A SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity

Abstract

The present invention provides a SAR image segmentation method based on spatial domain statistical similarity and frequency domain texture similarity, comprising quantizing the SAR image, calculating an edge strength map, obtaining an initial segmentation result, forcibly merging small regions, calculating frequency domain texture similarity and a merging cost function, and displaying and comparing the detection results. The present invention solves the problem of high computational complexity in texture region merging in the prior art. The present invention uses a spatial domain to frequency domain transformation to simplify texture similarity, and combines a statistical similarity measure based on the Bhattacharyya distance and an edge penalty term to obtain a new merging cost function, thereby achieving a higher accuracy SAR image segmentation method and reducing computational complexity.

Description

Technical Field

[0001] The present invention relates to the field of SAR image interpretation, and in particular to a SAR image segmentation method based on spatial domain statistical similarity and frequency domain texture similarity. Background Art

[0002] SAR image segmentation is a crucial step in SAR image interpretation. The goal of SAR image segmentation is to simplify a digital image by dividing it into blocks or regions, making it easier to analyze. Pixels belonging to the same region in a segmented image are similar in characteristics such as grayscale and texture, while pixels in adjacent regions have significantly different characteristics. The segmentation boundaries obtained through image segmentation can be used in fields such as 3D reconstruction and target recognition.

[0003] Among existing SAR image segmentation methods, edge-information-guided region merging algorithms have been widely studied due to their robustness. These methods consist of two stages. In the first stage, edge information is extracted from the image, which is then used to divide the SAR image into several small regions, which serve as the initial image segmentation result. The accuracy of the initial segmentation directly determines the correctness of the final merging result. In the second stage, the initial segmentation is iteratively merged according to predefined merging rules until the termination condition is met, resulting in the final image segmentation result. The key to region merging image segmentation algorithms lies in the extraction of edge information in the initial segmentation and the selection of the cost function during the merging process.

[0004] Existing region merging algorithms are exemplified by the relative common boundary length penalty (RCBLP) proposed by Shui et al. This algorithm uses the ratio of sample means within a vector region as a statistical similarity measure, effectively segmenting SAR images. Zhang et al. improved the merging rule based on this approach. While this algorithm achieved improved accuracy compared to previous methods, it ignored the large number of textured regions present in real SAR images, resulting in poor segmentation performance in these areas. To address this issue, Fan et al. proposed the texture pattern similarity measure (TPSM) algorithm, which constructs a merging cost function based on this texture pattern measure and improves the performance of the segmentation algorithm in textured regions. However, this algorithm requires the calculation of the spatial covariance matrix between a large number of pixels, which makes it computationally extremely complex and unable to meet the real-time requirements of practical applications. Summary of the Invention

[0005] The purpose of the present invention is to provide a SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity, so as to solve the problem of high computational complexity of texture region merging in the prior art. The present invention quantizes the SAR image and converts the region originally represented in the spatial domain into the frequency domain using Fourier transform. The texture similarity is simplified by transforming into the frequency domain, and a statistical similarity measure based on the Bhattacharyya distance and an edge penalty term are combined to obtain a new merging cost function, thereby achieving a higher accuracy SAR image segmentation method and reducing computational complexity.

[0006] The technical solutions for achieving the purpose of the present invention are as follows:

[0007] A SAR image segmentation method based on spatial domain statistical similarity and frequency domain texture similarity includes:

[0008] Quantize SAR image: read the SAR image to be segmented and quantize the SAR image into at least 16 gray levels;

[0009] Calculate the edge intensity map: Use the multi-scale Bhattacharyya distance to calculate the SAR image of the quantized gray level to obtain the edge intensity map;

[0010] Obtain the initial segmentation result: Use the watershed algorithm to perform watershed transformation on the edge intensity map to obtain the initial segmentation result of the SAR image;

[0011] Forced merging of small areas: Based on the initial segmentation results, a region adjacency graph is established. Two nodes are connected by a line segment to represent two adjacent regions. The weight of the line segment is used to represent the merging cost between the two adjacent regions. A directed nearest neighbor graph is extracted from the undirected region adjacency graph. Each region in the directed nearest neighbor graph points to the adjacent region corresponding to its minimum cost. The minimum cost value of the two-node ring is found in the directed nearest neighbor graph. If the minimum cost value is less than the preset threshold, the two nodes are merged. Then, the undirected region adjacency graph and the directed nearest neighbor graph are updated and merged in sequence until the cost value between all adjacent regions is greater than the preset threshold. The merge is continued until the number of pixels contained in all regions exceeds 50.

