CN105427291A - Method for detecting vector edges of multispectral remote sensing images - Google Patents

Abstract

The invention discloses a multi-spectral remote sensing image vector edge detection method, comprising the steps of: acquiring a high spatial resolution remote sensing image; performing gradient vector edge detection on the high spatial resolution remote sensing image to obtain a gradient vector feature image; acquiring the gradient vector feature image The angle value of the maximum change or discontinuous change position point and the change rate of the corresponding position point in the angle direction; according to the obtained angle value and change rate, use the non-extreme value suppression method to determine the initial edge point set of the gradient vector feature image; according to The initial edge point set is detected by double-threshold segmentation processing method to obtain the edge contour of the image; for each band of the edge contour, the edge vector is superimposed and synthesized in the multi-dimensional color space, and the edge detection result is output. The invention can present the normalized geographical environment where the remote sensing image edge information extraction project is located by detecting the vector edge of the multispectral remote sensing image, thereby ensuring the authenticity and effectiveness of the experiment.

Description

Vector Edge Detection Method of Multispectral Remote Sensing Image

technical field

The invention relates to a multi-spectral remote sensing image vector edge detection method.

Background technique

Edge detection generally uses a certain discontinuity feature of the target in the image (such as brightness, hue, saturation, etc.) to realize the extraction of edge information, and has a wide range of applications in remote sensing image object extraction. In multi-spectral high-spatial-resolution remote sensing images, due to the different material, orientation, geometric shape and lighting conditions of ground objects, there will be reflection edges, facing edges, occlusion edges, illumination (shadow) edges, and specular (highlight) edges; They usually cause a large number of false edges in the edge extraction results, which seriously affect the subsequent edge-based image analysis and understanding.

The famous edge detection Canny operator is famous for its good edge detection effect, but this operator can only be applied to the edge information extraction of single-channel grayscale images, and the necessary expand. Compared with single-band (single-channel) remote sensing images, multi-band (multi-channel or multi-spectral, generally the number of bands is not less than three) remote sensing images has more abundant spectral information and features that can be used for edge detection. The scheme needs to be determined flexibly according to the category of ground features.

The main problems that exist when traditional edge detection methods (such as Canny operator) are applied to multispectral remote sensing images are:

(1) Restrictions on the number of bands and processing strategies. Traditional edge detection methods can generally only deal with single-band grayscale images (or use the strategy of separately processing multi-band images one by one), and the detection response of each band corresponding to ground objects (generally affected by the difference in spectral range) and The relational features between them are rarely considered in the detection process, which splits the organic connection between the data of each band of the multispectral remote sensing image, resulting in partial loss of edge detection results, and cannot effectively extract edge information for various ground objects.

(2) The suitability of traditional methods for multispectral high spatial resolution remote sensing image data. Multi-spectral high-spatial-resolution remote sensing image data has the following characteristics: a large amount of data, rich information on the geometry and attribute details of ground objects, high spectral heterogeneity within the same ground objects or even the same ground objects, "same object with different spectrum" and "foreign object" The phenomenon of "same spectrum" is common, the spatial structure pattern of ground objects is complex, there are many boundary transition areas, and shadows and small ground objects interfere seriously. These problems lead to the fact that there are usually a large number of false edges in the edge detection results obtained by traditional methods, and sometimes it is impossible to effectively extract and layer (class) analyze and identify the edges of multiple types of objects at the same time, which seriously affects the multispectral based on edge information. Effective development of follow-up work such as remote sensing image analysis and understanding.

(3) The traditional edge detection method lacks the correlation analysis of the edge detection results corresponding to the remote sensing images of each band. Multi-band high-spatial-resolution remote sensing images are differential records of the spectral characteristics of various ground objects in different spectral ranges. The spectral responses of different ground objects in different bands are significantly different. Obviously, each band image will produce different edge detection results. Therefore, the lack of correlation analysis of the edge detection results of each band will lead to the loss of edge information and the decrease of detection accuracy.

Contents of the invention

In view of the above problems, the object of the present invention is to provide a method for edge detection of multi-spectral remote sensing image vectors.

