CN106934805A - SAR image superpixel segmentation method based on Gamma filtering - Google Patents

Abstract

The present invention proposes a kind of SAR image superpixel segmentation method based on Gamma filtering, the technical problem low for solving super-pixel segmentation result precision present in the existing superpixel segmentation method based on filtering, realizes that step is：The original SAR image being input into is filtered using mean filter, obtains the average value and standard deviation of pixel grey scale；The variation coefficient of the coefficient of variation, the coefficient of variation of coherent speckle noise and SAR image of pixel grey scale average value is calculated respectively；Judge the relation of each coefficient, it is determined whether carry out Gamma filtering；Coherent speckle noise noise reduction is carried out to the original SAR image being input into using Gamma filtering methods；Super-pixel segmentation is carried out to the SAR image after noise reduction, multiple super-pixel block is obtained and is exported.Present invention reduces the influence of coherent speckle noise in SAR image, the degree of accuracy of SAR image super-pixel segmentation is improve, can be used for the target detection to SAR image, identification and classify.

Description

SAR image super-pixel segmentation method based on Gamma filtering

Technical Field

The invention belongs to the technical field of image processing, relates to an SAR image superpixel segmentation method, and particularly relates to an SAR image superpixel segmentation method based on Gamma filtering, which can be used for target detection, identification and classification of an SAR image.

Background

Synthetic Aperture Radar (SAR) is a coherent imaging radar working in the microwave band, which becomes an important means for current remote sensing observation with high resolution and all-weather, all-day, large-area data acquisition capability, and is widely applied in the aspects of resources, environment, archaeology, military and the like. And forming an SAR image by backscattering of the electromagnetic wave transmitted by the radar by the ground target. With the massive acquisition of SAR images, intelligent image understanding and interpretation technology becomes a research hotspot today. As a key task for image understanding and interpretation, superpixel segmentation of an image is to divide a pair of input SAR images into image blocks by adjacent pixels according to different attributes, contents, features, and the like. The method has the advantages that the super-pixel segmentation is effectively and accurately carried out on an image, the accuracy of further understanding the image can be improved, more effective information can be obtained, and the complexity of a subsequent SAR image processing task is reduced to a great extent. However, the inherent coherent imaging mechanism of the SAR image causes speckle noise which cannot be eliminated in the SAR image, which causes the gray value of each pixel to change randomly, thereby affecting the super-pixel segmentation effect of the SAR image and the subsequent processing result of the SAR image.

Currently, most scholars mainly adopt the following three methods to solve the above problems: 1. a statistical model-based superpixel segmentation method; 2. a scale-based superpixel segmentation method. 3. A super-pixel segmentation method based on filtering.

The first method is to perform probability statistics on speckle noise in the SAR image, and to fuse a probability statistical model into a model of a conventional superpixel segmentation algorithm under the assumption that the probability distribution conforms to Rayleigh distribution, Gamma distribution, K distribution and the like, so as to obtain a superpixel segmentation result suitable for the SAR image.

The core idea of the second method is to regard the SAR image as textures with different scales, and to present different textures according to the property that different types of targets present different textures, so as to segment the superpixel of the SAR image and convert the segmentation into the identification of different textures in the SAR image. However, the realization process of the method is complex, and compared with other super-pixel segmentation methods, the obtained result has no obvious advantages.

The third method is that firstly, a filter is adopted to filter speckle noise in the SAR image, the influence of the speckle noise on SAR image processing is reduced, and therefore the SAR image after noise reduction is obtained, and then the SAR image after noise reduction is subjected to super-pixel segmentation. The method is simple to implement, and the super-pixel segmentation accuracy obtained under the condition that the complexity is smaller than that of the former two methods is the same. In such methods, the key is to filter speckle noise from the SAR image, and the technique is well-established, and includes using a median filter, a local filter, a Lee filter, a Sigma filter, a Frost filter, etc. to preprocess the SAR image to remove speckle noise. However, when the filters filter speckle noise of the SAR image, although the speckle noise can be filtered, edge information, texture information, linear characteristic information, and the like in the image cannot be well maintained.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, provides an SAR image superpixel segmentation method based on Gamma filtering, effectively maintains the edge information, texture information and linear characteristic information of an image under the condition of reducing the interference of speckle noise in an SAR image, and is used for solving the technical problem of low superpixel segmentation accuracy in the conventional SAR image superpixel segmentation method based on filtering.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:

(1) inputting an original SAR image;

(2) filtering the input original SAR image by adopting mean filtering to obtain the mean value of the pixel gray level of the original SAR imageAnd the standard deviation σ (i, j);

(3) using average value of pixel gray scale of original SAR imageAnd standard deviation sigma (i, j), calculating the average value of the gray scale of the pixels of the original SAR imageCoefficient of variance C_I：

(4) Calculating variance coefficient C of speckle noise in original SAR image_u：

Wherein L represents the view of the original SAR image;

(5) calculating the change coefficient C of the original SAR image_max:

(6) When C is present_u≤C_I≤C_maxThen, a Gamma filtering method is adopted to carry out speckle noise reduction on the input original SAR image to obtain a SAR image I after noise reduction_Gam(i,j)：

Wherein α is a heterogeneous parameter, and

(7) for SAR image I after noise reduction_Gam(i, j) performing superpixel division to obtain and output a plurality of superpixel blocks.