[0012] Calculate frequency domain texture similarity and merging cost function: first calculate the frequency domain texture similarity between adjacent regions, then calculate the merging cost function for all adjacent regions based on the frequency domain texture similarity to obtain the regional merging cost;

[0013] Display and compare detection results: Based on the region merging cost, the merging framework in the edge strength map is used to iteratively merge regions until the merging costs of all adjacent regions are less than a preset threshold. After the merging is completed, the final segmentation result is obtained.

[0014] The aforementioned nodes in the present invention treat each region in the initial segmentation result as a node. The adjacency relationship between regions is represented by a region adjacency graph, where nodes correspond to regions and adjacent regions are connected by line segments. The aforementioned extraction of a directed nearest neighbor graph from an undirected region adjacency graph in the present invention primarily involves adding arrows to the aforementioned line segments, with the arrows pointing from the region to the adjacent region corresponding to its minimum cost.

[0015] As a further improvement of the present invention, the quantized SAR image is an amplitude format SAR image to be segmented read by MATLAB software.

[0016] As a further improvement of the present invention, the specific calculation formula for obtaining the edge intensity map is as follows:

[0017]

[0018] Among them, S, M and L represent small scale, medium scale and large scale respectively, α(s) represents the weight of small scale, medium scale and large scale, satisfying α(S)+α(M)+α(L)=1, angle θ=kπ / K, k=0,1,…,K-1, K is the number of directions, usually taken as 8.

[0019] As a further improvement of the present invention, the obtaining of the initial segmentation result is specifically as follows:

[0020] The watershed algorithm in MATLAB software is used to perform watershed transformation on the edge intensity map to obtain the initial segmentation result of the SAR image.

[0021] As a further improvement of the present invention, the adjacent regions are represented as Ω1 and Ω2, and the frequency domain texture similarity between the two adjacent regions Ω1 and Ω2 is defined as follows:

[0022]

[0023] Among them, F1 and F2 represent the results of transforming the maximum inscribed square support area of ​​regions Ω1 and Ω2 into the frequency domain, respectively. The higher the texture similarity of the two regions, the smaller the FTSM value.

[0024] As a further improvement of the present invention, the FTSM process of the area Ω to be calculated includes: first, using a square of size 11×11 to erode it, and the eroded area is recorded as Ω * , then traverse Ω * For the pixels in , select an 11×11 image block with each pixel as the center, and transform the image block into the frequency domain as the frequency domain matrix at the pixel. The frequency domain representation of Ω is as follows:

[0025]

[0026] Where R(x,y) represents the square matrix centered at point (x,y), #Ω * It represents the number of pixels in the eroded area, and FFT2 represents the two-dimensional fast Fourier transform.

[0027] As a further improvement of the present invention, the calculation formula for obtaining the region merging cost is:

[0028]

[0029] Among them, the first item is the statistical similarity based on the Bhattacharyya coefficient, the second item is the boundary penalty item proposed in the RCBLP method, and the third item is the frequency domain texture similarity, where FTSM takes FTSM1 and FTSM2.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. The present invention uses a spatial-to-frequency domain transformation to simplify texture similarity, and combines a statistical similarity metric based on the Bhattacharyya distance with an edge penalty term to obtain a new merged cost function. This solves the problem of poor efficiency of the existing TPSM algorithm in segmenting texture regions, thereby improving the efficiency of SAR image texture region segmentation.

[0032] 2. The present invention adopts edge extraction based on multi-scale Bhattacharyya distance and adds a forced small region merging strategy before region merging to solve the problem of overly fragmented initial segmentation obtained based on the rectangular double-window mean ratio, thereby achieving efficient and accurate segmentation of SAR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of a SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity provided by the present invention;

[0034] Figure 2 Schematic diagram of the regional adjacency graph and nearest neighbor graph of the simulation image;

[0035] Figure 3 To update the image after the region adjacency graph and the nearest neighbor graph are merged;

[0036] Figure 4 A flowchart of an application example of a SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity provided by the present invention;

[0037] Figure 5 This is a schematic diagram of the initial segmentation of the synthetic cartoon SAR image;

[0038] Figure 6 Comparison of the segmentation results of synthetic texture SAR images using several methods;

[0039] Figure 7 This is a diagram of the segmentation results of real-scene SAR images. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in this field based on these embodiments are all within the scope of protection of the present invention.

