CN104794730B - SAR image segmentation method based on super-pixel - Google Patents

Abstract

The invention discloses a SAR image segmentation method based on superpixels, which mainly solves the problems of high computational complexity and inability to distinguish small targets in the prior art. The implementation steps are: 1. SAR image input, complete the input of the SAR image to be segmented and obtain image information; 2. Generate superpixels for the input SAR image, and then generate superpixel images; 3. Extract texture features and space of superpixel images Features; 4. By clustering texture features and combining superpixels with spatial features, the final segmentation result of the SAR image is output. The invention can effectively reduce the calculation complexity of the traditional algorithm, shorten the processing time of SAR image segmentation, can distinguish small targets, improve the accuracy of segmentation, and can be used for image processing of airport runways, farmland distribution and geological exploration.

Description

SAR Image Segmentation Method Based on Superpixel

technical field

The invention belongs to the technical field of image processing, in particular to an image segmentation method, which can be used for image processing of airport runways, farmland distribution and geological exploration.

Background technique

Synthetic Aperture Radar (SAR) is a coherent imaging radar operating in the microwave band. It has become an important means of remote sensing observation due to its high-resolution and all-weather, all-time, large-area data acquisition capabilities, and has been widely used in resources, environment, archaeology, and military affairs. SAR images contain rich object categories, but only some of them may be of interest in image understanding. The region of interest varies with the purpose of the application. For example, when detecting flood disasters, the area of interest is water, and in military activities, bridges and roads near military bases may become targets of interest. Therefore, segmenting the areas where these targets are located can not only effectively reduce computer The amount of calculation is improved, the real-time performance of the algorithm is improved, and it is of great significance to correctly identify the target. SAR image segmentation has always been the core problem and difficulty in SAR image understanding and interpretation.

SAR images originate from the backscattering of electromagnetic waves, so there is a large amount of coherent speckle noise on SAR images, which makes each pixel often far from its true value, so when conventional segmentation algorithms are applied to SAR images, the general effect is not very ideal . It can be said that coherent speckle noise is an important factor affecting the quality of SAR image segmentation. At present, there are mainly three methods to reduce the impact of coherent speckle noise on image segmentation: 1. Establish a statistical model of coherent speckle noise; 2. Carry out multi-scale SAR image segmentation methods; 3. Perform noise reduction preprocessing on SAR images .

When segmenting natural images, the additive Gaussian model is often used as the probability model of the image, but for SAR images, due to the different imaging mechanisms and the existence of coherent speckle noise, the additive Gaussian model in natural images cannot be used To represent the statistical distribution characteristics of SAR images. Therefore, some statistical probability models widely used in natural image segmentation, such as Markov random field model and Bayesian model, when segmenting SAR images, assume that the probability distribution conforms to Rayleigh distribution, Gamma distribution, K distribution, etc., see Song Jianshe, Zheng Yongan, and Yuan Lihai, "Synthetic Aperture Radar Image Understanding and Application", Beijing: Science Press, 2008. These algorithms have achieved relatively good segmentation results, but this type of method only operates on a single pixel, so the computational complexity is often large, and it still cannot solve the impact of coherent speckle noise on the segmentation results.

The second method is representative of the SAR image segmentation algorithm as a multi-scale analysis model. The core idea of this type of algorithm is to regard the SAR image as textures of different scales, and according to the nature of different textures for different types of targets, the SAR image will be The segmentation of SAR images is transformed into the recognition of different textures in SAR images. There are many algorithms that apply this method. For example, U.Kandaswamy et al. use co-occurrence matrix to segment SAR images. See U.Kandaswamy, D.A.Adjeroh, and M.C.Lee, "Efficient texture analysis of SAR imagery", IEEE Transactions on Geoscience and Remote Sensing, vol.43, no.9, pp.2075-2083, 2005. X. Zhang uses co-occurrence matrix and energy features of wavelet decomposition to segment SAR images, see X. Zhang, L. Jiao, F. Liu et al al., "Spectral clustering ensemble applied to SAR image segmentation", IEEE Transactions on Geoscience and Remote Sensing, vol.46, no.7, pp.2126-2136, 2008. Hou Biao et al. used the second-generation Bandelet domain hidden Markov Tree model to segment SAR images, see Hou Biao, Xu Jing, Liu Feng, etc., "Image Segmentation Based on the Second-Generation Bandelet Domain Hidden Markov Tree Model", Acta Automatica Sinica, vol.35, no.5, 2009 . But so far, there is still no method that can effectively model all textures, and this type of algorithm is still processed for a single pixel. For SAR images containing a large number of pixels, the computational efficiency of this type of algorithm is relatively low. Low, the calculation speed is relatively slow.

