CN115131373A - SAR image segmentation method based on texture features and SLIC - Google Patents

SAR image segmentation method based on texture features and SLIC Download PDF

Info

Publication number
CN115131373A
CN115131373A CN202210854382.9A CN202210854382A CN115131373A CN 115131373 A CN115131373 A CN 115131373A CN 202210854382 A CN202210854382 A CN 202210854382A CN 115131373 A CN115131373 A CN 115131373A
Authority
CN
China
Prior art keywords
sar image
edge
sequence
point
image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210854382.9A
Other languages
Chinese (zh)
Other versions
CN115131373B (en
Inventor
余航
刘志恒
郭玉茹
周绥平
蒋浩然
李晨阳
尹相杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202210854382.9A priority Critical patent/CN115131373B/en
Publication of CN115131373A publication Critical patent/CN115131373A/en
Application granted granted Critical
Publication of CN115131373B publication Critical patent/CN115131373B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/40—Analysis of texture
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G06T2207/10044—Radar image
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20048—Transform domain processing
    • G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30181—Earth observation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

本发明提出了一种基于纹理特征和SLIC的SAR图像分割方法，解决了SAR噪声和复杂纹理导致的分割精度低的技术问题。实现步骤为：将SAR图像分为K个超像素块；获得更新后的聚类中心C'i；获得SAR图像的边缘图像；获得搜索区域；计算搜索区域的纹理特征；基于SLIC算法对SAR图像进行分割。本发明对序列

Figure DDA0003747050460000011
三值化，降低噪声影响和计算量；通过三值化序列Hk的平均频谱AMP的收敛性获得搜索区域Ω，计算区域Ω的纹理特征；通过SLIC算法获得SAR图像的分割结果，用于SAR图像分割。

Figure 202210854382

The invention proposes a SAR image segmentation method based on texture features and SLIC, which solves the technical problem of low segmentation accuracy caused by SAR noise and complex textures. The realization steps are: dividing the SAR image into K superpixel blocks; obtaining the updated cluster center C'i ; obtaining the edge image of the SAR image; obtaining the search area; calculating the texture feature of the search area; to split. The sequence of the present invention

Figure DDA0003747050460000011
Ternarization to reduce noise impact and calculation amount; obtain the search area Ω through the convergence of the average spectrum AMP of the ternary sequence H k , and calculate the texture features of the area Ω; obtain the segmentation result of the SAR image through the SLIC algorithm, which is used for SAR Image segmentation.