[0041] Example 1:

[0042] See also Figure 1 The embodiment of the present invention provides a SAR image segmentation method based on spatial statistical similarity and frequency texture similarity, comprising:

[0043] Step S110 , quantizing the SAR image: reading the SAR image to be segmented, and quantizing the SAR image into at least 16 gray levels.

[0044] In step S110 , the SAR image is quantized by using MATLAB software to read the amplitude format SAR image to be segmented.

[0045] Step S120, calculating an edge intensity map: using a multi-scale Bhattacharyya distance to calculate the SAR image of the quantized grayscale level to obtain an edge intensity map.

[0046] In step S120, the specific calculation formula for obtaining the edge strength map is as follows:

[0047]

[0048] Among them, S, M and L represent small scale, medium scale and large scale respectively, α(s) represents the weight of small scale, medium scale and large scale, satisfying α(S)+α(M)+α(L)=1, angle θ=kπ / K, k=0,1,…,K-1, K is the number of directions, usually taken as 8.

[0049] Preferably, in step S120 of the embodiment of the present invention, the multi-scale Bhattacharyya distance used to calculate the edge intensity map of the image is calculated using the following formula (21):

[0050]

[0051] In formula (21), i=1,2,…,16 represents the gray level currently being calculated. There are 16 gray levels in total. u(x,y|s,θ) (i) and h l(x,y|s,θ) (i) represents the ratio of pixels of the i-th gray level to the total pixels in the rectangular upper window and rectangular lower window with a scale of s and a rotation angle of θ centered on the pixel to be detected n.

[0052] Step S130 , obtaining an initial segmentation result: performing a watershed transformation on the edge intensity map using a watershed algorithm to obtain an initial segmentation result of the SAR image.

[0053] In step S130 , the initial segmentation result is obtained by performing a watershed transformation on the edge intensity map using the watershed algorithm provided in the MATLAB software to obtain the initial segmentation result of the SAR image.

[0054] Step S140, forced merging of small areas: establishing a region adjacency graph based on the initial segmentation result, connecting two nodes with a line segment to represent two adjacent areas, using the weight of the line segment to represent the merging cost between the two adjacent areas, extracting a directed nearest neighbor graph from the undirected region adjacency graph, each area in the directed nearest neighbor graph points to the adjacent area corresponding to its minimum cost; finding the minimum cost value of the two-node ring in the directed nearest neighbor graph, merging if the minimum cost value is less than a preset threshold, and then updating the undirected region adjacency graph and the directed nearest neighbor graph, merging in sequence until the cost values ​​between all adjacent areas are greater than the preset threshold; merging in sequence until the number of pixels contained in all areas exceeds 50;.

[0055] In step S140, a region adjacency graph is established based on the initial segmentation result obtained in step S130 (the region adjacency graph in step S130 is an undirected region adjacency graph, RAG). The region adjacency graph uses a line segment to connect two nodes to represent two adjacent regions, and the weight of the line segment is used to represent the merging cost between the two regions.

[0056] A directed nearest neighbor graph (NNG) is extracted from the region adjacency graph created from the initial segmentation results. Each region in the nearest neighbor graph points to the adjacent region with the minimum cost. A large number of adjacent regions with mutually minimum costs in the image are called two-node rings. The minimum cost of the two-node ring is found in the nearest neighbor graph. If the cost is less than a preset threshold, the two-node rings are merged. The region adjacency graph and the nearest neighbor graph are then updated, and the two regions are merged sequentially until the cost between all adjacent regions exceeds the preset threshold, completing the forced merging of small regions.