In order to achieve the above object, a kind of multispectral remote sensing image vector edge detection method of the present invention comprises the following steps:

Obtain high spatial resolution remote sensing images with n bands;

Perform gradient vector edge detection on high spatial resolution remote sensing images to obtain gradient vector feature images;

Obtain the angle value of the maximum change or discontinuous change position point on the gradient vector feature image and the change rate of the corresponding position point in the angle direction;

Determine the initial edge point set of the gradient vector feature image by using the non-extreme value suppression method according to the obtained angle value and rate of change;

According to the initial edge point set, the image edge contour is obtained by using the double-threshold segmentation processing method to detect the edge contour;

Each band of the edge contour is superimposed in the multi-dimensional color space to synthesize the edge vector, and the edge detection result is output.

Preferably, the step of using the non-extreme value suppression method to determine the initial edge point set of the gradient vector feature image according to the obtained angle value and the rate of change includes:

Select all the pixel points in the n-dimensional gradient image band by band, and use the non-extreme value suppression method to refine the roof band formed by the pixels with larger gradient values;

If the gradient value on the direction angle θ of the pixel is a local maximum value, it is reserved as a preliminary edge point;

Otherwise, set the pixel as a non-edge point.

Preferably, according to the initial edge point set, the step of utilizing the double-threshold segmentation processing method to perform edge contour detection to obtain the image edge contour includes:

Select two gradient thresholds Y _h and Y _s ;

Remove the pixel points whose gradient value is less than Y _h from the non-extreme value suppression results band by band, and obtain the strong edge point set Q;

Connect the edge points into the initial contour based on Q;

Search the initial contour, and find edge points that can be connected to the current endpoint in the non-extreme value suppression results with gradient values between Y _h and Y _s ;

Use the recursive tracking method to collect edges in the gradient values between Y _h and Y _s until all discontinuities in Y _h are connected.

The beneficial effects of the present invention are:

The present invention can more comprehensively present the normalized geographical environment where the remote sensing image edge information extraction project is located by detecting the vector edge of the multispectral remote sensing image, thereby ensuring the authenticity and effectiveness of this experiment.

Description of drawings

Fig. 1 is the structural block diagram of detection method described in the embodiment of the present invention;

Fig. 2 is the satellite remote sensing image figure of QuickBird high spatial resolution of the embodiment of the present invention;

Fig. 3 is a schematic diagram of the image preprocessing results of the mean shift filter in Fig. 1;

Fig. 4 is the vector edge schematic diagram that the present invention extracts in RGB color space;

Fig. 5 is the vector edge schematic diagram that the present invention extracts in IHS color space;

Fig. 6 is the vector edge schematic diagram that the present invention extracts in YIQ color space;

Fig. 7 is the vector edge schematic diagram that the present invention extracts in YUV color space;

Fig. 8 is the vector edge schematic diagram that the present invention extracts in CIELUV color space;

Fig. 9 is the weighted vector edge schematic diagram that the present invention extracts in CIELUV (L)-YIQ (Y)-IHS (I) color space;

Fig. 10 is a schematic diagram of weighted vector edges extracted in the CIELUV.RGB(LB)-YIQ.RGB(YG)-IHS.RGB(IR) color space according to the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings.

A vector edge detection method of a multi-spectral remote sensing image according to the present invention includes the following steps.

(1) Obtain high spatial resolution remote sensing images of n bands and perform image preprocessing;

Due to its highly detailed spatial information representation ability, high spatial resolution remote sensing image effectively expresses the edge information of the semantic target of the feature, and at the same time its internal geometric detail information also appears in the form of noise (relative to the target size of the feature) , the phenomenon of "scale granularity" is very prominent; and the multispectral color information also shows obvious heterogeneity within the semantic target, and the phenomenon of "same object with different spectrum" is also very prominent, which brings a lot of inconvenience to the edge detection work. For this reason, image preprocessing is performed using mean shift filtering technology (here does not exclude other filtering technologies, such as bilateral filtering, anisotropic diffusion filtering, etc.), in order to achieve the purpose of removing the above noise.

(2) Perform gradient vector edge detection on high spatial resolution remote sensing images to obtain gradient vector feature images;

A multispectral remote sensing image with n bands (channels) can be expressed as G(x, y)=(B ₁ , B ₂ , . . . , B _n ) ^T . The n-dimensional gradient vector T(x, y) at position (x, y) can be expressed as:

$T(x,\mathrm{the\; y})=\left(\begin{array}{cc}\frac{\∂B_1}{\∂x}& \frac{\∂B_1}{\∂\mathrm{the\; y}}\\ \frac{\∂B_2}{\∂x}& \frac{\∂B_2}{\∂\mathrm{the\; y}}\\ \mathrm{...}& \mathrm{...}\\ \frac{\∂B_{\mathrm{no}}}{\∂x}& \frac{\∂B_{\mathrm{no}}}{\∂\mathrm{the\; y}}\end{array}\right)=\left(\begin{array}{c}t\left(B_1\right)\\ t\left(B_2\right)\\ \mathrm{...}\\ t\left(B_{\mathrm{no}}\right)\end{array}\right)$ (Formula 1)

