y both direction, the symbol ∑ is represented summation, p, coordinate (i among the q difference presentation video f, j) and the gray-scale value located of coordinate (i+ Δ x, j+ Δ y), also be the coordinate of gray level co-occurrence matrixes C; Wherein, at first image is carried out the gray level compression, the gray level gl of employing is 16, and compress mode is:

f(i，j)＝f(i，j)/256×(gl-1)

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), symbol * expression product, symbol/expression removes;

The displacement l that adopts is 6, and chooses 4 direction θ ₁, θ ₂, θ ₃, θ ₄, value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve four direction top offset amount Δ x respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates the gray level co-occurrence matrixes on all directions, calculates itself and probability distribution p then _I+j(t), computing formula is:

$p_{i+j}\left(t\right)=\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (t is and probability (i+j) for i, gradation of image value j), its scope be [0 ..., 2*N-2], the symbol ∑ is represented summation;

According to four direction with probability distribution p _I+j(t), calculate its entropy respectively, computing formula is:

$\mathrm{entropy}=-\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{p}_{i+j}\left(t\right)\×\log \left\{p_{i+j}\left(t\right)\right\}$

Wherein log represents to ask natural logarithm, and t is and probability (i+j), its scope be [0 ..., 2*N-2], ∑ is represented summation, symbol * expression product; According to the entropy that obtains, calculate and variance of probability distribution, computing formula is again:

$t_{\mathrm{sv}}=\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{(t-\mathrm{entropy})}^2\×p_{i+j}\left(t\right)$

T is and probability (i+j), its scope be [0 ..., 2*N-2], symbol * expression product, symbol ∑ are represented summation, 2 expressions of symbol subscript count square;

At last, with asking on average of the four direction that calculates, obtain eigenwert T with the probability distribution variance _Sv

The method of described automatic extraction remote sensing images reference mark section, its described step 3 is that the image feature value to step 2 gained carries out criterion: at first, for 5 ρ ₁, ρ ₂, ρ ₃, ρ ₄, ρ ₅Value as long as one of them value greater than 0.25, can enter next round and differentiate, is next judged the feature T of image slices _Sv, require greater than 1, greater than 1 promptly as the image slices of qualified reference mark (GCP).

The inventive method has been considered the textural characteristics of image essence, used image texture characteristic in analyzing coefficient of autocorrelation and based on the statistical nature of gray level co-occurrence matrixes, judge according to the image coefficient of autocorrelation that calculates with based on the size with the variance of probability distribution value of gray level co-occurrence matrixes, to obtain satisfactory reference mark (GCP) section.

The auto slice extractive technique of the inventive method has improved having now based on artificial and automanual extractive technique, has improved work efficiency and has guaranteed the real-time of handling.

Description of drawings

Fig. 1 is the method flow synoptic diagram of a kind of automatic extraction remote sensing images of the present invention reference mark section;

Fig. 2 is got the image synoptic diagram by the first step in the inventive method, second step; Wherein:

Fig. 2 A1 is the candidate image section synoptic diagram of section A;

Fig. 2 A2 is the normalized result schematic diagram of section A;

Fig. 2 B1 is the candidate image section synoptic diagram of section B;

Fig. 2 B2 is the normalized result schematic diagram of section B.

Embodiment

With reference to Fig. 1, the method that the section of a kind of remote sensing images of the present invention reference mark is extracted automatically, the specific implementation process is as follows:

Step 1: extract the candidate image section automatically, and carry out normalized; At first the traversing graph picture extracts image slices automatically, for image slices, tries to achieve its average μ earlier, and computing formula is:

$\mathrm{\μ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), ∑ represent the summation, * expression product.

With σ be gradation of image value standard deviation, computing formula is:

$\mathrm{\σ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{(f(i,j)-\mathrm{\μ})}^2}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], (i, j) coordinate is that (symbol * expression product, symbol ∑ are represented summation for i, gradation of image value j), and symbol subscript 2 expression counts square to f in the presentation video.Carry out normalized then, computing formula is:

$f(i,j)=\frac{f(i,j)-\mathrm{\μ}}{\mathrm{\σ}}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], (i, j) coordinate is (i, gradation of image value j) to f in the presentation video.