[0057] In step S140 of the embodiment of the present invention, the merging cost used for forcibly merging the small regions is preferably composed of a similarity measure and a shape cost. The similarity measure υ(Ω1, Ω2) of two adjacent small regions Ω1 and Ω2 is calculated using formula (41) as follows:

[0058]

[0059] In formula (41), and denotes the mean of the pixel amplitudes of regions Ω1 and Ω2, #Ω1 and #Ω2 denote the number of pixels within regions Ω1 and Ω2, and L denotes the number of views of the image. The more similar the two regions are, the smaller the similarity measure in the above formula, indicating that the two regions should be merged. Combining the above formula with the shape cost yields the cost function of formula (42):

[0060]

[0061] In formula (42), the second term is the shape cost, It represents the length of the common boundary between two adjacent regions. The parameter η is an adjustable parameter and is empirically selected as 1.2.

[0062] In actual use, the process of forcibly merging small areas in the embodiment of the present invention is exemplified as follows: Figure 2 The figure shows the RAG and NNG schematic diagram of a simulation image ( Figure 2 (a) represents the original image, Figure 2 (b) RAG diagram showing the image, Figure 2 (c) shows the NNG diagram of the image. The cost of the adjacent regions is only used for illustration and does not represent the actual cost. Assuming the preset threshold is 0, searching the NNG extracted from the RAG shows that the two-node ring cost of region 1 and region 2 is the smallest. Therefore, region 1 and region 2 are merged first, and the RAG and NNG are updated at the same time. The merged image is as follows: Figure 3 As shown ( Figure 3 (a) represents the original image, Figure 3 (b) RAG diagram showing the image, Figure 3 (c) Schematic diagram of the NNG of the image. After the information is updated, regions 3 and 4 are merged in the next iteration, and the process is recursive until the minimum cost is greater than 0, at which point the merger is complete.

[0063] Step S150 , calculating frequency domain texture similarity and merging cost function: first calculating frequency domain texture similarity between adjacent regions, then calculating merging cost function for all adjacent regions based on frequency domain texture similarity to obtain region merging cost.

[0064] In step S150, adjacent regions are represented as Ω1 and Ω2, and the frequency domain texture similarity between two adjacent regions Ω1 and Ω2 is defined as follows:

[0065]

[0066] Among them, F1 and F2 represent the results of transforming the maximum inscribed square support area of ​​regions Ω1 and Ω2 into the frequency domain, respectively. The higher the texture similarity of the two regions, the smaller the FTSM value.

[0067] The FSM process of the area Ω to be calculated includes: first, using a square of size 11×11 to erode it, and the area after the erosion is recorded as Ω * , then traverse Ω * For the pixels in , select an 11×11 image block with each pixel as the center, and transform the image block into the frequency domain as the frequency domain matrix at the pixel. The frequency domain representation of Ω is as follows:

[0068]

[0069] Where R(x,y) represents the square matrix centered at point (x,y), #Ω * It represents the number of pixels in the eroded area, and FFT2 represents the two-dimensional fast Fourier transform.

[0070] In step S150, the calculation formula for the region merging cost is obtained as follows:

[0071]

[0072] Among them, the first item is the statistical similarity based on the Bhattacharyya coefficient, the second item is the boundary penalty item proposed in the RCBLP method, and the third item is the frequency domain texture similarity, where FSM takes FSM1 and FSM2.

[0073] Step S160, display and compare the detection results: according to the region merging cost, use the merging framework in the edge strength map to iteratively merge regions until the merging costs of all adjacent regions are less than the preset threshold. After the merging is completed, the final segmentation result is obtained.

[0074] Based on the region merging costs calculated in step S150, the merging framework in the calculated edge strength map is used to iteratively merge regions until the merging costs of all adjacent regions are less than a preset threshold. Once the merging is complete, the final segmentation result is obtained. The SAR image segmentation results are displayed using MATLAB software, and the segmentation accuracy and computational complexity of this method are compared with those of existing RCBLP, OESW, and TPSM methods on simulated textured SAR images and real SAR images.

[0075] In step S160 of the embodiment of the present invention, BD(Ω1, Ω2) in the first term of the combined cost function is specifically performed using the following formula (61):

[0076]

[0077] Formula (61) represents the Bhattacharyya distance between regions Ω1 and Ω2, and It represents the proportion of pixels of the i-th gray level in two regions of the quantized image. The more similar the two regions are, the smaller the Bhattacharyya distance is.