Figure 202210854382

Description

SAR image segmentation method based on texture features and SLIC
Technical Field
The invention belongs to the technical field of image processing, mainly relates to SAR image segmentation, and particularly relates to an SAR image segmentation method based on texture features and SLIC, which can be applied to the fields of geological detection, disaster monitoring, resource exploration, urban planning and the like.
Background
Synthetic Aperture Radar (SAR) is an active earth observation system with characteristics of all-time, all-weather, high resolution, large breadth and the like by utilizing the Synthetic Aperture principle. An image shot by the SAR is called an SAR image, and the high-resolution SAR image can reflect the geometric characteristics of a target object according to the scattering characteristics of the target itself. In some special scenes, such as battlefields, oceans and the like, people can monitor and reconnaissance by utilizing all-weather and all-weather advantages of SAR.
The SAR image segmentation is to divide an SAR image into countable non-overlapping connected regions, and provides reliable information for SAR image classification identification and automatic interpretation. With the wide application of SAR in multiple fields, the SAR image segmentation is highly regarded as the basis and the premise of automatic understanding and interpretation of the SAR image, and the link between image processing and image analysis and understanding is connected, so that the quality and the efficiency of subsequent image analysis and understanding are directly influenced by the quality of a segmentation result. Currently, methods for SAR image segmentation are mainly classified into a threshold-based segmentation method, a region-based segmentation method, a superpixel-based segmentation method, and the like, wherein the superpixel-based segmentation method is considered as one of the most effective means for processing similarity and uncertainty in the SAR image. Simple Linear Iterative Clustering (SLIC) is a typical super-pixel generation method. The method uses the intensity and position iteration of the pixels to form a set for clustering, the calculation cost is low, and the performance is superior to most other algorithms.
Currently, improved methods based on SLIC image segmentation have been proposed. For example, a chinese patent with application publication number CN108596936A, entitled "a method for super-pixel segmentation of images based on multi-level SLIC" discloses an algorithm for super-pixel segmentation of images based on multi-level SLIC, which first performs super-pixel coarse segmentation on an input image by using a SLIC method; and then calculating the gray standard deviation of each super pixel in the segmentation result, if the gray standard deviation is higher than a preset threshold value, performing iterative SLIC segmentation on the super pixels until the gray standard deviation of all the super pixels in the image is less than or equal to the preset threshold value, and combining the over-segmented super pixels in the image. According to the method, the over-divided super pixels in the image are combined, data redundancy generated in the iterative division process is eliminated, the number of the super pixels is reduced, and the image division effect is improved.
A super-pixel is a block of pixels with irregular contours formed by several neighboring pixels with similar characteristics. The SLIC is the most effective means for processing similarity and uncertainty in an image segmentation algorithm. The method utilizes the intensity and position iteration of the pixels to form a set for clustering, has low calculation cost and performance superior to most other algorithms, but for SAR images with complex textures, the SLIC performance is not ideal due to the influence of image texture change and the influence of noise, and the segmentation precision is low.
In summary, the existing method only performs iterative optimization on the SLIC algorithm, eliminates data redundancy generated in the iterative process, and reduces erroneous segmentation in the super-pixel segmentation process. The accuracy of image segmentation is still low since no other pixel information of the image is taken into account.
Disclosure of Invention
The technical problem to be solved by the invention is to provide an SAR image segmentation method based on texture features and SLIC aiming at the defects in the prior art, so that the influence of noise is reduced, the image texture feature information is fused, and the segmentation precision of the SAR image is improved.
The invention relates to an SAR image segmentation method based on texture features and SLIC, which comprises the following steps:
(1) dividing the SAR image I into K superpixel blocks: inputting an SAR image I to be segmented, and setting the size of the pixel value of the input SAR image I to be L multiplied by W, wherein N is L multiplied by W, and the pixel matrix of the SAR image I is
Figure BDA0003747050440000021
Wherein, I (x) l ,y u ) Expressing the pixels of the ith row and the uth column in the SAR image I, and randomly distributing K original clustering centers C i ＝[l i ,a i ,b i ,x i ,y i ] T Wherein I denotes the ith cluster center, I is 1, … …, K, the SAR image I is divided into K super-patches, each super-pixel has a side length of
Figure BDA0003747050440000022
(2) Updating the clustering center C i Obtaining updated cluster center C' i : within 3-by-3 neighborhood of the original cluster center, the gradient value G is measured i The pixel with the smallest (x, y) value is set as the updated cluster center C' i Wherein i represents the ith cluster center, i ═ 1, … …, K;
(3) calculating each updated cluster center C' i The edge operator of (2): at the cluster center C 'updated with the first one' 1 In the (2S +1) × (2S +1) neighborhood of the central pixel, the neighborhood is divided into four parts, which are sequentially represented from left to right from top to bottom as
Figure BDA0003747050440000031
And
Figure BDA0003747050440000032
the four parts are subjected to horizontal weighted summation to obtain an updated clustering center C' 1 Edge operator of horizontal direction
Figure BDA0003747050440000033
The four parts are vertically weighted and summed to obtain a clustering center C' i Edge operator of vertical direction
Figure BDA0003747050440000034
Similarly, other updated cluster centers C 'are obtained through calculation' i Horizontal edge operator
Figure BDA0003747050440000035
And vertical edge operator
Figure BDA0003747050440000036
(4) Obtaining an edge image EI of the SAR image I: the first updated cluster center C' 1 In the horizontal direction of
Figure BDA0003747050440000037
And the vertical direction
Figure BDA0003747050440000038
Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient in the horizontal direction
Figure BDA0003747050440000039
And edge coefficient in vertical direction
Figure BDA00037470504400000310
Then the first updated cluster center C' 1 Has an edge coefficient of
Figure BDA00037470504400000311
Similarly, other updated cluster centers C 'are obtained through calculation' i Has an edge coefficient of
Figure BDA00037470504400000312
All the updated edge coefficients of the clustering centers form an edge image EI of the SAR image I;
(5) calculating each updated cluster center C' i Of different step sizes
Figure BDA00037470504400000313
With the first updated clustering center C 'on the SAR image I' 1 As a starting point, is counted as a point (x) 1 ,y 1 ) Point (x) 1 ,y 1 ) The sequence of (A) is
Figure BDA00037470504400000314
Along the y direction, with N/100 as the step size, the second point is obtained as (x) 1 ,y 1+N/100 ) Point (x) 1 ,y 1+N/100 ) The sequence of (A) is
Figure BDA00037470504400000315
Taking 2 XN/100 as a step length to obtain a third point of (x) 1 ,y 1+2×N/100 ) Point (x) 1 ,y 1+2×N/100 ) The sequence of (A) is
Figure BDA00037470504400000316
And the analogy is repeated, and the step length of (N-1) multiplied by N/100 is taken as the step length, so that the nth point is obtained as (x) 1 ,y 1+(n-1)×N/100 ) Point (x) 1 ,y 1+(n-1)×N/100 ) Is the sequence of
Figure BDA00037470504400000317
Figure BDA00037470504400000318
Is shown as
Figure BDA0003747050440000041
n represents the sequence of the nth point, and n is 1,2,3, …; similarly, all the cluster centers C 'are obtained by calculation' i Of different step sizes
Figure BDA0003747050440000042
Wherein K is 1,2, … …, K, which indicates the sequence of the kth updated cluster center, and K indicates the number of cluster centers;
(6) each sequence is divided into
Figure BDA0003747050440000043
Tri-valued to obtain a tri-valued sequence H k Spectrum of (a): the first sequence
Figure BDA0003747050440000044
Tri-valued to obtain a tri-valued sequence of
Figure BDA0003747050440000045
Calculation of a ternary sequence H by means of fast Fourier transform FFT 1 To obtain a ternary sequence H 1 Frequency spectrum FH 1 (ii) a In the same way, other sequences are added
Figure BDA0003747050440000046
Tri-valued to obtain a tri-valued sequence
Figure BDA0003747050440000047
Obtaining a ternary sequence H k Frequency spectrum FH k ；
(7) Computing a ternary sequence H k Obtaining the search area Ω: obtaining a ternary sequence H k Frequency spectrum FH k Then, calculate the frequency spectrum FH k If the AMP is at a point
Figure BDA0003747050440000048
To converge, the farthest point of the search area is the point
Figure BDA0003747050440000049
If AMP is not convergent, the farthest point of the search area is the extreme point of the image gradient, and a search area omega with the farthest point as the radius is obtained;
(8) calculating the texture feature FT of the search region omega: after the search region omega is obtained, the region is calculatedThe texture feature FT of omega is,
Figure BDA00037470504400000410
wherein,