The third method is relatively straightforward, that is, before segmenting the SAR image, it is firstly preprocessed for noise reduction to reduce the impact of coherent speckle noise on the segmentation algorithm. The simplest method in this class is to average multiple pixels, which is a very effective preprocessing method, but its disadvantage is that it cannot maintain the edge information of the image well while reducing noise.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art, and propose a SAR image segmentation method based on superpixels to effectively reduce the computational complexity, shorten the segmentation time, and maintain the edge of the image well while reducing noise information to improve segmentation accuracy.

To achieve the above object, the present invention provides a SAR image segmentation method based on superpixels, comprising the steps of:

1. A SAR image segmentation method based on superpixels, comprising the steps of:

(1) For the input SAR image I, extract its length L, width W and resolution R _I , and estimate the resolution interval R _s of the minimum target according to all images in the SAR image library;

(2) Estimate the number of superpixels N _s according to the image parameters, and generate a superpixel set S={s _i } of the SAR image to be segmented, where s _i is the ith superpixel i=1, 2,..., N _S ;

(3) Calculate the texture feature F _g (i) and spatial feature F _n (i, j) of the superpixel set S, where i, j=1, 2,..., N _s and i≠j;

(4) According to the SAR image segmentation number K, use the K-means algorithm to cluster the texture features F _g (i) of the superpixel set S;

(5) According to the clustering result of the texture feature F _g (i) of the superpixel set S, for any two superpixels s _i and s _j in the superpixel set S, use the nearest neighbor criterion to judge whether the superpixels correspond to the same cluster class, if s _i and s _j correspond to the same cluster, and their spatial features F _n (i, j) = 1, then they are merged, and the texture features F _g (i) and spatial features of the superpixel set S F _n (i, j) is updated; otherwise, no merge is performed;

(6) Count the superpixel number N _s (t) of the updated superpixel set S, and compare the superpixel number N _s (t) with the last statistical result N _s (t-1): if N _s (t)==N _s (t-1), the stop condition is met, and the SAR image segmentation result is output; if N _s (t)<N _s (t-1), the stop condition is not satisfied, and return to step (4) .

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

1. The present invention fully considers that SAR images contain a large amount of coherent speckle noise, and uses superpixels with a scale larger than a single pixel as the basic unit of image segmentation, effectively reducing the impact of SAR image coherent speckle noise on the segmentation results;

2. The present invention fully considers that the size of the SAR image is huge and contains a large number of pixels, so that the segmentation work faces a large number of calculations. The superpixel is used as the basic unit of segmentation, and the extracted texture features and spatial features are simple to operate and the amount of calculation is small. Advantages, effectively reducing the computational complexity of the segmentation algorithm and shortening the processing time of image segmentation;

3. The present invention uses an adaptive superpixel merging method, which can well maintain the edge information of the image while reducing noise, and improve the segmentation accuracy.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is the superpixel image produced by the present invention to a SAR image containing an airport;

Fig. 3 is the superpixel image produced by the present invention to a SAR image containing rivers and different vegetation;

Fig. 4 is the superpixel image produced by the present invention to a SAR image containing airports and buildings;

Fig. 5 is the original SAR image that contains airport, building and vegetation that simulation experiment of the present invention adopts;

Fig. 6 is the result figure of the simulation segmentation of Fig. 5 with the present invention;

Fig. 7 is the original SAR image that contains river, woods and farmland that simulation experiment of the present invention adopts;

Fig. 8 is a result diagram of the simulation segmentation of Fig. 7 by the present invention.

Detailed ways

With reference to Fig. 1, the SAR image segmentation method based on superpixel of the present invention comprises the steps:

Step 1: Obtain image information according to the input SAR image.

Input all the images I in the SAR image library, and extract the length L, width W and resolution R _I of each image from the input SAR image I, and estimate the resolution interval R _s of the minimum target according to these parameters.

Step 2: Estimate the number of superpixels and obtain a set of superpixels.

2a) According to the length L and width W of an image, the resolution R _I and the resolution interval R _s of the minimum target, estimate the number of superpixels N _s of this image according to the following two situations:

The first case is: according to the length L, width W and resolution R _I of the input SAR image to be segmented, the value interval of the number of superpixels N _s is estimated:

Wherein, Num(s) is the value interval of the number of pixels contained in each superpixel, the value interval is [50,200], and the number of superpixels N _s is an integer value arbitrarily selected in the value interval;

The second case is: when the resolution interval R _s of the minimum target cannot be estimated, according to the length L and width W of the input SAR image to be segmented, the estimated value interval of the number of superpixels N _s is:

2b) Generate the superpixel set S={s _i } of the SAR image to be segmented, where s _i is the ith superpixel i=1,2,...,N _s :

2b1) After obtaining the superpixel number N _s of the SAR image, within the SAR image to be segmented, uniformly select N _s pixel seeds, and use these seeds as the center, use the level set method to let these seeds expand according to the following seed expansion level The set evolution equation is expanded:

Φ ⁿ⁺¹ ＝Φ ⁿ +F|▽Φ ⁿ |Δt

Where Φ ⁿ represents the level set function after the nth evolution, |▽Φ ⁿ | represents the gradient value of the level set function after the nth evolution, F represents the velocity function, and Δt represents the time interval of the curve evolution;

2b2) When the edges of all seeds are coincident, the expansion of the seeds is stopped, and each seed is a superpixel s _i , thus completing the generation of the input SAR image superpixel set S={s _i }.

Step 3: Calculate the texture features F _g (i) and spatial features F _n (i, j) of the superpixel set S.

3a) For any superpixel s _i , calculate the texture feature F _g (i) of the superpixel set S by the following formula:

Among them, H(i,1:m) is the grayscale histogram of the superpixel _si , m represents the grayscale number of the image, and Num( _si ) represents the number of pixels contained in the superpixel _si ;

3b) For any two superpixels s _i and s _j , calculate the spatial feature F _n (i,j) of the superpixel set S by the following formula:

Step 4: Cluster the texture features and combine the superpixels with spatial features.

There are many existing clustering methods for texture features, such as K-means, fuzzy C-means method, nearest neighbor criterion, etc. This example uses but is not limited to K-means.

The specific implementation of this step is as follows:

4a) For any two superpixels s _i and s _j , calculate their distances to all cluster centers c _n according to the variable volume distance VBSD, dist(s _i ,c _n ), dist(s _j ,c _n ) , where n=1,2,...,K;

4b) Assign the cluster center corresponding to the minimum distance of the superpixel si to the _ith superpixel _si as the cluster _ci corresponding to the superpixel ^si :

4c) Assign the cluster center corresponding to the minimum value of the superpixel s _j distance to the jth superpixel s _j as the cluster c _j corresponding to the superpixel s ^j :

4d) Determine whether the two superpixels s _i and s _j correspond to the same cluster center, if the two superpixels s _i and s _j correspond to the same cluster center, then the superpixel s _i and s _j belong to the same cluster class, on the contrary, do not belong to the same cluster;

4e) Determine the spatial features of superpixels s _i and s _j in the same cluster, if their spatial features F _n (i,j)=1, merge superpixels s _i and s _j to obtain a new superpixel The set S ^* , on the contrary, is not merged.

Step 5: Update the spatial and texture features of the new superpixel set S ^* .

5a) After the two superpixels s _i and s _j are merged, the sequence number i of one of the two superpixels is removed, and the sequence number j of the other superpixel is used to represent the merged superpixel to obtain a new texture feature F _g '(j):

Among them, F _g ( _i ) represents the texture feature of superpixel si before merging, F _g (j) represents the texture feature of superpixel s _j before merging, and Num( _si ) represents the number of pixels contained in superpixel s _i before merging , Num(s _j ) represents the number of pixels contained in the superpixel s _j before merging;

5b) Delete the texture feature F _g (i) of s _i before merging, and replace the texture feature of s _j after merging with the newly generated texture feature F _g ′(j).

Step 6: Discriminate the result of superpixel merging.

Count the superpixel number N _s (t) of the new superpixel set S ^* after each combination, and compare the superpixel number N _s (t) with the last statistical result N _s (t-1): if N _s (t)==N _s (t-1), then satisfy the stop condition, and output the SAR image segmentation result; if N _s (t)<N _s (t-1), then do not meet the stop condition, return to step (4 ), continue to merge superpixels.

Effect of the present invention can be further illustrated by following simulation experiments to real SAR images:

1. Simulation experiment conditions

The emulation of the present invention is realized on windows XP, SP2, CPU Pentium (R) 4, basic frequency 2.4GHZ, and software platform is Matlab7.0.1. The real SAR images selected for the simulation experiment are two Ku-band SAR images, as shown in Figure 5 and Figure 7, where Figure 5 is the SAR image of the California airport in the United States, with a size of 522×446 and a resolution of 3 meters , containing ground targets such as airports, buildings, and vegetation; Figure 7 is a SAR image near Albuquerque, New Mexico, USA, with a size of 600×432 and a resolution of 1 meter, including rivers, woods, and farmland ground target.

2. Simulation content and results

Simulation 1, using the superpixels generated by the present invention in a SAR image containing an airport to form a superpixel image as shown in FIG. 2 .

In the second simulation, the present invention is used to generate superpixels in a SAR image containing rivers and different vegetation, forming a superpixel image as shown in FIG. 3 .

Simulation 3, using the present invention to generate superpixels in a SAR image containing airports and buildings to form a superpixel image as shown in FIG. 4 .

Simulation four, use the present invention to divide the simulation experiment of Fig. 5, and the result is shown in Fig. 6 .

It can be seen from Fig. 6 that the present invention can accurately locate the edge of the airport runway and accurately separate the runway from other ground targets. For the buildings near the airport, the present invention can find them, and even individual isolated buildings can be distinguished from the surrounding ground targets. For ground vegetation, the invention can effectively separate vegetation with different visual characteristics.

Simulation five, using the present invention to divide the simulation experiment of Fig. 7, the result is shown in Fig. 8.

As can be seen from Figure 8, the present invention can accurately locate the edge of the river and the edge of the road, and separate different vegetation on the ground, and the present invention can also effectively distinguish special targets, such as trees on the grass , small islands in rivers, etc. For the same type of ground objects, the segmentation results obtained by the present invention have good regional consistency and few misclassifications.

From the above simulation experiment results, it can be seen that the present invention can not only generate superpixel images, but also effectively segment SAR images based on superpixels, accurately locate the edges of different ground targets during segmentation, and keep similar targets regional consistency, and can accurately locate and discover isolated special targets in SAR images.

Claims (4)

1. a kind of SAR image segmentation method based on super-pixel, comprises the following steps：

(1) to the SAR image I of input, its length L, width W and resolution ratio R are extracted_I, and according to figure all in SAR image storehouse Resolution ratio section R as estimating minimum target_s；

(2) super-pixel number N is estimated according to image parameter_s, generate the super-pixel set S={ s of SAR image to be split_i, wherein s_iFor i-th of super-pixel i=1,2 ..., N_S；

(3) super-pixel set S textural characteristics F is calculated_gAnd space characteristics F (i)_n(i, j), wherein i, j=1,2 ..., N_sAnd i ≠ j；

(4) number K, the textural characteristics F using K-means algorithms to super-pixel set S are split according to SAR image_g(i) gathered Class；

(5) according to super-pixel set S textural characteristics F_g(i) cluster result, to any two super-pixel in super-pixel set S s_iAnd s_j, using Nearest neighbor rule, judge whether super-pixel corresponds to same cluster, if s_iAnd s_jCorresponding same cluster, and it Space characteristics F_n(i, j)=1, then merged, and to super-pixel set S textural characteristics F_gAnd space characteristics F (i)_n (i, j) is updated；Conversely, without merging；

It is described to use Nearest neighbor rule, judge whether super-pixel corresponds to same cluster, carry out as follows：

5a) for any two super-pixel s_iAnd s_j, it is calculated respectively according to variable volume distance VBSD arrives all cluster centre c_n Distance dist (s_i,c_n)、dist(s_j,c_n), wherein n=1,2 ..., K；

5b) by super-pixel s_iI-th of super-pixel s is distributed to apart from cluster centre corresponding to minimum value_i, as super-pixel s_iIt is corresponding Cluster centre cⁱ：

<mrow> <msup> <mi>c</mi> <mi>i</mi> </msup> <mo>=</mo> <munder> <mi>arg</mi> <mi>n</mi> </munder> <mi>m</mi> <mi>i</mi> <mi>n</mi> <mrow> <mo>(</mo> <mi>d</mi> <mi>i</mi> <mi>s</mi> <mi>t</mi> <mo>(</mo> <mrow> <msub> <mi>s</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>c</mi> <mi>n</mi> </msub> </mrow> <mo>)</mo> <mo>)</mo> </mrow> </mrow>

5c) by super-pixel s_jJ-th of super-pixel s is distributed to apart from cluster centre corresponding to minimum value_j, as super-pixel s_jIt is corresponding Cluster centre c^j：

<mrow> <msup> <mi>c</mi> <mi>j</mi> </msup> <mo>=</mo> <munder> <mi>arg</mi> <mi>n</mi> </munder> <mi>m</mi> <mi>i</mi> <mi>n</mi> <mrow> <mo>(</mo> <mi>d</mi> <mi>i</mi> <mi>s</mi> <mi>t</mi> <mo>(</mo> <mrow> <msub> <mi>s</mi> <mi>j</mi> </msub> <mo>,</mo> <msub> <mi>c</mi> <mi>n</mi> </msub> </mrow> <mo>)</mo> <mo>)</mo> </mrow> </mrow>

Judge super-pixel s_iAnd s_jWhether identical cluster centre is corresponded to, if super-pixel s_iAnd s_jCorresponding cluster centre is identical, Then super-pixel s_iAnd s_jBelong to same cluster；

The textural characteristics F to super-pixel set S_gAnd space characteristics F (i)_n(i, j) is updated, and is carried out as follows：

5d) in two super-pixel s_iAnd s_jAfter merging, the sequence number i of one of two super-pixel is removed, with the sequence of another super-pixel Number j represents the super-pixel after merging, and obtains new textural characteristics F_g′(j)：

<mrow> <msubsup> <mi>F</mi> <mi>g</mi> <mo>&amp;prime;</mo> </msubsup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <msub> <mi>F</mi> <mi>g</mi> </msub> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>&amp;times;</mo> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>+</mo> <msub> <mi>F</mi> <mi>g</mi> </msub> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>&amp;times;</mo> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> </mrow> <mrow> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>+</mo> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow>

Wherein F_g(i) super-pixel s before merging is represented_iTextural characteristics, F_g(j) super-pixel s before merging is represented_jTextural characteristics, Num (s_i) represent super-pixel s before merging_iComprising number of pixels, Num (s_j) represent super-pixel s before merging_jComprising number of pixels；

5e) delete s before merging_iTextural characteristics F_g(i), the textural characteristics F caused by new_g' (j) substitutes s_jTexture after merging is special Sign；

(6) the super-pixel number N of the super-pixel set S after statistical updating_s(t), and by the super-pixel number N_s(t) it is and last Statistical result N_s(t-1) it is compared：If N_s(t)==N_s(t-1) stop condition, output SAR image segmentation knot, are then met Fruit；If N_s(t) ＜ N_s(t-1) stop condition, return to step (4), are then unsatisfactory for.

2. the SAR image segmentation method according to claim 1 based on super-pixel, wherein the step (2) according to figure As parameter Estimation super-pixel number N_s, carried out by the following two kinds situation：

The first situation is：According to length L, width W and the resolution ratio R for inputting SAR image to be split_I, estimate super-pixel number N_sInterval：

<mrow> <msub> <mi>N</mi> <mi>s</mi> </msub> <mo>&amp;Element;</mo> <mfrac> <mrow> <mi>L</mi> <mo>&amp;times;</mo> <mi>W</mi> <mo>&amp;times;</mo> <msubsup> <mi>R</mi> <mi>I</mi> <mn>2</mn> </msubsup> </mrow> <msubsup> <mi>R</mi> <mi>s</mi> <mn>2</mn> </msubsup> </mfrac> <mo>&amp;cap;</mo> <mfrac> <mrow> <mi>L</mi> <mo>&amp;times;</mo> <mi>W</mi> </mrow> <mrow> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow>

Wherein, Num (s) is the interval of the number of pixels contained by each super-pixel, and interval is [50,200], super-pixel Number N_sTo appoint the integer value taken in interval；

Second of situation be：As the resolution ratio section R for being unable to estimate out minimum target_sWhen, according to input SAR image to be split Length L, width W, estimate super-pixel number N_sInterval be：

<mrow> <msub> <mi>N</mi> <mi>s</mi> </msub> <mo>&amp;Element;</mo> <mfrac> <mrow> <mi>L</mi> <mo>&amp;times;</mo> <mi>W</mi> </mrow> <mrow> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>.</mo> </mrow>

3. the SAR image segmentation method according to claim 1 based on super-pixel, wherein being calculated in the step (3) super Pixel set S textural characteristics F_g(i), for any one super-pixel s_i, calculated by equation below：

<mrow> <msub> <mi>F</mi> <mi>g</mi> </msub> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <mi>H</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mn>1</mn> <mo>:</mo> <mi>m</mi> <mo>)</mo> </mrow> </mrow> <mrow> <mi>N</mi> <mi>u</mi> <mi>m</mi> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> </mrow> </mfrac> </mrow>

Wherein, H (i, 1:M) it is super-pixel s_iGrey level histogram, m represent image number of greyscale levels, Num (s_i) represent s_iIt is contained Number of pixels.

4. the SAR image segmentation method according to claim 1 based on super-pixel, wherein being calculated in the step (3) super Pixel set S space characteristics F_n(i, j), for any two super-pixel s_iAnd s_j, calculated by equation below：