Step 2: the textural characteristics value of calculated candidate image slices, comprise coefficient of autocorrelation and based on the gradation of image co-occurrence matrix with the probability variance;

Wherein in the calculating of coefficient of autocorrelation, the displacement l of employing is 10, chooses 4 direction θ 1, and θ 2, and θ 3, and θ 4, and value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve the displacement Δ x on the four direction respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates four coefficient of autocorrelation ρ ₁, ρ ₂, ρ ₃, ρ ₄, then four coefficient of autocorrelation are averaged, obtain the 5th coefficient of autocorrelation ρ ₅,

In the calculating with the probability variance of gray level co-occurrence matrixes, at first image slices is carried out the gray level compression, the gray level gl that adopts in this method is 16, and compress mode is:

$f(i,j)=\frac{f(i,j)}{256}\×(\mathrm{gl}-1)$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], (i, j) coordinate is (i, gradation of image value j), symbol * expression product to f in the presentation video.

The displacement l that adopts is 6, and chooses 4 direction θ ₁, θ ₂, θ ₃, θ ₄, value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve four direction top offset amount Δ x respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates the gray level co-occurrence matrixes on all directions, calculates itself and probability distribution p then _I+j(t), computing formula is:

$p_{i+j}\left(t\right)=\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), t represent with probability distribution in the probable value of i+j, its scope be [0 ..., 2*N-2], the symbol ∑ is represented summation.

According to four direction with probability distribution p _I+j(t), calculate its entropy respectively, computing formula is:

$\mathrm{entropy}=-\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{p}_{i+j}\left(t\right)\log \left\{p_{i+j}\left(t\right)\right\}$

Wherein log represents to ask natural logarithm, and t is and probability (i+j), its scope be [0 ..., 2*N-2], the symbol ∑ is represented summation, symbol * expression product.

According to the entropy that obtains, calculate and variance of probability distribution, computing formula is again:

$t_{\mathrm{sv}}=\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{(t-\mathrm{entropy})}^2{p}_{i+j}\left(t\right)$

T is and probability (i+j), its scope be [0 ..., 2*N-2], symbol * expression product, symbol ∑ are represented summation, 2 expressions of symbol subscript count square.

At last, with asking on average of the four direction that calculates with the probability distribution variance, obtain gray level co-occurrence matrixes with probability variance eigenwert T _Sv

Step 3: according to the image feature value size that obtains, criterion be in five coefficient of autocorrelation any one greater than 0.25, and on this basis, require the gradation of image co-occurrence matrix with the probability variance greater than 1, judge whether have than the large texture feature, whether meet image reference mark slicing characteristics, determine satisfactory image slices.Referring to shown in Figure 2, wherein, Fig. 2 A1 is the candidate image section synoptic diagram of section A, and Fig. 2 A2 is the normalized result schematic diagram of section A; Fig. 2 B1 is the candidate image section synoptic diagram of section B, and Fig. 2 B2 is the normalized result schematic diagram of section B.Fig. 2 A2 and Fig. 2 B2 make its gradation of image value be applicable to subsequent calculations and differentiation through normalized.

Illustrate as following table 1:

Table 1

Auto slice extractive technique of the present invention is the picture material of expressing according to the texture information of image slices, and image sheet is effectively screened, and finally obtains the reference mark sectioning image that is fit to.Both guaranteed the validity of the reference mark sectioning image of extraction, promptly had significantly terrestrial object information, also improved the efficient of the work of extracting greatly, with reference to following table 2.

Table 2

Claims (3)

1. a method of extracting the section of remote sensing images reference mark automatically is characterized in that, comprising:

Step 1: extract the candidate image section automatically, and carry out normalized;

Step 2: the textural characteristics value of calculated candidate image slices, comprise image five coefficient of autocorrelation and based on the gradation of image co-occurrence matrix with the probability variance;

Step 3:, judge whether have than the large texture feature, whether meet image reference mark slicing characteristics, determine satisfactory image slices according to the image feature value size that obtains; Wherein in five coefficient of autocorrelation of candidate section any one greater than 0.25, and the gradation of image co-occurrence matrix with the probability variance greater than 1, think that then this section has than the large texture feature, meet image reference mark slicing characteristics, be satisfactory image slices;

Wherein, the textural characteristics value of described step 2 calculated candidate image slices is:

A) five of the calculated candidate image slices coefficient of autocorrelation, for piece image section f, its coefficient of autocorrelation computing formula is:

$\mathrm{\ρ}(\mathrm{\Δx},\mathrm{\Δy})=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)\×f(i+\mathrm{\Δx},j+\mathrm{\Δy})}{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{f}^2(i,j)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) coordinate is (i, gradation of image value j), Δ x in the presentation video, Δ y is an image displacement in the x and y direction, and the symbol ∑ is represented summation, and symbol * expression product, symbol subscript 2 are represented to count square; Wherein the displacement l of Cai Yonging is 10, chooses 4 direction θ ₁, θ ₂, θ ₃, θ ₄, value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve the displacement Δ x on the four direction respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates four coefficient of autocorrelation ρ ₁, ρ ₂, ρ ₃, ρ ₄, then four coefficient of autocorrelation are averaged, obtain the 5th coefficient of autocorrelation ρ ₅

B) the calculated candidate image slices based on gray level co-occurrence matrixes with the probability variance, at first ask for the gray level co-occurrence matrixes of image f, computing formula is:

Wherein, (i, j) be image f at x, the coordinate figure of y both direction, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), Δ x, Δ y be image at x, the displacement on the y both direction, symbol ∑ represent the summation; P, q respectively coordinate among the presentation video f (i j) and the gray-scale value located of coordinate (i+ Δ x, j+ Δ y), also is the coordinate of gray level co-occurrence matrixes C; Wherein, at first image is carried out the gray level compression, the gray level gl of employing is 16, and compress mode is:

f(i，j)＝f(i，j)/256×(gl-1)

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), symbol * expression product, symbol/expression removes;

The displacement l that adopts is 6, and chooses 4 direction θ ₁, θ ₂, θ ₃, θ ₄, value is respectively: 0 °, and 45 °, 90 °, 135 °, try to achieve four direction top offset amount Δ x respectively, Δ y, computing formula is:

Δx＝l×cosθ _m

Δy＝l×sinθ _m

M=1 wherein, 2,3,4, symbol * expression product calculates the gray level co-occurrence matrixes on all directions, calculates itself and probability distribution p then _I+j(t), computing formula is:

$p_{i+j}\left(t\right)=\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (t is and probability (i+j) for i, gradation of image value j), its scope be [0 ..., 2*N-2], the symbol ∑ is represented summation;

According to four direction with probability distribution p _I+j(t), calculate its entropy respectively, computing formula is:

$\mathrm{entropy}=-\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{p}_{i+j}\left(t\right)\×\log \left\{p_{i+j}\left(t\right)\right\}$

Wherein log represents to ask natural logarithm, and t is and probability (i+j), its scope be [0 ..., 2*N-2], ∑ is represented summation, symbol * expression product; According to the entropy that obtains, calculate and variance of probability distribution, computing formula is again:

$t_{\mathrm{sv}}=\underset{t=0}{\overset{2N-2}{\mathrm{\Σ}}}{(t-\mathrm{entropy})}^2\×p_{i+j}\left(t\right)$

T is and probability (i+j), its scope be [0 ..., 2*N-2], symbol * expression product, symbol ∑ are represented summation, 2 expressions of symbol subscript count square;

At last, with asking on average of the four direction that calculates, obtain eigenwert T with the probability distribution variance _Sv

2. the method for automatic extraction remote sensing images as claimed in claim 1 reference mark section is characterized in that, described step 1 is that traversal is extracted image slices from original remote sensing images, cuts into slices as the candidate; For piece image section f, the normalized formula is:

f(i，j)＝(f(i，j)-μ)/σ

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (symbol/expression removes for i, gradation of image value j), and μ is a gradation of image value average, and computing formula is:

$\mathrm{\μ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}f(i,j)}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], f (i, j) in the presentation video coordinate be (i, gradation of image value j), the symbol ∑ represent the summation, symbol * expression product;

σ is a gradation of image value standard deviation, and computing formula is:

$\mathrm{\σ}=\frac{\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{N-1}{\mathrm{\Σ}}}{(f(i,j)-\mathrm{\μ})}^2}{(N-1)\×(N-1)}$

Wherein (i, j) presentation video pixel coordinate, coordinate range be [0 ..., N-1], (i, j) coordinate is that (the symbol ∑ is represented summation for i, gradation of image value j), and symbol * expression product, symbol subscript 2 are represented to count square to f in the presentation video.

3. the method for automatic extraction remote sensing images as claimed in claim 1 reference mark section is characterized in that, described step 3 is that the image feature value to step 2 gained carries out criterion: at first, and for 5 ρ ₁, ρ ₂, ρ ₃, ρ ₄, ρ ₅Value as long as one of them value greater than 0.25, can enter next round and differentiate, is next judged the feature T of image slices _Sv, require greater than 1, greater than 1 promptly as the image slices at qualified reference mark.

Images