Figure BDA00037470504400000411
is the average intensity of the region omega, alpha 4 Denotes the kurtosis, α 4 ＝μ 4 /σ 4 ，μ 4 Fourth moment, δ, representing the mean of the region Ω 4 Is the variance of the region Ω;
(9) segmenting the SAR image I based on the SLIC algorithm: fusing the SLIC algorithm with the texture features, and segmenting the SAR image I;
(9a) setting the pixel with the minimum boundary coefficient e as the clustering center of the boundary coefficients
Figure BDA00037470504400000412
At the cluster center C 'updated with the first one' 1 In the 3 x 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA0003747050440000051
Boundary coefficient clustering center
Figure BDA0003747050440000052
Is (x' 1 ,y′ 1 ) Boundary coefficient clustering center
Figure BDA0003747050440000053
Gray value of g 1 Boundary coefficient clustering center
Figure BDA0003747050440000054
Is characterized by a texture of FT 1 (ii) a Similarly, the cluster center C 'updated with the other clusters' i In 3 × 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA0003747050440000055
Boundary coefficient aggregationClass center
Figure BDA0003747050440000056
Is (x' i ,y′ i ) Boundary coefficient clustering center
Figure BDA0003747050440000057
Gray value of g i Boundary coefficient clustering center
Figure BDA0003747050440000058
Is characterized by a texture of FT i ；
(9b) Calculating the clustering center of pixels and boundary coefficients on the SAR image I
Figure BDA0003747050440000059
The segmented image I' is obtained: calculating the clustering center of pixels and boundary coefficients on the SAR image I
Figure BDA00037470504400000510
The distance between the pixels is D, the pixels are allocated to the nearest boundary coefficient cluster center, wherein D is alpha D w +d g +d t ，d g To represent the distance of the grey value between the pixel on the SAR image I and the cluster center, d t Representing the texture distance between the pixel and the cluster center on the SAR image I, d w Representing the spatial distance between the pixels on the SAR image I and the clustering center, alpha is an influence coefficient determining the quality of the super-pixels, the pixels at the edge and the pixels not positioned at the edge should select different alpha,
Figure BDA00037470504400000511
wherein e is l,u Is a pixel point (x) l ,y u ) The edge coefficient of (e), max (EI), is the maximum value of the edge image EI, and q is 91%, finally obtaining the segmented image I' of the SAR image I.
The method considers the influence of the image texture change on the image segmentation, and solves the technical problem of low segmentation precision caused by complex texture in the SAR image segmentation process.
Compared with the prior art, the invention has the following advantages:
the segmentation precision is improved: due to the existence of complex texture changes in the SAR image, the existing superpixel algorithm only uses single pixel information, such as Lab color space information, and cannot realize accurate segmentation. The texture features of the image are introduced into an SLIC algorithm, the texture features in the region are firstly calculated, then the texture distance between the pixels and the clustering center is increased when the distance D between the pixels and the clustering center on the SAR image is calculated, and finally the SAR image is accurately segmented. Compared with the segmentation precision of the existing method, the segmentation precision of the SAR image is improved by 15%.
The calculation cost is low, and the noise influence is reduced: synthetic aperture radar coherent imaging mechanisms result in the generation of speckle noise, which is an inherent characteristic of SAR systems. The echoes of scattering points in each ground resolution unit are received by the SAR and then are subjected to coherent superposition, so that the condition of random fading of amplitude and phase can occur in the resolution unit, speckle noise with staggered brightness and darkness can occur on the SAR image, and the segmentation precision of the SAR image can be influenced by the speckle noise. The invention aims to reduce the influence of noise on image segmentation precision and to reduce the updated clustering center C' i Of different step sizes
Figure BDA0003747050440000061
By carrying out the three-valued processing, the influence of image noise can be reduced, and the calculation amount of the algorithm can be reduced.
Compared with the existing segmentation algorithm, the SAR image segmentation method based on the SLIC algorithm can better solve the segmentation problem of the SAR image under the influence of the image texture, reduces the influence of noise and calculation cost in the segmentation process, and improves the segmentation precision of the SAR image by combining the texture features of the image with the SLIC algorithm and adding the texture distance in clustering.
Drawings
FIG. 1 is a schematic block flow diagram of the present invention;
fig. 2 is an original image of an SAR image, which is also an experimental map of the present invention;
FIG. 3 is a graph of the segmentation results of FIG. 2 using the method of the present invention.
Detailed Description
Example 1
Because SAR images often contain speckle noise and complex textures, for SAR images containing complex textures, the effect of SAR image segmentation directly by using the SLIC method is not ideal, and segmentation is inaccurate due to texture intensity change inside an image target and noise at the edge of the image.
The existing SAR image segmentation method only considers the influence of image noise on image segmentation and does not consider the influence of complex textures. When processing image noise, common image filtering algorithms, such as median filtering, gaussian filtering algorithm, etc., are used to eliminate the influence of noise on the segmentation precision. However, the detail information of the image is reduced by the common image filtering algorithm. In order to overcome the defects, the invention provides an SAR image segmentation method based on texture features and SLIC.
The invention relates to an SAR image segmentation method based on texture features and SLIC, and referring to fig. 1, fig. 1 is a schematic flow chart of the invention, and the method comprises the following steps:
(1) initializing a clustering center, and dividing the SAR image I into K superpixel blocks: inputting a SAR image I to be segmented, and setting the size of the input SAR image I to be L multiplied by W to be 640 multiplied by 480(L and W respectively represent the width and the height of the SAR image I), wherein N is the total number of pixels in the SAR image I, and N is L multiplied by W to be 307200; the SAR image may be represented in a matrix form, so in a computer image processing program, a two-dimensional array is usually used to store image data, and then the pixel matrix of the SAR image I may be represented as
Figure BDA0003747050440000071
Wherein, I (x) l ,y u ) Which represents the pixel of the ith row and the uth column in the SAR image I. And randomly setting K initialized clustering centers on the SAR image, wherein K determines the average number of pixels contained in each super pixel and influences the precision of super pixel segmentation. K original cluster centers are denoted C i ＝[l i ,a i ,b i ,x i ,y i ] T Wherein i is the serial number of the clustering center, i is 1, … …, K, l is the brightness component of the Lab color space, a, b are the color opponent dimensions of the Lab color space, and K original clustering centers C are obtained i Dividing the SAR image I into K uniform super-image blocks with side length of S multiplied by S, wherein
Figure BDA0003747050440000072
(2) Updating the clustering center C i Obtaining an updated cluster center C' i : in calculating the gradient values of the image, the size of the image neighborhood may also be set to 3 × 3, 5 × 5, 7 × 7, 9 × 9, etc., and a larger neighborhood may result in a larger amount of calculation in calculating the gradient values, and in order to reduce the amount of calculation of the algorithm, the size of the image neighborhood is selected to be 3 × 3 in this example. Obtaining the original clustering center C i Then, the first original clustering center C on the SAR image I 1 Calculating gradient value G of each pixel in the neighborhood within 3 x 3 neighborhood 1 (x, y) dividing the gradient G in 3-by-3 neighborhood 1 The pixel with the smallest (x, y) value is set as the updated cluster center C' 1 (ii) a Second original clustering center C on SAR image I 2 Within 3 x 3 neighborhood, the gradient values G of all pixels within the neighborhood are calculated 2 (x, y) converting the gradient value G 2 The pixel with the smallest (x, y) is set as the updated cluster center C' 2 (ii) a A third original clustering center C on the SAR image I 3 Within 3 x 3 neighborhood, the gradient values G of all pixels within the neighborhood are calculated 3 (x, y) dividing the gradient value G 3 The pixel with the smallest (x, y) value is set as the updated cluster center C' 3 (ii) a And in the same way, by analogy, the original clustering center C is clustered on the SAR image I i In the 3 x 3 neighborhood, the gradient value G of all pixels in the neighborhood is calculated i (x, y) dividing the gradient value G i The pixel with the smallest (x, y) value is set as the updated cluster center C' i Wherein i is the updated serial number of the cluster center, i is 1, … …, K is the total number of the cluster centers.
(3) Calculating each updated cluster center C' i The edge operator of (2): at the first updated cluster center C' 1 Within the (2S +1) × (2S +1) neighborhood of the central pixel,the neighborhood is uniformly divided into four parts which are sequentially expressed as
Figure BDA0003747050440000081
And
Figure BDA0003747050440000082
will be in the horizontal direction