Compared with the prior art, the invention has the following advantages:

according to the invention, before the SAR image is subjected to the superpixel segmentation, the speckle noise in the original SAR image is filtered by adopting Gamma filtering, the SAR image after noise reduction is obtained, and then the superpixel segmentation is carried out on the SAR image.

Drawings

FIG. 1 is a block diagram of an implementation flow of the present invention;

FIG. 2 is a superpixel image generated by simulating an SAR image according to the present invention;

FIG. 3 is a superpixel image generated from a simulated SAR image containing texture in accordance with the present invention;

fig. 4 is a superpixel image generated from a real SAR image of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Example 1

Referring to fig. 1: the invention comprises the following steps:

step 1: an original SAR image is input.

The original SAR image is divided into two categories, namely a simulated SAR image and a real SAR image, and in the embodiment, the original SAR image is a simulated SAR image without texture.

Step 2: filtering the input original SAR image by adopting mean filtering to obtain the mean value of the pixel gray level of the original SAR imageAnd the standard deviation σ (i, j);

the mean filtering is a typical linear filtering algorithm, and when performing the mean filtering, a filtering template is needed, and the original pixel value is replaced by the average value of all pixels in the template. The selection of the filtering template generally includes 3 × 3 templates, 5 × 5 templates, or a larger window template, and the selection of the template in the present invention is not limited, but since the larger the window is, the influence may be on the filtered result, in the embodiment of the present invention, a weight coefficient matrix template of 3 × 3 is selected, and its expression is as follows:

wherein s is a row of the template matrix expression, and t is a column of the template matrix expression;

(2a) filtering the input original SAR image by adopting a weight coefficient matrix template w (s, t) to obtain the average value of the pixel gray level of the original SAR image

(2b) Calculating a standard deviation sigma (i, j) of the pixel gray level of the original SAR image:

wherein, N represents the number of pixels on the abscissa of the input original SAR image, and M represents the number of pixels on the ordinate of the input original SAR image.

And step 3: using average value of pixel gray scale of original SAR imageAnd standard deviation sigma (i, j), calculating the average value of the gray scale of the pixels of the original SAR imageCoefficient of variance C_I：

And 4, step 4: calculating variance coefficient C of speckle noise in original SAR image_u：

Wherein L represents the view of the original SAR image;

and 5: calculating the change coefficient C of the original SAR image_max:

Step 6: variance coefficient C for judging pixel gray level average value of original SAR image_ICoefficient of variance C with speckle noise in original SAR image_uAnd the original SAR image change coefficient C_maxWhen C is a relationship of_u≤C_I≤C_maxThen, a Gamma filtering method is adopted to carry out speckle noise reduction on the input original SAR image to obtain a SAR image I after noise reduction_Gam(i,j)：

Wherein α is a heterogeneous parameter, and

when C is present_I＜C_uThen the pixel gray value of the SAR image is the average value of the pixel gray values of the original SAR image

When C is present_I＞C_maxAnd if so, the pixel gray value of the SAR image is the pixel gray value I (I, j) of the original SAR image. For both casesIt is not Gamma filtered.

And 7: for SAR image I after noise reduction_Gam(i, j) performing superpixel division to obtain and output a plurality of superpixel blocks. The specific implementation of this step is as follows:

(7a) calculating SAR image I after noise reduction_Gam(i, j) difference W (i, j) of pixel values of adjacent pixels in the horizontal direction:

W(i,j)＝|I_Gam(i,j)-I_Gam(i-1,j)|

wherein, I_Gam(I-1, j) represents the pixel value I_Gam(i, j) pixel values of horizontally adjacent pixels;

(7b) calculating SAR image I after noise reduction_Gam(i, j) difference V (i, j) in pixel values of adjacent pixels in the vertical direction:

V(i,j)＝|I_Gam(i,j)-I_Gam(i,j-1)|

wherein, I_Gam(I, j-1) represents a pixel value I_Gam(i, j) pixel values of vertically adjacent pixels;

(7c) calculating SAR image I after noise reduction_Gam(i, j) the gradient value G (i, j) of (i, j) is calculated by the formula:

G(i,j)＝W(i,j)+V(i,j)

(7d) setting SAR image I after noise reduction_Gam(i, j) the number K of superpixel blocks, and calculating the average distance d of the superpixel initial seed points:

wherein S is SAR image I after noise reduction_Gam(ii) total area of (i, j), and S ═ MN;

(7e) for SAR image I_GamDisturbing the position of the super-pixel initial seed point on the (I, j) edge to deviate from the SAR image I_Gam(I, j) the edge, i.e. the gradient value at this initial seed point, is smaller than the SAR image I_Gam(i, j) gradient values G (i, j) at the edges, obtaining superpixel seed points;

(7f) and (4) performing boundary growth on the super-pixel seed points, and stopping the boundary growth until the adjacent super-pixel seed points expand to the boundary and collide, so as to obtain a plurality of super-pixel blocks and output the super-pixel blocks.

Example 2

In embodiment 2, as in the other steps of embodiment 1, only the type of the original SAR image used in step 1 is adjusted, and the original SAR image of this embodiment uses a simulated SAR image containing texture.

Example 3

In embodiment 3, the same as other steps in embodiment 1, only the type of the original SAR image used in step 1 is adjusted, and the original SAR image in this embodiment uses a real SAR image.

The effect of the present invention will be further described below with reference to simulation experiments.

1. Simulation conditions are as follows:

the simulation of the invention is carried out under the Intel (R) core (TM)2Duo of the main frequency 2.4GHZ, the hardware environment of the memory 4GB and the software environment of MATLAB R2015 a. In the invention, the adopted SAR images are all SAR images containing 1-view speckle noise, the generated superpixel block is set to be 2500, and the SAR images are subjected to superpixel segmentation, wherein the number of the superpixel blocks is usually set to be between 1000 and 5000.

2. Simulation content and result analysis:

simulation one, the present invention is used to generate superpixel segmentation on a simulated SAR image, resulting in the superpixel segmentation result shown in fig. 2 (c).

With reference to figure 2 of the drawings,

fig. 2(a) is an input simulated SAR image, with dimensions of 512 × 512 pixels; fig. 2(b) is a denoised SAR image obtained by Gamma filtering an input simulated SAR image, and it can be seen from fig. 2(b) that speckle noise is filtered out to a great extent and the edge of the image is effectively maintained after Gamma filtering; fig. 2(c) illustrates the super-pixel image generated from the denoised SAR image, and it can be seen from fig. 2(c) that the super-pixel segmentation obtained by Gamma filtering effectively maintains the edge engagement degree, so that the accuracy of the super-pixel segmentation is improved.

Simulation two, the invention is used for generating the superpixel segmentation on the simulated SAR image containing texture, and the superpixel segmentation result shown in fig. 3(c) is generated.

With reference to figure 3 of the drawings,

FIG. 3(a) is an input texture-containing simulated SAR image with dimensions of 512 × 512 pixels; fig. 3(b) is a denoised SAR image obtained by Gamma filtering an input simulated SAR image containing texture, and it can be seen from fig. 3(b) that speckle noise is filtered out to a great extent after Gamma filtering, and edge and texture detail information of the image is effectively maintained; fig. 3(c) illustrates the super-pixel image generated from the denoised SAR image, and it can be seen from fig. 3(c) that the super-pixel segmentation obtained by Gamma filtering effectively maintains the edge information and texture detail information of the image, so that the accuracy of the super-pixel segmentation is improved.

And thirdly, generating superpixel segmentation on the real SAR image by using the method, and generating a superpixel segmentation result as shown in fig. 4 (c).

With reference to figure 4 of the drawings,

FIG. 4(a) is an input real SAR image from a real time processed airborne SAR in the United states Sangia national Laboratories (Sandia national Laboratories) showing the China Lake airport area of California, USA with a resolution of 3 meters and a size of 522 x 466 pixels; fig. 4(b) is a denoised SAR image obtained by Gamma filtering an input real SAR image, and it can be seen from fig. 4(b) that speckle noise is filtered out to a great extent after Gamma filtering, and edge information, texture detail information and linear feature information of the image are effectively maintained; fig. 4(c) illustrates the super-pixel image generated from the SAR image after noise reduction, and it can be seen from fig. 4(c) that the super-pixel segmentation obtained by Gamma filtering effectively maintains the edge information, texture detail information, and linear feature information of the image, so that the accuracy of the super-pixel segmentation is also improved.

The simulation experiment results show that the method can effectively keep the edge of the image, and can also effectively keep the linear characteristic and the texture detail information when the SAR image contains the linear characteristic and the texture information, so that the accuracy of the superpixel segmentation result is improved.