[0078] In this embodiment of the present invention, the aforementioned nodes refer to treating each region in the initial segmentation result as a node. The adjacency relationship between regions is represented by a region adjacency graph, where nodes correspond to regions and adjacent regions are connected by line segments. In this embodiment of the present invention, extracting a directed nearest neighbor graph from an undirected region adjacency graph primarily involves adding arrows to the aforementioned line segments, with the arrows pointing from the region to the adjacent region corresponding to its minimum cost.

[0079] This embodiment of the present invention uses a spatial-to-frequency domain transformation to simplify texture similarity. Combining this with a statistical similarity metric based on the Bhattacharyya distance and an edge penalty term, a new merging cost function is derived. This solves the problem of poor efficiency in segmenting texture regions in existing TPSM algorithms, thereby improving the efficiency of SAR image texture region segmentation. This embodiment of the present invention employs edge extraction based on the multi-scale Bhattacharyya distance and adds a forced small region merging strategy before region merging, addressing the problem of overly fragmented initial segmentation based on the rectangular double-window mean ratio. This allows for efficient and accurate SAR image segmentation.

[0080] Example 2:

[0081] Based on the solution disclosed in Example 1, this embodiment uses a SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity. Figure 4 , the implementation steps of this embodiment are as follows:

[0082] Step 1: Quantize the input SAR image: Use MATLAB software to read the amplitude format SAR image to be segmented and quantize the image into 16 gray levels;

[0083] Step 2: Calculate the edge strength map: Use the existing multi-scale Bhattacharyya distance to calculate the edge strength map (ESM) of the image. Formula (1) is used as follows:

[0084]

[0085] In formula (1), S, M, and L represent small scale, medium scale, and large scale, respectively. α(s) represents the weight of small scale, medium scale, and large scale, satisfying α(S)+α(M)+α(L)=1. The angle θ=kπ / K, k=0,1,…,K-1, and K is the number of directions, usually 8. θ,s (n) represents the Bhattacharyya distance at the pixel to be detected, which is calculated using formula (2) as follows:

[0086]

[0087] In formula (2), i=1,2,…,16 represents the gray level currently being calculated. There are 16 gray levels in total. u(x,y|s,θ) (i) and h l(x,y|s,θ) (i) represents the ratio of pixels of the i-th gray level to the total pixels in the rectangular upper window and rectangular lower window with a scale of s and a rotation angle of θ centered on the pixel to be detected n.

[0088] Step 3: Watershed transform to obtain initial segmentation: The ESM obtained in step 2 is subjected to watershed transform using the watershed algorithm provided in MATLAB software to obtain the initial segmentation result of the SAR image.

[0089] Step 4. Force merge small regions: Based on the initial segmentation result obtained in step 3, a region adjacency graph (RAG) is established. RAG is an undirected region adjacency graph. Two adjacent regions are represented by connecting two nodes with a line segment, and the weight of the line segment is used to represent the merging cost between the two regions. A directed nearest neighbor graph (NNG) can be extracted from RAG. Each region in NNG points to the adjacent region corresponding to its minimum cost. There may be a large number of adjacent regions with minimum costs to each other in the image. Such regions are called two-node rings. The minimum cost value of the two-node ring is found in NNG. If the cost value is less than the preset threshold, it is merged. Then the RAG and NNG are updated and merged in sequence until the cost value between all adjacent regions is greater than the preset threshold. At this point, the forced merger of small regions is completed.

[0090] The merging cost consists of a similarity measure and a shape cost. The similarity measure υ(Ω1,Ω2) of two adjacent small regions Ω1 and Ω2 is calculated using formula (3) as follows:

[0091]

[0092] In formula (3), and denotes the mean of the pixel amplitudes of regions Ω1 and Ω2, #Ω1 and #Ω2 denote the number of pixels within regions Ω1 and Ω2, and L denotes the number of views of the image. The more similar the two regions are, the smaller the similarity measure in the above formula, indicating that the two regions should be merged. Combining formula (3) with the shape cost yields the cost function of formula (4):

[0093]

[0094] In formula (4), the second term is the shape cost, It represents the length of the common boundary between two adjacent regions. The parameter η is an adjustable parameter and is empirically selected as 1.2.