Figure BDA0003747050440000083
And
Figure BDA0003747050440000084
Figure BDA0003747050440000085
and
Figure BDA0003747050440000086
weighted summation is carried out to obtain an updated clustering center C' 1 Edge operator of the horizontal direction
Figure BDA0003747050440000087
Will be in the vertical direction
Figure BDA0003747050440000088
And
Figure BDA0003747050440000089
and
Figure BDA00037470504400000810
weighted summation is carried out to obtain a clustering center C' 1 Edge operator of vertical direction
Figure BDA00037470504400000811
At the second updated cluster center C' 2 In the (2S +1) × (2S +1) neighborhood of the central pixel, the neighborhood is uniformly divided into four parts, which are sequentially represented as
Figure BDA00037470504400000812
And
Figure BDA00037470504400000813
will be in the horizontal direction
Figure BDA00037470504400000814
And
Figure BDA00037470504400000815
and
Figure BDA00037470504400000816
weighted summation is carried out to obtain an updated clustering center C' 2 Edge operator of the horizontal direction
Figure BDA00037470504400000817
Will be in the vertical direction
Figure BDA00037470504400000818
And
Figure BDA00037470504400000819
and
Figure BDA00037470504400000820
weighted summation is carried out to obtain a clustering center C' 2 Edge operator of vertical direction
Figure BDA00037470504400000821
At cluster center C 'updated with the third' 3 In the (2S +1) × (2S +1) neighborhood of the central pixel, the neighborhood is uniformly divided into four parts, which are sequentially represented as
Figure BDA00037470504400000822
And
Figure BDA00037470504400000823
will be in the horizontal direction
Figure BDA00037470504400000824
And
Figure BDA00037470504400000825
and
Figure BDA00037470504400000826
weighted summation is carried out to obtain an updated clustering center C' 3 Edge operator of horizontal direction
Figure BDA00037470504400000827
Will be in the vertical direction
Figure BDA00037470504400000828
And
Figure BDA00037470504400000829
and
Figure BDA00037470504400000830
weighted summation is carried out to obtain a cluster center C' 3 Edge operator of vertical direction
Figure BDA00037470504400000831
Similarly, all updated cluster centers C 'are calculated by analogy' i Horizontal edge operator
Figure BDA00037470504400000832
And vertical edge operator
Figure BDA00037470504400000833
Wherein i is the updated serial number of the clustering center, i is 1, … …, K is the total number of the clustering centers.
(4) Obtaining an edge image EI of the SAR image I: the first updated cluster center C' 1 Horizontal direction edge operator of
Figure BDA00037470504400000834
And vertical direction edge operators
Figure BDA00037470504400000835
Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient in the horizontal direction
Figure BDA00037470504400000836
And edge coefficient in vertical direction
Figure BDA00037470504400000837
Then the first updated cluster center C' 1 Has an edge coefficient of
Figure BDA0003747050440000091
Second updated cluster center C' 2 Of the edge operator in the horizontal direction
Figure BDA0003747050440000092
And vertical direction edge operators
Figure BDA0003747050440000093
Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient in the horizontal direction
Figure BDA0003747050440000094
And edge coefficient in vertical direction
Figure BDA0003747050440000095
Then the second updated cluster center C' 2 Has an edge coefficient of
Figure BDA0003747050440000096
The third updated cluster center C' 3 Of the edge operator in the horizontal direction
Figure BDA0003747050440000097
And vertical direction edge operators
Figure BDA0003747050440000098
Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient in the horizontal direction
Figure BDA0003747050440000099
And edge coefficient in vertical direction
Figure BDA00037470504400000910
Then the third updated cluster center C' 3 Has an edge coefficient of
Figure BDA00037470504400000911
Similarly, calculating all updated clustering centers C 'by analogy' i Has an edge coefficient of
Figure BDA00037470504400000912
Wherein i is the updated serial number of the clustering center, i is 1, … …, K is the total number of the clustering centers; edge coefficient e of all updated cluster centers i An edge image EI constituting the SAR image I.
(5) Calculating each updated cluster center C' i Of different step sizes
Figure BDA00037470504400000913
With the first updated clustering center C 'on the SAR image I' 1 As a starting point, is counted as a point (x) 1 ,y 1 ) Point (x) 1 ,y 1 ) The sequence of (A) is
Figure BDA00037470504400000914
Along the y direction, with N/100 as the step size, the second point is obtained as (x) 1 ,y 1+N/100 ) Point (x) 1 ,y 1+N/100 ) The sequence of (A) is
Figure BDA00037470504400000915
Taking 2 XN/100 as a step length to obtain a third point of (x) 1 ,y 1+2×N/100 ) Point (x) 1 ,y 1+2×N/100 ) The sequence of (A) is
Figure BDA00037470504400000916
And the analogy is repeated, and the step length of (N-1) multiplied by N/100 is taken as the step length, so that the nth point is obtained as (x) 1 ,y 1+(n-1)×N/100 ) Point (x) 1 ,y 1+(n-1)×N100 ) Is the sequence of
Figure BDA00037470504400000917
Figure BDA00037470504400000918
Is shown as
Figure BDA00037470504400000919
n represents the sequence of the nth point, and n is 1,2,3, …; with the second updated clustering center C 'on the SAR image I' 2 As a starting point, is counted as a point (x) 2 ,y 2 ) Point (x) 2 ,y 2 ) The sequence of (A) is
Figure BDA00037470504400000920
Along the y direction, with N/100 as the step size, the second point is obtained as (x) 2 ,y 2+N/100 ) Point (x) 2 ,y 2+N/100 ) The sequence of (A) is
Figure BDA00037470504400000921
Taking 2 XN/100 as a step length to obtain a third point of (x) 2 ,y 2+2×N/100 ) Point (x) 2 ,y 2+2×N/100 ) The sequence of (A) is
Figure BDA00037470504400000922
And the analogy is repeated, and the step length of (N-1) multiplied by N/100 is taken as the step length, so that the nth point is obtained as (x) 2 ,y 2+(n-1)×N/100 ) Point (x) 2 ,y 2+(n-1)×N/100 ) Is the sequence of
Figure BDA00037470504400000923
Figure BDA00037470504400000924
Is shown as
Figure BDA0003747050440000101
n represents the sequence of the nth point, and n is 1,2,3, …; with the third updated clustering center C 'on the SAR image I' 3 As a starting point, is counted as a point (x) 3 ,y 3 ) Point (x) 3 ,y 3 ) The sequence of (A) is
Figure BDA0003747050440000102
Along the y direction, with N/100 as the step size, the second point is obtained as (x) 3 ,y 3+N/100 ) Point (x) 3 ,y 3+N/100 ) The sequence of (A) is
Figure BDA0003747050440000103
Taking 2 XN/100 as a step length to obtain a third point of (x) 3 ,y 3+2×N/100 ) Point (x) 3 ,y 3+2×N/100 ) The sequence of (A) is
Figure BDA0003747050440000104
And the analogy is repeated, and the step length of (N-1) multiplied by N/100 is taken as the step length, so that the nth point is obtained as (x) 3 ,y 3+(n-1)×N/100 ) Point (x) 3 ,y 3+(n-1)×N/100 ) Is the sequence of
Figure BDA0003747050440000105
Figure BDA0003747050440000106
Is shown as
Figure BDA0003747050440000107
n represents the sequence of the nth point, and n is 1,2,3, …; similarly, all updated cluster centers C 'are calculated and obtained by analogy' i Of different step sizes
Figure BDA0003747050440000108
Where K is the sequence of the kth updated cluster center, and K is 1,2,3, … …, where K denotes the total number of cluster centers.
(6) Each sequence is divided into
Figure BDA0003747050440000109
Tri-valued to obtain a tri-valued sequence H k Spectrum of (a): in order to reduce the calculation amount of the algorithm and the influence of noise in the SAR image on the segmentation precision, the invention sequences
Figure BDA00037470504400001010
Carrying out ternary processing, which specifically comprises the following steps: the first sequence
Figure BDA00037470504400001011
Thresholded, if sequence
Figure BDA00037470504400001012
Of the first point in (a)
Figure BDA00037470504400001013
If the value of (A) is greater than 1, the ternary sequence H is formed 1 Equal to 1 if the sequence
Figure BDA00037470504400001014
Of the first point in (a)
Figure BDA00037470504400001015
Is less than-1, the ternary sequence H 1 Is equal to-1 if the sequence
Figure BDA00037470504400001016
Of the first point in (a)
Figure BDA00037470504400001017
Is less than-1 and greater than 1, the ternary sequence H 1 Equals 0, then the sequence H is thresholded 1 Can be expressed as
Figure BDA00037470504400001018
Calculation of a ternary sequence H by means of fast Fourier transform FFT 1 To obtain a ternary sequence H 1 Frequency spectrum FH 1 (ii) a The second sequence is added
Figure BDA00037470504400001019
Thresholded, if sequence
Figure BDA00037470504400001020
Of the first point in (a)
Figure BDA00037470504400001021
Is greater than 1, the ternary sequence H 2 Equal to 1 if the sequence
Figure BDA00037470504400001022
Of the first point in (a)
Figure BDA00037470504400001023
Is less than-1, the ternary sequence H 2 Is equal to-1 if the sequence
Figure BDA00037470504400001024
Of the first point in (a)
Figure BDA0003747050440000111
Is less than-1 and greater than 1, the ternary sequence H 2 Equals 0, the ternary sequence H 2 Can be expressed as
Figure BDA0003747050440000112
Calculation of a ternary sequence H by means of fast Fourier transform FFT 2 To obtain a ternary sequence H 2 Frequency spectrum FH 2 (ii) a The third sequence
Figure BDA0003747050440000113
Thresholded, if sequence
Figure BDA0003747050440000114
Of the first point in (a)
Figure BDA0003747050440000115
If the value of (A) is greater than 1, the ternary sequence H is formed 3 Equal to 1 if the sequence
Figure BDA0003747050440000116
Of the first point in (a)
Figure BDA0003747050440000117
Is less than-1, the ternary sequence H 3 Is equal to-1 if the sequence
Figure BDA0003747050440000118
Of the first point in (a)
Figure BDA0003747050440000119
Is less than-1 and greater than 1, then the sequence H is thresholded 3 Equals 0, the ternary sequence H 3 Can be expressed as