(3) Obtain the angle value of the maximum change or discontinuous change position point on the gradient vector feature image and the change rate of the corresponding position point in the angle direction;

In G(x, y), the direction with the largest change or discontinuity (generally the position of the edge ⁾ can be expressed as:

$V^{T}V=\left(\begin{array}{cc}{u}^{T}u& u^{T}v\\ v^{T}u& v^{T}v\end{array}\right)$ (Formula 2)

Among them, the n-dimensional vectors u and v can be expressed as:

$u={\left(\begin{array}{cccc}\frac{\∂B_1}{\∂x}& \frac{\∂B_2}{\∂x}& \mathrm{...}& \frac{\∂B_{\mathrm{no}}}{\∂x}\end{array}\right)}^T$ (Formula 3)

$v={\left(\begin{array}{cccc}\frac{\∂B_1}{\∂\mathrm{the\; y}}& \frac{\∂B_2}{a\mathrm{the\; y}}& \mathrm{...}& \frac{\∂B_{\mathrm{no}}}{\∂\mathrm{the\; y}}\end{array}\right)}^T$ (Formula 4)

The direction of maximum change or discontinuity in G(x,y) can be expressed by angle θ as:

$\mathrm{\θ}(x,\mathrm{the\; y})=\frac{1}{2}arcta\mathrm{no}\frac{u^{T}v+v^{T}u}{u^{T}u-v^{T}v}$ (Formula 5)

The rate of change Grad(x, y) of point G(x, y) in the θ direction is expressed as:

$Grad(x,\mathrm{the\; y})={\{\frac{1}{2}\[(u^{T}u+v^{T}v)+(u^{T}u-v^{T}v)\cos 2\mathrm{\θ}+(u^{T}v+v^{T}u)\sin 2\mathrm{\θ}\]\}}^{\frac{1}{2}}$ (Formula 6)

(4) Utilize the non-extreme value suppression method to determine the initial edge point set of the gradient vector feature image according to the obtained angle value and rate of change;

Traverse all the pixel points in the n-dimensional gradient image band by band, and use the non-extreme value suppression method to refine the roof band formed by the pixel with a large gradient value; if the gradient value at the direction angle θ of the pixel is the local maximum value, it will be reserved as a preliminary edge point, otherwise it will be set as a non-edge point. Due to the difference in the response of each band to the spectral characteristics of different ground object types, the initial edge points formed by each band generally also show differences.

(5) According to the initial edge point set, the edge contour detection is carried out by using the double-threshold segmentation processing method to obtain the image edge contour;

Two gradient thresholds Y _h and Y _s are selected, generally Y _s =0.4Y _h , and Y _h is obtained by statistical analysis of the "initial edge point set". Firstly, the pixel points whose gradient value is smaller than Y _h are removed from the non-extreme value suppression results band by band, and the strong edge point set Q is obtained. Then, on the basis of Q, the edge points are connected into the initial contour, and there are generally discontinuities on the initial contour. When the contour endpoint is found, the algorithm continues to search for edge points that can be connected to the current endpoint in the non-extreme value suppression results with gradient values between Y _h and _{Y s ;} The edge is continuously collected in the value until all the discontinuities in Y _h are connected, and the discontinuity threshold can be given according to the actual situation.

(6) Overlay and synthesize edge vectors in the multi-dimensional color space for each band of the edge contour, and output the edge detection result.

The Canny operator is improved based on the output fusion strategy. The vector edge in the multi-dimensional color space is obtained from the edge components of each band through feature-level "matching and fusion" in a form similar to superimposed synthesis; the calculation of the edge components of each band needs to be done in a multi-color space, and obtain the edge detection results corresponding to the meaning of the color components.