[0095] The following takes a simulation image as an example to illustrate the merging process. Figure 2 A schematic diagram of RAG and NNG of a simulated image is given, where the cost of adjacent regions is only used for illustration and does not represent the actual cost ( Figure 2 (a) represents the original image, Figure 2 (b) RAG diagram showing the image, Figure 2 (c) Schematic diagram of NNG of the image. Assuming the preset threshold is 0, searching the NNG extracted from RAG shows that the two-node ring cost of region 1 and region 2 is the smallest. Therefore, region 1 and region 2 are merged first, and RAG and NNG are updated at the same time. The merged image is as follows: Figure 3 As shown ( Figure 3 (a) represents the original image, Figure 3 (b) RAG diagram showing the image, Figure 3 (c) Schematic diagram of the NNG of the image. After the information is updated, regions 3 and 4 are merged in the next iteration, and the process is recursive until the minimum cost is greater than 0, at which point the merger is complete.

[0096] Step 5: Calculate the frequency domain texture similarity between adjacent regions: The frequency domain texture similarity measure (FSM) between two adjacent regions Ω1 and Ω2 is defined using formula (5) as follows:

[0097]

[0098] In formula (5), F1 and F2 represent the results of transforming the maximum inscribed square support area of ​​regions Ω1 and Ω2 into the frequency domain, respectively. The higher the texture similarity of the two regions, the smaller the FTSM value. Here, the maximum inscribed square of the region is used to approximate the original region, and the obtained texture similarity is recorded as FSM1. This approximation is to improve computational efficiency, and the accuracy of the method is sacrificed to a certain extent. In addition, another more accurate method is also provided here, recorded as FSM2. For the region Ω to be calculated for FSM, first use an 11×11 square to corrode it. The corroded area is recorded as Ω*. Then, traverse the pixels in Ω*, select an 11×11 image block with each pixel as the center, and transform the image block into the frequency domain as the frequency domain matrix at the pixel. Then, the frequency domain of Ω is expressed as follows using formula (6):

[0099]

[0100] In formula (6), R(x,y) represents the square matrix centered at point (x,y), #Ω *It represents the number of pixels in the eroded area, and FFT2 represents the two-dimensional fast Fourier transform.

[0101] Step 6: Calculate the merging cost function based on frequency domain texture similarity: Based on step 5, calculate the merging cost function for all adjacent regions using the following formula (7):

[0102]

[0103] In formula (7), the first term is the statistical similarity based on the Bhattacharyya coefficient, using the following formula (8):

[0104]

[0105] Formula (8) represents the Bhattacharyya distance between regions Ω1 and Ω2, and represents the proportion of pixels of the i-th grayscale level in two regions of the quantized image. The more similar the two regions are, the smaller the Bhattacharyya distance. The second term is the boundary penalty term proposed in the RCBLP method, and the third term is the frequency domain texture similarity proposed in step 5. The FSM can be FSM1 or FSM2.

[0106] Step 7: Iteratively merge regions: Based on the region merging cost calculated in the previous step, iteratively merge regions according to the process in step 4 until the merging cost of all adjacent regions is less than the preset threshold. After the merging is completed, the final segmentation result is obtained.

[0107] Step 8. Display and compare the detection results: Use MATLAB software to display the SAR image segmentation results, and compare the segmentation accuracy and computational complexity of this method with the existing RCBLP, OESW and TPSM methods on simulated texture SAR images and real SAR images.

[0108] Based on steps 1 to 7, this embodiment implements a SAR image segmentation method based on spatial domain statistical similarity and frequency domain texture similarity.

[0109] Example 3:

[0110] Based on the solutions disclosed in Examples 1 and 2, this embodiment simulates a SAR image segmentation method based on spatial statistical similarity and frequency texture similarity as follows:

[0111] The data used in the simulation experiment included both simulated and real images. The simulated images were two simulated three-view SAR images. Image 1 was simulated by adding speckle noise to a cartoon image, while Image 2 was a simulated image composed of five different textures. The real images were three real SAR images. Image 1 was a single-view SAR image measuring 401×401 and containing elements such as street lights. Images 2 and 3 were three-view SAR images with relatively low noise levels and containing elements such as farmland and residential areas. They were sized 501×501 and 512×512, respectively. Both images contained some difficult-to-segment texture areas, providing a good demonstration of the texture segmentation capabilities of different methods.