Figure BDA00037470504400001110
Calculation of a ternary sequence H by means of fast Fourier transform FFT 3 To obtain a ternary sequence H 3 Frequency spectrum FH 3 (ii) a For the same reason, and so on, the other sequences are repeated
Figure BDA00037470504400001111
Tri-valued to obtain a tri-valued sequence
Figure BDA00037470504400001112
Obtaining a ternary sequence H by performing Fast Fourier Transform (FFT) on the ternary sequence k Frequency spectrum FH k 。
(7) Computing a ternary sequence H K Obtaining the search area Ω: in order to reduce the complexity of the algorithm and improve the processing efficiency of the algorithm, a ternary sequence H is obtained k Frequency spectrum FH k Then, by applying to all frequency spectrums FH k Summing and averaging to obtain an average frequency spectrum AMP; judging convergence of the average spectrum AMP if AMP is at the point
Figure BDA00037470504400001113
When the point is converged, the farthest point of the search area is the point
Figure BDA00037470504400001114
If AMP is not convergent, the farthest point of the search area is the extreme point of the image gradient, and finally the search area Ω with the farthest point as the radius can be obtained.
(8) Calculating the texture feature FT of the search region omega: textureThe visual feature reflects the homogeneity phenomenon in the image, embodies the surface structure organization arrangement attribute with slow change or periodic change of the object surface, and can reflect the spatial distribution attribute of the pixels. After the search region omega is obtained, by calculating the texture feature FT of the region omega,
Figure BDA0003747050440000121
wherein,
Figure BDA0003747050440000122
is the average intensity of the region omega, alpha 4 The peak-to-peak ratio is expressed,
Figure BDA0003747050440000123
μ 4 fourth order moment, δ, representing the mean of the region Ω 4 Is the variance of the region omega.
(9) Segmenting the SAR image I based on the SLIC algorithm: and fusing the SLIC algorithm with the texture features to segment the SAR image I.
(9a) Setting the pixel with the minimum boundary coefficient e as the clustering center of the boundary coefficients
Figure BDA0003747050440000124
At cluster center C 'updated with the first one' 1 In 3 × 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA0003747050440000125
Boundary coefficient clustering center
Figure BDA0003747050440000126
Is (x' 1 ,y′ 1 ) Edge coefficient clustering center
Figure BDA0003747050440000127
Gray value of g 1 Boundary coefficient clustering center
Figure BDA0003747050440000128
Is characterized by the texture ofFT 1 (ii) a At cluster center C 'updated with the second' 2 In the 3 x 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA0003747050440000129
Boundary coefficient clustering center
Figure BDA00037470504400001210
Is (x' 2 ,y' 2 ) Boundary coefficient clustering center
Figure BDA00037470504400001211
Gray value of g 2 Boundary coefficient clustering center
Figure BDA00037470504400001212
Is characterized by a texture of FT 2 (ii) a At cluster center C 'updated with the third' 3 In 3 × 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA00037470504400001213
Boundary coefficient clustering center
Figure BDA00037470504400001214
Is (x' 3 ,y' 3 ) Boundary coefficient clustering center
Figure BDA00037470504400001215
Gray value of g 3 Boundary coefficient clustering center
Figure BDA00037470504400001216
Is characterized by a texture of FT 3 (ii) a Similarly, and so on, at the updated cluster center C' i In 3 × 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure BDA00037470504400001217
Boundary systemNumber clustering center
Figure BDA00037470504400001218
Is (x' i ,y′ i ) Boundary coefficient clustering center
Figure BDA00037470504400001219
Gray value of g i Boundary coefficient clustering center
Figure BDA00037470504400001220
Is characterized by a texture of FT i 。
(9b) Calculating clustering center of pixels and boundary coefficients on SAR image I
Figure BDA00037470504400001221
The segmented image I' is obtained: for each pixel point on the SAR image, calculating the clustering center of the pixels and the boundary coefficients on the SAR image I
Figure BDA00037470504400001222
The distance D between the pixel points and the nearest boundary coefficient cluster center according to the distance measurement result obtained by calculation
Figure BDA00037470504400001223
Keeping the categories consistent. Wherein D is α · D w +d g +d t ，d g D is a distance representing the gray value between the pixel on the SAR image I and the cluster center t Representing the texture distance between the pixel and the cluster center on the SAR image I, d w The space distance between the pixels on the SAR image I and the clustering center is represented, alpha is an influence coefficient for determining the quality of the super-pixels, different alpha should be selected for the pixels at the edge and the pixels not at the edge,
Figure BDA0003747050440000131
wherein e is l,u Is a pixel point (x) l ,y u ) Max (EI) is the maximum value in the edge image EI, and in the present invention, q is the scale factor and is an experimentAnd determining a value, and finally obtaining a segmentation image I' of the SAR image I, wherein the final q is 91% after a plurality of times of experimental verification.
The invention provides an integral technical scheme of an SAR image segmentation method based on texture features and SLIC, and in view of the most effective means of similarity and uncertainty of the SLIC method in processing images, the SAR image segmentation method based on SLIC uses the SLIC method and fuses the texture features of the images, so that the segmentation precision of the SAR images can be effectively improved.
The scheme of the invention firstly calculates the clustering center C 'in the superpixel segmentation' i Obtaining the edge coefficient e of the clustering center i Obtaining an edge image EI of the SAR image I; then, the updated cluster center C 'is calculated' i Of different step sizes
Figure BDA0003747050440000132
In order to reduce the calculation amount of the algorithm and reduce the influence of noise, the sequence is processed
Figure BDA0003747050440000133
Ternary to obtain new sequence H k (ii) a Then, the sequence H is obtained by fast Fourier transform calculation K The average frequency spectrum AMP is obtained by analyzing the convergence and the convergence point of the AMP, a search region omega is obtained, and then the texture feature FT of the region omega is obtained; and finally, combining the texture feature FT of the image with an SLIC algorithm to obtain a segmented image. According to the scheme of the invention, while the influence of noise is reduced, the textural feature information of the image is fused, and the segmentation precision of the SAR image is improved.
Example 2
The SAR image segmentation method based on texture features and SLIC is the same as that of embodiment 1, and the original clustering center C on the SAR image I in the step (2) of the invention i Within 3 x 3 neighborhood, the gradient values G of all pixels within the neighborhood are calculated i (x, y) dividing the gradient value G i The pixel with the smallest (x, y) value is set as the updated cluster center C' i Therein gradient value G i The formula for the calculation of (x, y) is:
G i (x,y)＝dx i (x,y)+dy i (x,y)
dx i (x, y) is the derivation of the point (x, y) in the x-direction:
dx i (x,y)＝I(x+1,y)-I(x,y)
dy i (x, y) is the derivative of point (x, y) in the y-direction:
dy i (x,y)＝I(x,y+1)-I(x,y)
wherein, I (x, y) represents the gray value of the pixel (x, y), and in the 3 × 3 neighborhood, the gradient values G (x, y) of the pixels are sequentially calculated, the pixel with the minimum gradient value G (x, y) is set as the new cluster center, and (x, y) represents the coordinates of the calculated pixel.
The gradient at each point in the image points in the direction in which the pixel grey value increases the most, with a magnitude corresponding to the rate of change in this direction. When the image edge is obvious, the image edge has a larger gradient value; when the image is smoother, the gray value change is smaller, and the corresponding gradient value is also smaller. In the invention, the pixel with the minimum gradient value G (x, y) in the 3 multiplied by 3 neighborhood is taken as a new clustering center, so that the super-pixel clustering effect can be improved, and the segmentation precision is improved.
Example 3
The SAR image segmentation method based on texture features and SLIC is the same as that in embodiment 1-2, and all updated cluster centers C 'are obtained through calculation in step (3) of the invention' i Horizontal edge operator
Figure BDA0003747050440000141
And vertical edge operator
Figure BDA0003747050440000142
Edge operator in horizontal direction
Figure BDA0003747050440000143
And edge operator in vertical direction
Figure BDA0003747050440000144
The calculation formulas are respectively as follows:
Figure BDA0003747050440000145
Figure BDA0003747050440000146
wherein p represents an updated cluster center C' i The p-th part of the neighborhood, p is 1,2,3,4, i is the updated serial number of the cluster center, i is 1, … …, K, H is the S × S matrix,
Figure BDA0003747050440000151
when calculating only cluster center C' i Horizontal direction edge operator A x When only the cluster center C 'is calculated without considering the vertical characteristic of the image edge' i Vertical edge operator A y When the horizontal characteristics of the image edges are not taken into account. The invention calculates the clustering center C 'respectively' i Edge operator A in horizontal and vertical directions x And A y The method avoids the unicity of the edge operator only considering the horizontal direction or the vertical direction, the invention considers more factors, and the obtained edge coefficient is more accurate.