Let the scalar E _i be the weighted integrated edge component of the vector edge of the remote sensing image, then the vector edge E _v (E _i ) and the weighted integrated scalar edge E _s obtained by superposition and synthesis of all weighted integrated edge components E _i can be expressed as:

E _v (E _i )=(E ₁ E ₂ . . . E _i )i=1, 2, . . . , n; (Formula 7)

β ₁ +β ₂ +...+β _i = 1, β _i ≥ 0, k ≤ n; (Formula 8)

α ₁ +α ₂ +...+α _i =1, α _i ≥ 0, k ≤ n; (Formula 9)

Wherein the edge component B _i is the edge detection result corresponding to each band (or different color space feature components); α _i is the weight corresponding to the edge component B _i , that is, E _i is obtained by weighted synthesis of k band edge components B _i ; β _i is the weight corresponding to the weighted comprehensive edge component when extracting the scalar edge; the edge component E _i contained in E _v can be composed of E _s .

In order to demonstrate the technical effect of this method in practical engineering applications, the QuickBird high-spatial-resolution remote sensing image data is used to demonstrate and explain engineering experiments.

1. Experimental data

As shown in Figure 2, the original experimental data is a 1024×1024 pixel 3-channel (red, green, blue) multi-band QuickBird high-spatial resolution satellite remote sensing image, the data size is 3M; there are roads, buildings, rivers, Trees, bare land, grassland and other ground objects comprehensively present the normalized geographical environment of the remote sensing image edge information extraction project, thus ensuring the authenticity and effectiveness of this experiment.

2. Experimental results

Image preprocessing. Figure 3 below is the result obtained after image preprocessing of the mean shift filter in Figure 2.

The following figures 4-8 are the experimental results obtained by this method for vector edge detection in RGB, IHS, YIQ, YUV, and CIELUV color spaces.

The following figures 9-10 are the experimental results obtained by this method in RGB, IHS, YIQ, YUV, and CIELUV color spaces for weighted vector edge detection.

In Figure 9, the edge detection results corresponding to the color component L in the CIELUV color space, the edge detection results corresponding to the color component Y in the YIQ color space, and the edge detection results corresponding to the color component I in the IHS color space are respectively passed through the feature level The matching fusion of the weighted vector edge LYI is formed, and the weights among the three components are equal to each other when the components L, Y, and I are synthesized.

In Figure 10, the edge detection results corresponding to the color components L and B in the CIELUV and RGB color spaces, the edge detection results corresponding to the color components Y and G in the YIQ and RGB color spaces, and the color components in the IHS and RGB color spaces The edge detection results corresponding to I and R form a weighted vector edge CIELUV.RGB(LB)-YIQ.RGB(YG)-IHS.RGB(IR) through feature-level matching and fusion, and the weights of each component L and B are equal when combined , Y and G are combined with equal weights, I and R are combined with equal weights, and LB, YG, and IR are combined with equal weights.

The above are only preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention are all Should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be defined by the claims.

Claims (3)

1. a multi-spectrum remote sensing image vector edge detection method, is characterized in that, comprises the following steps:

Obtain the high spatial resolution remote sense image with n wave band;

Gradient vector rim detection is carried out to high spatial resolution remote sense image and obtains gradient vector characteristic image;

Obtain angle value and the rate of change of correspondence position point on this angle direction of maximum change on gradient vector characteristic image or discontinuous change location point;

Non-extreme value is utilized to suppress the initial edge point set of method determination gradient vector characteristic image according to the angle value obtained and rate of change;

Utilize Double Thresholding Segmentation facture to carry out edge contour according to initial edge point set and detect acquisition image edge profile;

Each wave band stacked synthesis edge vectors in multi-dimensional color space of edge profile, and export edge detection results.

2. detection method according to claim 1, is characterized in that, utilizes non-extreme value to suppress the step of the initial edge point set of method determination gradient vector characteristic image to comprise according to the angle value obtained and rate of change:

Tie up all pixel points in gradient image by band selection n, suppress method to carry out refinement to the non-extreme value of ridge band that the larger pixel of Grad is formed;

If the Grad on this pixel orientation angle θ is local maximum, be then left preliminary edge point;

Otherwise, this pixel is set to non-edge point.

3. detection method according to claim 1, is characterized in that, comprises according to the step that initial edge point set utilizes Double Thresholding Segmentation facture to carry out edge contour detection acquisition image edge profile:

Selected two Grads threshold Y _hand Y _s;

Suppress to remove Grad in result by wave band in non-extreme value and be less than Y _hpixel point, and obtain strong edge point set Q;

Based on Q, marginal point is connected into initial profile;

Initial profile is searched for, at Grad between Y _hwith Y _snon-extreme value suppress in result, to find the marginal point that can be connected to current endpoint;

Utilize recurrence tracking between Y _hwith Y _sgrad in collect edge, until by Y _hin all discontinuous phases connect.