[0112] Simulation experiment content: In the simulation experiment, simulated images are used to verify the effectiveness of the initial segmentation of the embodiment of the present invention and the overall efficiency of the algorithm, and real images are used to verify the accuracy of the segmentation results of the embodiment of the present invention.

[0113] Simulation experiment 1: For the simulation image 1, the method of the embodiment of the present invention is used to perform initial segmentation, and compared with the existing rectangular double window mean ratio and super pixel algorithm, the initial segmentation result is as follows: Figure 5 As shown, Figure 5 in Figure 5 (a) represents the original image, Figure 5 (b) shows the result of the rectangular double window mean ratio algorithm, Figure 5 (c) represents the result of the method of the present invention, Figure 5 (d) shows the result of superpixel algorithm. Figure 5 The boundaries of the regions are marked in . The method based on the rectangular double-window mean ratio generates 3148 regions, and the method according to the embodiment of the present invention generates 845 regions. While retaining all the correct region boundaries, the method according to the embodiment of the present invention effectively reduces the redundancy of the initial segmentation. Although the initial segmentation time of the method according to the embodiment of the present invention is slightly longer than that of the method according to the rectangular double-window mean ratio, the improvement in accuracy brought about by this also brings great advantages to the subsequent region merging. By observing the results of the superpixel initial segmentation method, it can be seen that this method misses a lot of correct region boundaries, and the segmentation results do not match the real boundaries, which cannot be corrected in the subsequent merging. In addition, the computational complexity of the method according to the embodiment of the present invention is extremely high, so it is not suitable for initial segmentation.

[0114] Simulation experiment 2: The simulation image 2 is segmented using the method of the embodiment of the present invention and compared with the existing RCBLP, OESW, and TPSM methods. The segmentation results are shown in the figure below. Figure 6 As shown, Figure 6 (a) represents the original image, Figure 6 (b) shows the results of the RCBLP method, Figure 6 (c) shows the results of the OESW method, Figure 6(d) shows the results of the TPSM method, Figure 6 (e) shows the result of FSM1 method, Figure 6 (f) shows the results of the FSM2 method.

[0115] like Figure 6 As shown, neither RCBLP nor OESW can separate regions with different textures, with RCBLP performing worse. OESW can segment the circular region in the center. TPSM's segmentation result is relatively accurate, but the segmentation lines of the two regions on the right do not align closely with the true boundaries. The merging method proposed in this embodiment of the present invention achieves very accurate segmentation results. While there are minor errors in other areas, it performs better overall than the other methods. The efficiency of several methods is compared below: RCBLP, OESW, TPSM, TFSM1, and TFSM2 take 0.70s, 13.15s, 40.73s, 13.81s, and 20.24s to segment this image, respectively. It can be seen that while RCBLP has the lowest accuracy, it has the highest efficiency due to its simple merging cost form. OESW has slightly better accuracy and relatively higher efficiency. While TPSM has good accuracy, its computational complexity is too high. While maintaining comparable accuracy, the method in this embodiment of the present invention doubles efficiency, with TFSM1 being slightly more efficient than TFSM2, which is consistent with the previous analysis.

[0116] Simulation experiment 3: The real images 1-3 are segmented using the method of the embodiment of the present invention and compared with the existing IRGS, RCBLP, OESW, TPSM and methods. The segmentation results are shown in the figure. Figure 7 As shown, Figure 7 (a) shows the segmentation result on the real image 1, Figure 7 (b) shows the segmentation result on the real image 2. Figure 7 (c) shows the segmentation result on the real image 3. IRGS is a method based on region growing. The edges of its segmentation results are very uneven and there are many small areas. This also makes its segmentation accuracy on real SAR images relatively general. RCBLP can segment more accurately in some homogeneous areas such as farmland and streets, but for images such as the real image 2 that contain large texture areas in the middle, the under-segmentation of the entire image is more obvious. The segmentation results of OESW are relatively accurate overall, but because there is no measurement to distinguish textures, it does not accurately separate the large texture areas of the real image 2. The accuracy of TPSM and the method of the present invention is good.

[0117] This embodiment of the present invention employs edge extraction based on the multi-scale Bhattacharyya distance and adds a forced small region merging strategy before region merging, effectively reducing the redundancy of the initial segmentation and improving its accuracy. It also employs a spatial-to-frequency domain transformation to simplify texture similarity. Combining a statistical similarity metric based on the Bhattacharyya distance with an edge penalty term, a new merging cost function is derived. This overcomes the inefficiency of existing TPSM algorithms in segmenting texture regions, thereby improving the efficiency of texture region segmentation in SAR images.