Example 4
The SAR image segmentation method based on texture features and SLIC is the same as in embodiments 1-3, and all updated clustering centers C 'are obtained through calculation in step (4) of the invention' i Has an edge coefficient of
Figure BDA0003747050440000152
Edge coefficient e of all updated cluster centers i Edge image EI forming SAR image I, in which edge coefficients in horizontal direction
Figure BDA0003747050440000153
And edge coefficient in vertical direction
Figure BDA0003747050440000154
The calculation formulas of (A) and (B) are respectively as follows:
Figure BDA0003747050440000155
Figure BDA0003747050440000156
wherein G is a convolution kernel,
Figure BDA0003747050440000157
Figure BDA0003747050440000158
is the convolution operator.
Since the edge operator is a matrix, the cluster center C 'cannot be expressed' i By applying an edge operator A in the horizontal direction x And edge operator A in the vertical direction y Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient e in the horizontal direction x And edge coefficient e in the vertical direction y And a specific value is obtained, so that the concretization and visualized expression of the edge feature are realized.
Example 5
The SAR image segmentation method based on texture features and SLIC is the same as the embodiment 1-4, and other sequences are added in the step (6) of the invention
Figure BDA0003747050440000161
Tri-valued to obtain a tri-valued sequence
Figure BDA0003747050440000162
Obtaining a ternary sequence H by performing Fast Fourier Transform (FFT) on the ternary sequence k Frequency spectrum FH k In which the ternary sequence H k Frequency spectrum FH k The calculation formula of (c) is:
FH k ＝FFT(H k )
wherein, the FFT is fast Fourier transform;
the FFT is odd, even, or imaginary based on discrete Fourier transformAnd real characteristics, etc., and the multiplication times required by the discrete Fourier transform DFT can be greatly reduced by improving the algorithm of the discrete Fourier transform. The invention uses fast Fourier transform FFT to calculate new sequence H k The frequency spectrum of (2) can reduce the calculation amount of the algorithm and improve the efficiency.
Example 6
The SAR image segmentation method based on texture features and SLIC is the same as the embodiments 1-5, and the method of the invention in step (7) is implemented by performing frequency spectrum FH on all frequency spectrums k Summing and averaging to obtain an average frequency spectrum AMP, judging the convergence of the average frequency spectrum AMP, and obtaining a search area omega, wherein the calculation formula of the AMP is as follows:
Figure BDA0003747050440000163
for all frequency spectrums FH k Summing and averaging to obtain an average frequency spectrum AMP, so that the calculation amount of an algorithm can be reduced; and secondly, according to the convergence of the average frequency spectrum AMP, extreme points of the average frequency spectrum AMP can be obtained, the boundary of the region is obtained, and then the whole search region can be obtained, so that the limitation of manually setting the search region is avoided, the self-adaption of the algorithm is realized, and the segmentation precision of the SAR image is improved.
The invention discloses an SAR image segmentation method based on textural features and SLIC, and belongs to the technical field of image processing. Since the synthetic aperture radar SAR is a coherent system, images shot by the synthetic aperture radar SAR contain speckle noise, which results in low segmentation accuracy. The invention calculates the edge coefficient of the image, and in order to reduce the calculation amount of the algorithm and reduce the influence of noise, the updated clustering center C' i Of different step sizes
Figure BDA0003747050440000164
Ternary to obtain new sequence H k And calculating by using fast Fourier transform to obtain a sequence H k The average frequency spectrum AMP is obtained by analyzing the convergence and the convergence point of the AMP, a search region omega is obtained, and then the texture feature FT of the region omega is obtained; texture feature of the final imageAnd combining FT and SLIC algorithm to obtain a segmented image. The method can effectively reduce the influence of speckle noise on image segmentation, and remarkably improve the segmentation precision of the SAR image. The method is used for SAR image segmentation.
The technical effects of the present invention will be described below with reference to the following experiments:
example 7
SAR image segmentation method based on texture features and SLIC (Long-range computed tomography) as in embodiments 1-6
Conditions of the experiment
The experimental hardware conditions were: intel (R) core (TM) i5-10400F 2.9GHz CPU, memory 16 GB; the experimental software conditions were: matlab 2019 a.
Content of the experiment
Fig. 2 is an original image of an SAR image, which is also an experimental map of the present invention; fig. 2 contains airports, buildings and vegetation, and the size of the image is 640 x 480, which contains a total of 307200 pixels. The method of the invention is adopted to segment the graph 2, firstly, the graph 2 is divided into 2500 superpixel blocks to obtain 2500 clustering centers C i The side length of each super pixel block is 123; at the clustering center C i Calculates gradient values of all pixels in the 3-to-3 neighborhood, and sets the pixel with the minimum gradient value as a new cluster center C' i (ii) a By calculating the clustering center C' i Respectively subjecting the horizontal direction edge operator and the vertical direction edge operator to Gaussian convolution kernels to obtain a clustering center C' i The edge coefficient in the horizontal direction and the edge coefficient in the vertical direction, so that an edge image EI in the figure 2 can be obtained; by calculating each updated cluster center C' i Of different step sizes
Figure BDA0003747050440000171
Will be sequenced
Figure BDA0003747050440000172
Obtaining a new sequence H after the ternary reaction k And calculating to obtain a sequence H by utilizing fast Fourier change k Of all sequences H k Is averaged to obtain a flatA homogeneous spectrum AMP; obtaining a search region omega according to the convergence of AMP, and calculating the texture feature of the region omega; and finally, combining the texture features with the SLIC algorithm based on the SLIC algorithm to obtain the segmentation image map 3 of the figure 2.
Results and analysis of the experiments
FIG. 3 is a graph of the segmentation results of FIG. 2 using the method of the present invention. Segmentation results fig. 3 is divided into two parts: the super pixel blocks in the left half part of the image are surrounded by lines, so that the shapes of the super pixel blocks can be displayed more clearly; the right half part of the image has no lines, and the segmentation quality of the image target boundary can be shown according to the texture intensity of different areas. As can be seen from FIG. 3, the airport of the target area in the image is accurately and independently segmented, the buildings and the vegetation of the non-target area are segmented into a whole, and the boundary area of the airport has no wrong segmentation and wrong segmentation. Comparing the invention with the existing published classical methods, the comparison methods are respectively as follows: SLIC, SNIC, and SEEDS; the adopted evaluation index is a boundary recall ratio (BR), the BR range is 0-1, and the larger the value is, the higher the segmentation precision is. Referring to FIG. 3, the BR of the SLIC process is 0.9016, the BR of the SNIC process is 0.9266, the BR of the SEEDS process is 0.9402, and the BR of the process of the present invention is 0.9623. Therefore, the SAR image segmented by the method has more regular shape and higher compactness of each target region and higher segmentation precision.
In conclusion, the invention provides an SAR image segmentation method based on texture features and SLIC, and solves the technical problem of low segmentation precision caused by noise and complex texture in the SAR image segmentation process. The method comprises the following concrete implementation steps: firstly, dividing an SAR image I into K superpixel blocks; second step updating clustering center C i Obtaining an updated cluster center C' i (ii) a The third step is to obtain an updated clustering center C' i Then, each updated cluster center C 'is calculated' i The edge operator of (1); fourth step, calculating a clustering center C' i Obtaining an edge image EI of the SAR image I; the fifth step is to calculate each updated cluster center C' i Of different step sizes
Figure BDA0003747050440000181
The sixth step is to reduce the calculation amount of the algorithm and the influence of noise by each sequence
Figure BDA0003747050440000182
Tri-valued to obtain a tri-valued sequence H k The frequency spectrum of (a); the seventh step calculates a ternary sequence H k Obtaining a search area Ω; eighthly, calculating the texture feature FT of the search area omega according to the search area omega; and finally, segmenting the SAR image I based on the SLIC algorithm by combining the textural features of the image. Inventive sequences
Figure BDA0003747050440000183
Carrying out ternary operation to reduce noise influence and calculation amount of an algorithm; by means of a ternary sequence H k The convergence of the average frequency spectrum AMP obtains a search region omega, and the texture feature of the region omega is calculated; and obtaining a segmentation result of the SAR image through an SLIC algorithm. Compared with the prior method, the method has the advantages that the method has higher segmentation precision, and the generated super-pixel has higher regularity and compactness.