Claims (3)

1. A SAR image super-pixel segmentation method based on Gamma filtering comprises the following steps:

(1) inputting an original SAR image;

(2) filtering the input original SAR image by adopting mean filtering to obtain the mean value of the pixel gray level of the original SAR imageAnd the standard deviation σ (i, j);

(3) using average value of pixel gray scale of original SAR imageAnd standard deviation sigma (i, j), calculating the average value of the gray scale of the pixels of the original SAR imageCoefficient of variance C_I：

$C_I=\frac{\mathrm{\σ}(i,j)}{\stackrel{\‾}{I}(i,j)};$

(4) Calculating variance coefficient C of speckle noise in original SAR image_u：

$C_u=\frac{1}{L}$

Wherein L represents the view of the original SAR image;

(5) calculating the change coefficient C of the original SAR image_max:

$C_{max}=\sqrt{2/L-1};$

(6) When C is present_u≤C_I≤C_maxThen, a Gamma filtering method is adopted to carry out speckle noise reduction on the input original SAR image to obtain a SAR image I after noise reduction_Gam(i,j)：

$I_{Gam}(i,j)=\frac{(\mathrm{\α}-L-1)\stackrel{\‾}{I}(i,j)+\sqrt{{(\mathrm{\α}-L-1)}^2{\stackrel{\‾}{I}}^2(i,j)+4\mathrm{\α}L\stackrel{\‾}{I}(i,j)I(i,j)}}{2\mathrm{\α}}$

Wherein α is a heterogeneous parameter, and

(7) for SAR image I after noise reduction_Gam(i, j) performing superpixel division to obtain and output a plurality of superpixel blocks.

2. The SAR image superpixel segmentation method based on Gamma filtering according to claim 1, characterized in that: filtering the input original SAR image by adopting mean filtering in the step (2), wherein the implementation steps are as follows:

(2a) filtering the input original SAR image by adopting a weight coefficient matrix template w (s, t) to obtain the average value of the pixel gray level of the original SAR image

$\stackrel{\‾}{I}(i,j)=\underset{s=-1}{\overset{1}{\Σ}}\underset{t=-1}{\overset{1}{\Σ}}w(s,t)I(i+s,j+t)$

Wherein,s is a row of the template matrix expression, and t is a column of the template matrix expression;

(2b) calculating a standard deviation sigma (i, j) of the pixel gray level of the original SAR image:

$\mathrm{\σ}(i,j)=\sqrt{\frac{1}{N}\frac{1}{M}\underset{i=1}{\overset{N}{\Σ}}\underset{j=1}{\overset{M}{\Σ}}{\left(I\right(i,j)-\stackrel{\‾}{I}(i,j\left)\right)}^2}$

wherein, N represents the number of pixels on the abscissa of the input original SAR image, and M represents the number of pixels on the ordinate of the input original SAR image.

3. The SAR image superpixel segmentation method based on Gamma filtering according to claim 1, characterized in that: the SAR image I subjected to noise reduction in the step (7)_Gam(i, j) performing superpixel segmentation, and realizing the following steps:

(7a) calculating SAR image I after noise reduction_Gam(i, j) difference W (i, j) of pixel values of adjacent pixels in the horizontal direction:

W(i,j)＝|I_Gam(i,j)-I_Gam(i-1,j)|

wherein, I_Gam(I-1, j) represents the pixel value I_Gam(i, j) pixel values of horizontally adjacent pixels;

(7b) calculating SAR image I after noise reduction_Gam(i, j) difference V (i, j) in pixel values of adjacent pixels in the vertical direction:

V(i,j)＝|I_Gam(i,j)-I_Gam(i,j-1)|

wherein, I_Gam(I, j-1) represents a pixel value I_Gam(i, j) pixel values of vertically adjacent pixels;

(7c) calculating SAR image I after noise reduction_Gam(i, j) the gradient value G (i, j) of (i, j) is calculated by the formula:

G(i,j)＝W(i,j)+V(i,j)

(7d) setting SAR image I after noise reduction_Gam(i, j) the number K of superpixel blocks, and calculating the average distance d of the superpixel initial seed points:

$d=\frac{S}{K}$

wherein S is SAR image I after noise reduction_Gam(ii) total area of (i, j), and S ═ MN;

(7e) for SAR image I_GamDisturbing the position of the super-pixel initial seed point on the (I, j) edge to deviate from the SAR image I_Gam(I, j) the edge, i.e. the gradient value at this initial seed point, is smaller than the SAR image I_Gam(i, j) gradient values G (i, j) at the edges, obtaining superpixel seed points;

(7f) and (4) performing boundary growth on the super-pixel seed points, and stopping the boundary growth until the adjacent super-pixel seed points expand to the boundary and collide, so as to obtain a plurality of super-pixel blocks and output the super-pixel blocks.