[0118] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

[0119] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0120] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A SAR image segmentation method based on spatial statistical similarity and frequency texture similarity, characterized in that: include: Quantize SAR image: read the SAR image to be segmented and quantize the SAR image into at least 16 gray levels; Calculate the edge intensity map: Use the multi-scale Bhattacharyya distance to calculate the SAR image of the quantized gray level to obtain the edge intensity map; Obtain the initial segmentation result: Use the watershed algorithm to perform watershed transformation on the edge intensity map to obtain the initial segmentation result of the SAR image; Forced merging of small regions: Based on the initial segmentation results, a region adjacency graph is established. Two adjacent regions are represented by connecting two nodes with a line segment. The weight of the line segment represents the merging cost between the two adjacent regions. A directed nearest neighbor graph is extracted from the undirected region adjacency graph. In the directed nearest neighbor graph, each region points to the adjacent region corresponding to its minimum cost. Find the minimum cost of a two-node ring in a directed nearest neighbor graph. If the minimum cost is less than a preset threshold, merge them. Then update the undirected region adjacency graph and the directed nearest neighbor graph, merging them in sequence until the cost between all adjacent regions is greater than the preset threshold. The two-node ring refers to a large number of adjacent regions in the SAR image that have the minimum cost to each other. Calculate frequency domain texture similarity and merging cost function: first calculate the frequency domain texture similarity between adjacent regions, then calculate the merging cost function for all adjacent regions based on the frequency domain texture similarity to obtain the regional merging cost; Display and compare detection results: Based on the region merging cost, iteratively merge regions using the merging framework in the calculated edge strength map until the merging cost of all adjacent regions is less than a preset threshold. After the merging is completed, the final segmentation result is obtained; The adjacent regions are represented as Ω1 and Ω2, and the frequency domain texture similarity between two adjacent regions Ω1 and Ω2 is defined as follows: Where F1 and F2 represent the results of transforming the maximum inscribed square support area of ​​regions Ω1 and Ω2 into the frequency domain. The higher the texture similarity between the two regions, the smaller the FTSM value. The FSM process of the area Ω to be calculated includes: first, using a square of size 11×11 to erode it, and the area after the erosion is recorded as Ω * , then traverse Ω * For the pixels in , select an 11×11 image block with each pixel as the center, and transform the image block into the frequency domain as the frequency domain matrix at the pixel. Then the frequency domain representation of Ω is as follows: Where R(x,y) represents the square matrix centered at point (x,y), #Ω * Indicates the number of pixels in the eroded area, and FFT2 indicates two-dimensional fast Fourier transform; The calculation formula for obtaining the region merging cost is: Among them, the first term is the statistical similarity based on the Bhattacharyya coefficient, the second term is the boundary penalty term proposed in the RCBLP method, and the third term is the frequency domain texture similarity, where FSM takes FSM1 and FSM2; The BD(Ω1,Ω2) in the first term of the merge cost function in the merge cost is specifically performed by the following formula (61): Formula (61) represents the Bhattacharyya distance between regions Ω1 and Ω2, and It represents the proportion of pixels of the i-th gray level in two regions of the quantized image. The more similar the two regions are, the smaller the Bhattacharyya distance is.

2. The SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity according to claim 1, characterized in that: The quantized SAR image is an amplitude format SAR image to be segmented that is read using MATLAB software.

3. The SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity according to claim 1, characterized in that: The specific calculation formula for obtaining the edge intensity map is as follows: Among them, S, M and L represent small scale, medium scale and large scale respectively, α(s) represents the weight of small scale, medium scale and large scale, satisfying α(S)+α(M)+α(L)=1, angle θ=kπ / K, k=0,1,…,K-1, K is the number of directions, usually taken as 8.

4. The SAR image segmentation method based on spatial statistical similarity and frequency domain texture similarity according to claim 1, characterized in that: The specific steps of obtaining the initial segmentation result are as follows: The watershed algorithm in MATLAB software is used to perform watershed transformation on the edge intensity map to obtain the initial segmentation result of the SAR image.

Images