Claims (6)

1. A SAR image segmentation method based on texture features and SLIC is characterized by comprising the following steps:
(1) dividing the SAR image I into K superpixel blocks: inputting an SAR image I to be segmented, and setting the size of the pixel value of the input SAR image I to be L multiplied by W, wherein N is L multiplied by W, and the pixel matrix of the SAR image I is
Figure FDA0003747050430000011
Wherein, I (x) l ,y u ) The pixel of the ith row and the uth column in the SAR image I is represented, and K original clustering centers C are randomly distributed i ＝[l i ,a i ,b i ,x i ,y i ] T Wherein I denotes the ith cluster center, I is 1, … …, K, and the SAR image I is divided into K super-image blocks, each super-pixel having a side length of
Figure FDA0003747050430000012
(2) Updating the clustering center C i Obtaining updated cluster center C' i : in the 3-3 neighborhood of the original cluster center, the pixel with the minimum gradient value G (x, y) is set as the updated cluster center C' i Wherein i represents the ith cluster center, i ═ 1, … …, K;
(3) calculating each updated cluster center C' i The edge operator of (2): at the cluster center C 'updated with the first one' 1 In the (2S +1) × (2S +1) neighborhood of the central pixel, the neighborhood is divided into four parts, which are sequentially represented from left to right from top to bottom as
Figure FDA0003747050430000013
And
Figure FDA0003747050430000014
the four parts are subjected to horizontal weighted summation to obtain an updated clustering center C' 1 Edge operator of horizontal direction
Figure FDA0003747050430000015
The four parts are vertically weighted and summed to obtain a cluster center C' i Edge operator of vertical direction
Figure FDA0003747050430000016
Similarly, other updated cluster centers C 'are obtained through calculation' i Horizontal edge operator
Figure FDA0003747050430000017
And vertical edge operators
Figure FDA0003747050430000018
(4) Obtaining an edge image EI of the SAR image I: the first updated cluster center C' 1 In the horizontal direction of
Figure FDA0003747050430000019
And the vertical direction
Figure FDA00037470504300000110
Respectively combined with the Gaussian convolution kernels G to obtain the edge coefficient in the horizontal direction
Figure FDA00037470504300000111
And edge coefficient in vertical direction
Figure FDA00037470504300000112
Then the first updated cluster center C' 1 Has an edge coefficient of
Figure FDA0003747050430000021
Similarly, other updated cluster centers C 'are obtained through calculation' i Has an edge coefficient of
Figure FDA0003747050430000022
All the updated edge coefficients of the clustering centers form an edge image EI of the SAR image I;
(5) calculating each updated cluster center C' i Of different step sizes
Figure FDA0003747050430000023
With the first updated cluster center C on the SAR image I 1 ' As starting Point, it is counted as Point (x) 1 ,y 1 ) Point (x) 1 ,y 1 ) The sequence of (A) is
Figure FDA0003747050430000024
Along the y direction, with N/100 as the step size, the second point is obtained as (x) 1 ,y 1+N/100 ) Point (x) 1 ,y 1+N/100 ) The sequence of (A) is
Figure FDA0003747050430000025
At 2 XNStep size of/100, and the third point is obtained as (x) 1 ,y 1+2×N/100 ) Point (x) 1 ,y 1+2×N/100 ) The sequence of (A) is
Figure FDA0003747050430000026
And the analogy is repeated, and the step length of (N-1) multiplied by N/100 is taken as the step length, so that the nth point is obtained as (x) 1 ,y 1+(n-1)×N/100 ) Point (x) 1 ,y 1+(n-1)×N/100 ) Is the sequence of
Figure FDA0003747050430000027
Figure FDA0003747050430000028
Is shown as
Figure FDA0003747050430000029
n represents the sequence of the nth point, and n is 1,2,3, …; similarly, all the cluster centers C 'are obtained through calculation' i Of different step sizes
Figure FDA00037470504300000210
Wherein K is 1,2, … …, K, which indicates the sequence of the kth updated cluster center, and K indicates the number of cluster centers;
(6) each sequence is divided into
Figure FDA00037470504300000211
Tri-valued to obtain a tri-valued sequence H k Spectrum of (a): the first sequence is
Figure FDA00037470504300000212
Tri-valued to obtain a tri-valued sequence of
Figure FDA00037470504300000213
Calculation of a ternary sequence H by means of fast Fourier transform FFT 1 To obtain a ternary sequence H 1 Frequency spectrum FH 1 (ii) a In the same way, other sequences are added
Figure FDA00037470504300000214
Tri-valued to obtain a tri-valued sequence
Figure FDA00037470504300000215
Obtaining a ternary sequence H k Frequency spectrum FH k ；
(7) Computing a ternary sequence H k Obtaining the search area Ω: obtaining a ternary sequence H k Frequency spectrum FH k Then, calculate the frequency spectrum FH k If the AMP is at a point
Figure FDA0003747050430000031
When the point is converged, the farthest point of the search area is the point
Figure FDA0003747050430000032
If AMP is not convergent, the farthest point of the search area is the extreme point of the image gradient, and a search area omega with the farthest point as the radius is obtained;
(8) calculating the texture feature FT of the search region omega: after the search region omega is obtained, calculating the texture feature FT of the region omega,
Figure FDA0003747050430000033
wherein,
Figure FDA0003747050430000034
is the average intensity of the region omega, alpha 4 Denotes the kurtosis, α 4 ＝μ 4 /σ 4 ，μ 4 Fourth moment, δ, representing the mean of the region Ω 4 Is the variance of the region Ω;
(9) segmenting the SAR image I based on the SLIC algorithm: fusing the SLIC algorithm with the texture features, and segmenting the SAR image I;
(9a) setting the pixel with the minimum boundary coefficient e as the clustering center of the boundary coefficients
Figure FDA0003747050430000035
At the cluster center C 'updated with the first one' 1 In the 3 x 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure FDA0003747050430000036
Boundary coefficient cluster center
Figure FDA0003747050430000037
Is (x' 1 ,y′ 1 ) Edge coefficient clustering center
Figure FDA0003747050430000038
Gray value of g 1 Edge coefficient clustering center
Figure FDA0003747050430000039
Is characterized by a texture of FT 1 (ii) a Similarly, the cluster center C 'updated with the other cluster center' i In 3 × 3 neighborhood of the central pixel, the pixel with the minimum boundary coefficient e is set as the clustering center of the boundary coefficient
Figure FDA00037470504300000310
Boundary coefficient cluster center
Figure FDA00037470504300000311
Is (x' i ,y′ i ) Edge coefficient clustering center
Figure FDA00037470504300000312
Gray value of g i Boundary coefficient clustering center
Figure FDA00037470504300000313
Is characterized by a texture of FT i ；
(9b) Calculating the clustering center of pixels and boundary coefficients on the SAR image I
Figure FDA00037470504300000314
The segmented image I' is obtained: calculating the clustering center of pixels and boundary coefficients on the SAR image I
Figure FDA00037470504300000315
The distance between the pixels is D, the pixels are allocated to the nearest boundary coefficient cluster center, wherein D is alpha D w +d g +d t ，d g To represent the distance of the grey value between the pixel on the SAR image I and the cluster center, d t Representing the texture distance between the pixel and the cluster center on the SAR image I, d w Representing the spatial distance between the pixels on the SAR image I and the clustering center, alpha is an influence coefficient determining the quality of the super-pixels, the pixels at the edge and the pixels not positioned at the edge should select different alpha,
Figure FDA00037470504300000316
wherein e is l,u Is a pixel point (x) l ,y u ) The edge coefficient of (e), max (EI), is the maximum value of the edge image EI, and q is 91%, finally obtaining the segmented image I' of the SAR image I.
2. The method for segmenting the SAR image based on the texture feature and the SLIC according to the claim 1, characterized in that the step (2) comprises: updating the clustering center C i Obtaining updated cluster center C' i Therein gradient value G i The formula for the calculation of (x, y) is:
G i (x,y)＝dx i (x,y)+dy i (x,y)
dx i (x, y) is the derivation of the point (x, y) in the x-direction:
dx i (x,y)＝I(x+1,y)-I(x,y)
dy i (x, y) is the derivative of point (x, y) in the y-direction:
dy i (x,y)＝I(x,y+1)-I(x,y)
wherein, I (x, y) represents the gray value of the pixel point (x, y), and (x, y) represents the calculated coordinates of the pixel point.
3. The SAR image segmentation method based on texture features and SLIC of claim 1, wherein the calculation in step (3) obtains other updated cluster centers C' i Horizontal edge operator
Figure FDA0003747050430000041
And vertical edge operator
Figure FDA0003747050430000042
Wherein the edge operator in the horizontal direction
Figure FDA0003747050430000043
And edge operator in vertical direction
Figure FDA0003747050430000044
The calculation formulas are respectively as follows:
Figure FDA0003747050430000045
Figure FDA0003747050430000046
wherein p represents an updated cluster center C' i The p-th part of the neighborhood, p is 1,2,3,4, i is the updated serial number of the cluster center, i is 1, … …, K, H is the S × S matrix,
Figure FDA0003747050430000047
4. the texture feature and SLIC-based SAR image segmentation method as claimed in claim 1, characterized in that the step (4) obtains the edge image EI of the SAR image I, wherein the edge coefficient in horizontal direction
Figure FDA0003747050430000051
And edge coefficient in vertical direction
Figure FDA0003747050430000052
The calculation formulas of (A) and (B) are respectively as follows:
Figure FDA0003747050430000053
Figure FDA0003747050430000054
wherein G is a convolution kernel,
Figure FDA0003747050430000055
Figure FDA0003747050430000056
is the convolution operator.
5. The texture feature and SLIC-based SAR image segmentation method as claimed in claim 1, wherein the step (6) is that the sequence is divided
Figure FDA0003747050430000057
Ternary to obtain new sequence H k Of the new sequence H, wherein k Frequency spectrum FH k The calculation formula of (2) is as follows:
FH k ＝FFT(H k )
wherein the FFT is a fast Fourier transform.
6. The texture feature and SLIC based SAR image segmentation method according to claim 1, wherein said calculating the new sequence H in step (7) k Obtaining the search area Ω, wherein the calculation formula of AMP is:
Figure FDA0003747050430000058
wherein, FH k Is a new sequence H k K denotes the kth sequence, K ∈ [1, K ∈ K]。
CN202210854382.9A 2022-07-14 2022-07-14 SAR image segmentation method based on texture features and SLIC Active CN115131373B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210854382.9A CN115131373B (en) 2022-07-14 2022-07-14 SAR image segmentation method based on texture features and SLIC

