CN101968885A - Method for detecting remote sensing image change based on edge and grayscale - Google Patents

Abstract

The invention discloses a method for detecting remote sensing image change based on an edge and grayscale, which is used for solving the technical problem of low precision of the conventional remote sensing image change detection method. The technical scheme comprises the following steps of: performing multitemporal image edge characteristic extraction by using a bilateral filtering-based Canny algorithm; performing OTSU threshold segmentation and edge extraction on grayscale interpolation images to obtain grayscale characteristics; and integrating the extracted edge and grayscale characteristics, and detecting change areas of remote sensing images. The linear characteristic of images is fully utilized, and the grayscale interpolation images compensate broken lines caused by registration errors, so the precision of the remote sensing image change detection method is improved to 90.32 percent from 87.75 percent of the prior art, and the detection accuracy is improved.

Description

A kind of method for detecting change of remote sensing image based on edge and gray scale

Technical field

The present invention relates to a kind of method for detecting change of remote sensing image, particularly a kind of method for detecting change of remote sensing image based on edge and gray scale.

Background technology

Document " a kind of new multiband remote sensing image change detecting method, Chinese image graphics journal, 2009, Vol.14 (4), p572-578 " discloses a kind of remote sensing image change detecting method based on fuzzy C-means clustering (FCM) algorithm.Fuzzy C average (FCM) algorithm that this method at first utilizes a kind of improved time complexity to reduce carries out the remote sensing image unsupervised classification to be handled, and introduces the method for multiband comprehensive variation mask then and carries out change-detection.Carry out remote sensing image unsupervised classification processing promptly the carrying out judgement whether each pixel changes, belong to Pixel-level but not the change-detection of feature level, and artificial target comprises road, bridge, airport, railway, house class building etc., is the extremely important linear target of a class.This method is when carrying out change-detection to the remote sensing images that comprise this type of target, because do not consider its linear feature, very easily be subjected to The noise, cause that to detect target imperfect, the overall accuracy of final change-detection only can reach 87.75%, so the document disclosed method has bigger limitation when being applied to handle the remote sensing images that comprise artificial target.

Summary of the invention

In order to overcome the low deficiency of existing method for detecting change of remote sensing image precision, the invention provides a kind of method for detecting change of remote sensing image based on edge and gray scale, this method utilization is carried out the extraction of multidate picture edge characteristic based on the Canny algorithm of bilateral filtering, then the gray scale difference value image is carried out OSTU Threshold Segmentation and edge extracting, obtain gray feature.Again edge and the gray feature that is extracted carried out comprehensively, detect the region of variation of remote sensing images.Owing to when having made full use of the linear characteristic in the image, utilize the gray scale difference value image to remedy the broken string that causes because of registration error, can improve the precision of method for detecting change of remote sensing image.

The technical solution adopted for the present invention to solve the technical problems: a kind of method for detecting change of remote sensing image based on edge and gray scale is characterized in comprising the steps:

(a) establishing the denoising operator is D _h, set up the bilateral filtering model

$D_h\left(x\right)=\underset{\mathrm{\Ω}}{\∫}w(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}(i=\mathrm{1,2})---\left(1\right)$

In the formula, X _TiThe remote sensing images that expression ti sensor obtains when the ti for same position on the ground, X _Ti(y) presentation video X _TiThe gray-scale value at y place, position, Ω represents integral domain;

In the low-pass filtering model

$\mathrm{LPF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(2\right)$

Weight function w (x, y) only relevant with space distribution, handle afterwards in the image, the gray-scale value LPF (x) at each pixel x place does the intensity-weighted computing on the pixel in its small neighbourhood Ω to obtain; In the formula, w _D(x y) is the weight function of weighing pixel x and y proximity relations on how much;

Based on the low-pass filtering model, the bilateral filtering model has increased the weight factor of the gray-scale value degree of closeness of weighing this pixel and its neighborhood interior pixel:

$\mathrm{BF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)w_R(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(3\right)$

In the formula, w _R(x is a weight factor of weighing the close degree of gray-scale value between pixel x and the y y), gets

$w_D(x,y)=\frac{1}{Z_D\left(x\right)}{e}^{-\frac{\left|\right|x-y\left|\right|}{2{\mathrm{\σ}}_D^2}}---\left(4\right)$

$w_R(x,y)=\frac{1}{Z_R\left(x\right)}{e}^{-\frac{{\{X_{\mathrm{ti}}\left(x\right)-X_{\mathrm{ti}}\left(y\right)\}}^2}{2{\mathrm{\σ}}_R^2}}---\left(5\right)$

In the formula, Ω is an integral domain, is taken as 8 and is communicated with neighborhood;

With

Be respectively distance variance and gray variance; Z _D(x) and Z _R(x) be respectively normalized factor, computing method are

The edge of phasor was a benchmark in the past, the edge of unanimity and removal with it in the phasor after the neighborhood search, and the edge result that two sub-pictures are kept merges, and obtains the edge variation testing result after the removal of zonule;

(b) be provided with pixel vector X _T1And X _T2Be respectively the remote sensing images that sensor obtains at t1 and t2 constantly for same position on the ground, the variation of terrain object shows as vectorial X _T1And X _T2Between difference:

δ＝|X _t1-X _t2| (6)

When gray scale＞(the image average * scale-up factor) of δ, think that then the zone of this pixel representative is a region of variation among the current δ, scale-up factor is made as 1.0;

When the gray scale of δ＞Thresh, think that then the zone of this pixel representative is a region of variation among the current δ;

The OTSU threshold value adopts

$\mathrm{Thresh}=\max (\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)*\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)*{(m_1-m_2)}^2)---\left(7\right)$

Calculate; In the formula, hist is that gray scale is the pixel count of δ in the image, the value when promptly gray scale is i in the histogram;

$m_1=\frac{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}$ $m_2=\frac{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}---\left(8\right)$

Utilize morphologic filtering that the gray scale difference value image is carried out rim detection, extract the result as gray feature;

(c) be that benchmark is done the edge tracking with the edge variation testing result, end points for edge that can't be closed, in grey scale change result's outline map, make neighborhood search, adopt " or " logic replenishes edge change-detection result, obtains closed change-detection edge and region of variation.

The invention has the beneficial effects as follows: owing to utilize Canny algorithm to carry out the extraction of multidate picture edge characteristic, then the gray scale difference value image is carried out OSTU Threshold Segmentation and edge extracting, obtain gray feature based on bilateral filtering.Again edge and the gray feature that is extracted carried out comprehensively, detect the region of variation of remote sensing images.Because when having made full use of the linear characteristic in the image, utilize the gray scale difference value image to remedy the broken string that causes because of registration error, the precision of method for detecting change of remote sensing image brings up to 90.32% by 87.75% of background technology, has improved detection accuracy.

Below in conjunction with embodiment the present invention is elaborated.

Embodiment

1, the extraction of edge feature.

Edge detection algorithm among the present invention adopts improved Canny algorithm.The Canny operator can be obtained balance preferably between squelch and rim detection.Traditional Canny operator adopts Gauss filtering to weaken local edge to a certain extent.In the present invention, adopt bilateral filtering (Bilateral filter) rather than Gauss filtering, can better obtain the image border characteristic.The bilateral filtering method is as follows:

Note denoising operator is D _h, Filtering Model can be write as following general type:

$D_h\left(x\right)=\underset{\mathrm{\Ω}}{\∫}w(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}(i=\mathrm{1,2})---\left(1\right)$

In the formula, X _TiThe remote sensing images that expression ti sensor obtains when the ti for same position on the ground, X _Ti(y) presentation video X _TiThe gray-scale value at y place, position, Ω represents integral domain.In traditional low-pass filtering model, weight function w (x, y) only relevant with space distribution, handle afterwards in the image, the gray-scale value LPF (x) at each pixel x place does the intensity-weighted computing on the pixel in its small neighbourhood Ω to obtain, and expression formula is:

$\mathrm{LPF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(2\right)$

In the formula, w _D(x y) is the weight function of weighing pixel x and y proximity relations on how much.On this basis, the bilateral filtering model has increased the weight factor of the gray-scale value degree of closeness of weighing this pixel and its neighborhood interior pixel.Its expression formula is as follows:

$\mathrm{BF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)w_R(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(3\right)$

In the formula, w _R(x is a weight factor of weighing the close degree of gray-scale value between pixel x and the y y), gets usually

$w_D(x,y)=\frac{1}{Z_D\left(x\right)}{e}^{-\frac{\left|\right|x-y\left|\right|}{2{\mathrm{\σ}}_D^2}}---\left(4\right)$

$w_R(x,y)=\frac{1}{Z_R\left(x\right)}{e}^{-\frac{{(X_{\mathrm{ti}}\left(x\right)-X_{\mathrm{ti}}\left(y\right))}^2}{2{\mathrm{\σ}}_R^2}}---\left(5\right)$

In the formula, Ω is an integral domain, be taken as 8 and be communicated with neighborhood,

With

Be respectively distance variance and gray variance, be taken as 1 and 4 here respectively, Z _D(x) and Z _R(x) be respectively normalized factor, computing method are

$Z_R\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{e}^{-\frac{{\{X_{\mathrm{ti}}\left(x\right)-X_{\mathrm{ti}}\left(y\right)\}}^2}{2{\mathrm{\σ}}_R^2}}\mathrm{dy}.$

The bilateral filtering model is in the smoothed image process, both considered the proximity relations on how much, considered the similarity on the gray scale again, according to the distribution situation of neighborhood interior pixel gray-scale value, make the ranking operation of different weights, embodied the anisotropic character of algorithm well.At the boundary of image, common neighborhood convolution algorithm can cause image boundary to thicken.The power that adds the expression grey value difference is because of w _R(x, y) after, at image boundary place gray-scale value a bigger drop is arranged, this moment, weights were less, on one side intensity profile can not have influence on the intensity profile of another side on the border.So just can in smooth noise, can effectively keep the border.

Bilateral filtering model biggest advantage is its non-iteration, this arthmetic statement succinctly is easy to program and realizes, and all reached good denoising effect in a lot of situations, this makes the calculated amount of denoising model that significantly reduction arranged, and this algorithm also has good concurrency.

Size two width of cloth registering images all identical with ratio that the different times areal is taken are made rim detection in the present invention, utilize rim detection extraction marginal information, relatively and detect the linear feature of variation; Simultaneously two width of cloth difference gray level images are done the profile that change-detection obtains region of variation; The linear feature of detection and the contour feature of region of variation are comprehensively obtained final result of variations.

2, the extraction of gray feature.

The processing image of image difference δ as change-detection, understanding theoretically is the most intuitively, simultaneously it for parallel processing not simultaneously the change-detection between mutually also be very effective.When using this method, must consider that will change pixel and no change pixel of selected threshold makes a distinction, being chosen in the change detection algorithm based on gray scale difference value of threshold value is extremely important.The present invention adopts big Tianjin method (OTSU) algorithm to carry out adaptive threshold and selects.

Be provided with pixel vector X _T1And X _T2Be respectively sensor for the obtainable remote sensing images when t1 and the t2 of same position on the ground.Do not consider other factors, as the influence of atmospheric conditions, sensor, promptly in ideal conditions, the variation of terrain object shows as vectorial X _T1And X _T2Between difference:

δ＝|X _t1-X _t2| (6)

Usually, t1 and t2 are identical or the close moon or the date in not the same year, to get rid of variation that season, equal factor caused etc. as far as possible.

Algorithmic descriptions is as follows: the difference with image before and after changing is discussed, and when gray scale＞(the image average * scale-up factor) of δ, thinks that then the zone that this pixel is represented among the current δ is a region of variation, and scale-up factor is made as 1.0.

The OTSU method judges that the δ region of variation is described as follows: when the gray scale of δ＞Thresh, think that then the zone of this pixel representative is a region of variation among the current δ.The threshold value calculation method of OTSU is:

$\mathrm{Thresh}=\max (\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)*\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right){(m_1-m_2)}^2)---\left(7\right)$

In the formula, hist is that gray scale is the pixel count of δ in the image, the value when promptly gray scale is i in the histogram.

$m_1=\frac{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}$ $m_2=\frac{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}---\left(8\right)$

Utilize morphologic filtering that the gray scale difference value image is carried out rim detection, extract the result as gray feature.

3, the fusion of edge and gray feature.

Edge change-detection result is carried out the edge follow the tracks of, scanning can't closed edge end points, and method is as follows: scan a point, if its on every side the difference of 8 all adjacent two points of point absolute value and be 2 * 255, then it is an end points.

With the edge end points is benchmark, does neighborhood search in grey scale change result's outline map, adopt " or " logic replenishes edge change-detection result, obtains closed change-detection edge and region of variation.

After tested, the precision that the inventive method detects Remote Sensing Imagery Change has reached 90.32%, has improved detection accuracy.

Claims (1)

1. the method for detecting change of remote sensing image based on edge and gray scale is characterized in that comprising the steps:

(a) establishing the denoising operator is D _h, set up the bilateral filtering model

$D_h\left(x\right)=\underset{\mathrm{\Ω}}{\∫}w(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}(i=\mathrm{1,2})---\left(1\right)$

In the formula, X _TiThe remote sensing images that expression ti sensor obtains when the ti for same position on the ground, X _Ti(y) presentation video X _TiThe gray-scale value at y place, position, Ω represents integral domain;

In the low-pass filtering model

$\mathrm{LPF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(2\right)$

Weight function w (x, y) only relevant with space distribution, handle afterwards in the image, the gray-scale value LPF (x) at each pixel x place does the intensity-weighted computing on the pixel in its small neighbourhood Ω to obtain; In the formula, w _D(x y) is the weight function of weighing pixel x and y proximity relations on how much;

Based on the low-pass filtering model, the bilateral filtering model has increased the weight factor of the gray-scale value degree of closeness of weighing this pixel and its neighborhood interior pixel:

$\mathrm{BF}\left(x\right)=\underset{\mathrm{\Ω}}{\∫}{w}_D(x,y)w_R(x,y)X_{\mathrm{ti}}\left(y\right)\mathrm{dy}---\left(3\right)$

In the formula, w _R(x is a weight factor of weighing the close degree of gray-scale value between pixel x and the y y), gets

$w_D(x,y)=\frac{1}{Z_D\left(x\right)}{e}^{-\frac{\left|\right|x-y\left|\right|}{2{\mathrm{\σ}}_D^2}}---\left(4\right)$

$w_R(x,y)=\frac{1}{Z_R\left(x\right)}{e}^{-\frac{{(X_{\mathrm{ti}}\left(x\right)-X_{\mathrm{ti}}\left(y\right))}^2}{2{\mathrm{\σ}}_R^2}}---\left(5\right)$

In the formula, Ω is an integral domain, is taken as 8 and is communicated with neighborhood;

With

Be respectively distance variance and gray variance; Z _D(x) and Z _R(x) be respectively normalized factor, computing method are

The edge of phasor was a benchmark in the past, the edge of unanimity and removal with it in the phasor after the neighborhood search, and the edge result that two sub-pictures are kept merges, and obtains the edge variation testing result after the removal of zonule;

(b) be provided with pixel vector X _T1And X _T2Be respectively the remote sensing images that sensor obtains at t1 and t2 constantly for same position on the ground, the variation of terrain object shows as vectorial X _T1And X _T2Between difference:

δ＝|X _t1-X _t2| (6)

When gray scale＞(the image average * scale-up factor) of δ, think that then the zone of this pixel representative is a region of variation among the current δ, scale-up factor is made as 1.0;

When the gray scale of δ＞Thresh, think that then the zone of this pixel representative is a region of variation among the current δ;

The OTSU threshold value adopts

$\mathrm{Thresh}=\max (\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)*\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right){(m_1-m_2)}^2)---\left(7\right)$

Calculate; In the formula, hist is that gray scale is the pixel count of δ in the image, the value when promptly gray scale is i in the histogram;

$m_1=\frac{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=1}{\overset{\mathrm{thresh}}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}$ $m_2=\frac{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}i*\mathrm{hist}\left(i\right)}{\underset{i=\mathrm{thresh}+1}{\overset{255}{\mathrm{\Σ}}}\mathrm{hist}\left(i\right)}---\left(8\right)$

Utilize morphologic filtering that the gray scale difference value image is carried out rim detection, extract the result as gray feature;

(c) be that benchmark is done the edge tracking with the edge variation testing result, end points for edge that can't be closed, in grey scale change result's outline map, make neighborhood search, adopt " or " logic replenishes edge change-detection result, obtains closed change-detection edge and region of variation.