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210854382.9A CN115131373B (en) 2022-07-14 2022-07-14 SAR image segmentation method based on texture features and SLIC

Publications (2)

Publication Number Publication Date
CN115131373A true CN115131373A (en) 2022-09-30
CN115131373B CN115131373B (en) 2023-10-27

Family

ID=83384099

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210854382.9A Active CN115131373B (en) 2022-07-14 2022-07-14 SAR image segmentation method based on texture features and SLIC

Country Status (1)

Country Link
CN (1) CN115131373B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116679080A (en) * 2023-05-30 2023-09-01 广州伏羲智能科技有限公司 River surface flow velocity determining method and device and electronic equipment
CN117765008A (en) * 2023-12-26 2024-03-26 西北核技术研究所 An image segmentation method based on adaptive feature perception

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6631212B1 (en) * 1999-09-13 2003-10-07 Eastman Kodak Company Twostage scheme for texture segmentation based on clustering using a first set of features and refinement using a second set of features
CN104794730A (en) * 2015-05-07 2015-07-22 西安电子科技大学 Superpixel-based SAR image segmentation method
KR101694697B1 (en) * 2015-08-03 2017-01-10 안동대학교 산학협력단 IMAGE PARTITIONING METHOD USING SLIC(Simple Linear Iterative Clustering) INCLUDING TEXTURE INFORMATION AND RECORDING MEDIUM
CN109712153A (en) * 2018-12-25 2019-05-03 杭州世平信息科技有限公司 A kind of remote sensing images city superpixel segmentation method
CN110349160A (en) * 2019-06-25 2019-10-18 电子科技大学 One kind is based on super-pixel and fuzzy C-means clustering SAR image segmentation method
CN113313672A (en) * 2021-04-28 2021-08-27 贵州电网有限责任公司 Active contour model image segmentation method based on SLIC superpixel segmentation and saliency detection algorithm

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6631212B1 (en) * 1999-09-13 2003-10-07 Eastman Kodak Company Twostage scheme for texture segmentation based on clustering using a first set of features and refinement using a second set of features
CN104794730A (en) * 2015-05-07 2015-07-22 西安电子科技大学 Superpixel-based SAR image segmentation method
KR101694697B1 (en) * 2015-08-03 2017-01-10 안동대학교 산학협력단 IMAGE PARTITIONING METHOD USING SLIC(Simple Linear Iterative Clustering) INCLUDING TEXTURE INFORMATION AND RECORDING MEDIUM
CN109712153A (en) * 2018-12-25 2019-05-03 杭州世平信息科技有限公司 A kind of remote sensing images city superpixel segmentation method
CN110349160A (en) * 2019-06-25 2019-10-18 电子科技大学 One kind is based on super-pixel and fuzzy C-means clustering SAR image segmentation method
CN113313672A (en) * 2021-04-28 2021-08-27 贵州电网有限责任公司 Active contour model image segmentation method based on SLIC superpixel segmentation and saliency detection algorithm

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
HUANXIN ZOU ET AL: "A PDF-based SLIC superpixel algorithm for SAR images", 2016 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS), pages 6823 - 6826 *
康红宴: "基于SLIC的SAR图像超像素分割方法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》, no. 02, pages 136 - 1981 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116679080A (en) * 2023-05-30 2023-09-01 广州伏羲智能科技有限公司 River surface flow velocity determining method and device and electronic equipment
CN117765008A (en) * 2023-12-26 2024-03-26 西北核技术研究所 An image segmentation method based on adaptive feature perception
CN117765008B (en) * 2023-12-26 2024-09-06 西北核技术研究所 An image segmentation method based on adaptive feature perception

Also Published As

Publication number Publication date
CN115131373B (en) 2023-10-27

Similar Documents

Publication Publication Date Title
CN108830870B (en) Satellite image high-precision farmland boundary extraction method based on multi-scale structure learning
Xiang et al. Superpixel generating algorithm based on pixel intensity and location similarity for SAR image classification
CN107067405B (en) Remote sensing image segmentation method based on scale optimization
CN106886980B (en) Point cloud density enhancement method based on three-dimensional laser radar target identification
CN102402685B (en) Method for segmenting three Markov field SAR image based on Gabor characteristic
CN106251332B (en) SAR image airport target detection method based on edge feature
CN120451655B (en) Image-based 3D point cloud traffic sign classification method and system
CN111145337B (en) Linear array SAR three-dimensional imaging method based on resolution approximation and rapid sparse reconstruction
CN112529910B (en) SAR image rapid superpixel merging and image segmentation method
CN109461127B (en) SAR image sparse regularization feature enhancement method with interpretation as purpose
CN109633646A (en) A Bistatic ISAR Imaging Method Based on Weighted L1 Norm Constraint
CN103871039B (en) Generation method for difference chart in SAR (Synthetic Aperture Radar) image change detection
CN105549009B (en) A kind of SAR image CFAR object detection methods based on super-pixel
CN112946601B (en) Efficient distributed target phase optimization method based on Gauss-Seidel
CN109886218A (en) SAR image Ship Target Detection method based on super-pixel statistics diversity
CN107273813A (en) Geographical space elements recognition system based on high score satellite remote sensing date
CN115131373B (en) SAR image segmentation method based on texture features and SLIC
CN109712149B (en) Image segmentation method based on wavelet energy and fuzzy C-means
CN115951350B (en) Methods, apparatus, equipment and media for extracting permanent scatterer points
CN111738278B (en) Underwater multi-source acoustic image feature extraction method and system
CN119758335B (en) Double-frequency chromatographic SAR precise three-dimensional reconstruction method and system
Tuo et al. Radar forward-looking super-resolution imaging using a two-step regularization strategy
CN117809044A (en) Methods, devices and equipment for array interferometric synthetic aperture radar three-dimensional point cloud denoising
CN113343819B (en) Efficient unmanned airborne SAR image target segmentation method
CN110751652A (en) SAR Image Segmentation Method Based on Bap Distance and Texture Pattern Metrics

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant