Fuzzy iterative SAR image superpixel segmentation method fusing edge information

Abstract

The application provides a fuzzy iterative SAR image superpixel segmentation method fusing edge information, which is based on a local iterative clustering principle, uses an edge detection algorithm of a SAR image to obtain edge information in the image, optimizes an initial clustering center step, updates a distance measurement method, introduces a fuzzy theory to form a fuzzy iterative clustering, and finally obtains a superpixel segmentation result with good boundary preservation and pixel identity through a post-processing step of fusing the edge information. Compared with other superpixel segmentation methods, the application mainly faces a target detection task of a SAR image under a complex condition, and especially when the sizes of targets in the SAR image greatly differ, existing superpixel segmentation methods cannot well keep the boundaries of small targets and have a problem of integrating small targets into superpixels, and the application introduces edge information to obtain an ideal superpixel segmentation result.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and relates to SAR images, specifically to a fuzzy iterative SAR image superpixel segmentation method that fuses edge information. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active type of microwave imaging sensor, can operate in all weather conditions, regardless of weather conditions. SAR plays an increasingly important role in both civilian and military fields, with target detection being a key application.

[0003] In SAR images, superpixels are small regions composed of a series of adjacent pixels with similar intensity, texture, and other characteristics. These superpixels retain valuable information for further target detection and generally do not destroy the boundary information of targets in the image. Replacing a large number of pixels with a small number of superpixels to represent image features significantly reduces the complexity of image processing.

[0004] Currently, superpixel segmentation methods for SAR images are mainly used to solve SAR image classification problems. However, when dealing with target detection in SAR images with complex backgrounds, especially when the target sizes differ significantly, existing superpixel segmentation methods cannot maintain the independence of small targets when the superpixel size S is large. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a fuzzy iterative SAR image superpixel segmentation method that integrates edge information, thereby solving the technical problem that existing superpixel segmentation methods cannot effectively maintain the boundaries of small targets when the target sizes differ significantly.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for superpixel segmentation of fuzzy iterative SAR images by fusing edge information, the method comprising the following steps:

[0008] Step 1: Use an edge detector based on a Gaussian gamma-shaped dual window to perform edge detection on the SAR image to obtain an edge enhancement map (ESM) and an edge map (EM).

[0009] Step 2: Based on the set superpixel size S, initialize the cluster centers using the edge enhancement map ESM and the edge map EM;

[0010] Step 3: Calculate the metric distance D(i,j) between the cluster center and each pixel within the 2s×2S neighborhood of the cluster center;

[0011] Step 4: Metric distance comparison update. Each pixel has a metric distance with its surrounding cluster centers. The smaller l metric distances between each pixel and its surrounding cluster centers are retained.

[0012] Step 5: Calculate the membership degree U and cluster center C;

[0013] Step 6: Repeat steps 3, 4, and 5 until the set number of iterations is reached;

[0014] Step 7: Select the cluster center category with the highest pixel membership degree as the superpixel category of the pixel to obtain the initial superpixel segmentation result;

[0015] Step 8: Use the edge information of the edge map EM for post-processing to obtain the final superpixel segmentation result.

[0016] Compared with the prior art, the present invention has the following technical effects:

[0017] (I) The superpixel segmentation effect of the present invention is good. The present invention uses a ratio-based Gaussian gamma-shaped dual-window edge detector to obtain edge information in the image, optimizes the initialization of cluster centers, updates the distance metric method, and introduces fuzzy theory to form a fuzzy local iterative clustering principle. The post-processing step fuses edge information, and finally obtains a superpixel segmentation result with good boundary preservation and pixel identity.

[0018] (II) This invention is applicable to detection tasks with complex scenes and large differences in target size. Currently, superpixel segmentation methods for SAR images are mainly geared towards SAR image classification tasks. In complex situations, especially when the target sizes differ greatly, existing superpixel segmentation methods cannot maintain the boundaries of small targets well, resulting in the problem of small targets being incorporated into superpixels. This invention utilizes edge information to maintain the independence of small targets, making it suitable for detection tasks with complex scenes and large differences in target size.

[0019] (III) The method of the present invention can be used as a superpixel segmentation step in a superpixel-based target detection algorithm, providing excellent superpixel segmentation results in complex situations, especially when the target size difference is large. Attached Figure Description

[0020] Figure 1 This is a flowchart of the fuzzy iterative SAR image superpixel segmentation method that fuses edge information according to the present invention.

[0021] Figure 2(a) is a simulation diagram of the SAR image of Simulation 1.

[0022] Figure 2(b) is a simulation diagram of the SLIC segmentation result in Simulation 1.

[0023] Figure 2(c) is a simulation diagram of the SLIC-EDGE segmentation result of Simulation 1.

[0024] Figure 2(d) is a simulation diagram of the segmentation results of SLIC-GLR in Simulation 1.

[0025] Figure 2(e) is a simulation diagram of the SREP segmentation results of Simulation 1.

[0026] Figure 2(f) is a simulation diagram of the superpixel segmentation result of the present invention in Simulation 1.

[0027] Figure 3(a) is a simulation diagram of the SAR image of Simulation 2.

[0028] Figure 3(b) is a simulation diagram of the SLIC segmentation result in Simulation 2.

[0029] Figure 3(c) is a simulation diagram of the SLIC-EDGE segmentation result in Simulation 2.

[0030] Figure 3(d) is a simulation diagram of the SLIC-GLR segmentation results in Simulation 2.

[0031] Figure 3(e) is a simulation diagram of the SREP segmentation results of Simulation 2.

[0032] Figure 3(f) is a simulation diagram of the superpixel segmentation results of the present invention in Simulation 2.

[0033] The specific content of the present invention will be further explained in detail below with reference to the embodiments. Detailed Implementation

[0034] It should be noted that, unless otherwise specified, all devices and algorithms in this invention are based on devices and algorithms known in the prior art.

[0035] SAR refers to Synthetic Aperture Radar.

[0036] SLIC stands for Simple Linear Iterative Clustering Algorithm (doi:10.1109 / TPAMI.2012.120).

[0037] SLIC-EDGE refers to the synthetic aperture radar image superpixel segmentation algorithm that fuses edge information (doi:10.3969 / j.issn.1671-1815.2020.24.038).

[0038] SLIC-GLR refers to a single-polarization SAR image superpixel segmentation algorithm based on Gamma distribution (doi:10.1109 / LGRS.2018.2805714).

[0039] SREP refers to the SAR image superpixel segmentation algorithm based on Euclidean distance similarity ratio (doi:10.1049 / el.2016.0020).

[0040] In superpixel-based target detection methods, the boundary preservation and identity of superpixels are crucial to the final target detection result, while the number of superpixels has a significant impact on the efficiency of target detection. Therefore, excellent superpixel segmentation methods can greatly improve the performance of target detection in SAR images.

[0041] Compared to existing superpixel segmentation methods for SAR images, this invention is mainly aimed at SAR image target detection tasks. It addresses the problem that existing superpixel segmentation methods cannot well preserve the boundaries of small targets and have the problem of small targets being incorporated into superpixels when the targets are in complex situations, especially when the target sizes differ greatly.

[0042] This invention discloses a fuzzy iterative SAR image superpixel segmentation method that fuses edge information. The general framework of the method is consistent with the Simple Linear Iterative Clustering (SLIC) algorithm. Based on the principle of local iterative clustering, this method utilizes SAR image edge detection algorithms to obtain edge information in the image, optimizes the initialization of cluster centers, updates the distance metric, introduces fuzzy theory to construct fuzzy iterative clustering, and finally, through a post-processing step that fuses edge information, obtains a superpixel segmentation result with good boundary preservation and pixel identity. Compared to other superpixel segmentation methods, this invention is mainly aimed at target detection tasks in SAR images under complex conditions. In particular, when the target sizes in the SAR image differ significantly, existing superpixel segmentation methods cannot well preserve the boundaries of small targets, resulting in the problem of small targets being incorporated into superpixels. This invention achieves ideal superpixel segmentation results by introducing edge information.

[0043] The following are specific embodiments of the present invention. It should be noted that the present invention is not limited to the following specific embodiments. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of the present invention.

[0044] Example:

[0045] This embodiment presents a fuzzy iterative SAR image superpixel segmentation method that fuses edge information, such as... Figure 1 As shown, the method includes the following steps:

[0046] Step 1: Use an edge detector based on a Gaussian gamma-shaped dual window to perform edge detection on the SAR image to obtain an edge enhancement map (ESM) and an edge map (EM).

[0047] Step 2: Based on the set superpixel size S, initialize the cluster centers using the edge enhancement map ESM and the edge map EM.

[0048] In step 2, the method for initializing cluster centers is as follows:

[0049] Step 201: Based on the superpixel size S, distribute the cluster centers evenly on a regular grid spaced S pixels apart within the SAR image.

[0050] Step 202: Locate the contour on the edge map EM to obtain approximate target contour information.

[0051] Step 203: Based on the target contour information, determine whether the target contour contains a cluster center each time. If it does, skip it. If it does not, take the pixel at the coordinate value of the average of the maximum and minimum values ​​of the target contour on the x-axis and y-axis of the two-dimensional plane as the new cluster center.

[0052] Step 204: To avoid cluster centers falling on edge or noisy pixels, the cluster centers are moved to the position with the minimum edge intensity in the 3×3 neighborhood based on the edge enhancement map (ESM) information.

[0053] Step 3: Calculate the metric distance D(i,j) between the cluster center and each pixel within the 2S×2S neighborhood of the cluster center.

[0054] In step 3, the method for calculating the metric distance D(i,j) between the cluster center and each pixel is as follows:

[0055] The formula for calculating the distance metric D(i,j) is as follows:

[0056] D(i,j)=(D intensity (i, j) + αD spatial (i,j))γ af

[0057] In the formula:

[0058] i is the cluster center;

[0059] j represents the pixel number;

[0060] D intensity (i, j) represents the intensity metric distance between the cluster center and the pixel;

[0061] D spatial (i, j) represents the spatial distance between the cluster center and the pixel;

[0062] α is a balance factor, and the balance intensity measures the weight of distance and spatial distance;

[0063] γ afThe edge information coefficient.

[0064] Furthermore, in step 3, the specific definitions of each term in the calculation formula for the distance metric D(i,j) are as follows:

[0065]

[0066] In the formula:

[0067] These represent the average intensity values ​​of cluster center i, pixel j, and their 8 neighbors, respectively.

[0068] M is the number of pixels for which the average intensity value is taken, which is 9 in this case;

[0069] Ln() performs the logarithmic operation.

[0070]

[0071] In the formula:

[0072] (x i y i () represents the coordinates of cluster center i;

[0073] (x j y j ) represents the coordinates of pixel j.

[0074] γ af = 1 + β·E(i, j)

[0075]

[0076] In the formula:

[0077] β is the amplification factor;

[0078] E(i,j) is the edge information factor;

[0079] The linear path from cluster center i to pixel j;

[0080] k represents the number of pixels on the linear path.

[0081] For existence;

[0082] It can be arbitrary;

[0083] EM(k) is the value of the edge graph EM at point k.

[0084] Step 4: Metric distance comparison update. Each pixel has a metric distance with its surrounding cluster centers. The smaller l metric distances between each pixel and its surrounding cluster centers are retained.

[0085] In step 4, the method for updating the metric distance comparison is as follows:

[0086] The search range for cluster centers is 2S×2S. Pixels will be within the search range of multiple cluster centers, so each pixel will have multiple metric distances. Only l metric distances of a pixel are retained. When the number of metric distances between a pixel and its surrounding cluster centers is greater than l, the new metric distance is compared with the largest metric distance that has been saved. If it is less than l, it is replaced.

[0087] Step 5: Calculate the membership degree U and the cluster center C.

[0088] Step 6: Repeat steps 3, 4 and 5 until the set number of iterations is reached.

[0089] Step 7: Select the cluster center category with the highest pixel membership degree as the superpixel category of the pixel to obtain the initial superpixel segmentation result.

[0090] Step 8: Use the edge information of the edge map EM for post-processing to obtain the final superpixel segmentation result.

[0091] In step 8, the post-processing steps are as follows:

[0092] The initial superpixel segmentation result contains many isolated pixels and some small pixel regions. During the post-processing step, the image is traversed, and isolated pixels and small pixel regions are merged into their adjacent superpixels. However, when merging small pixel regions, the edge map (EM) is used to check whether the small pixel region contains edge information. If it does, the small pixel region is considered a small target region and is treated as a separate superpixel, not merged into adjacent superpixels. If it does not contain edge information, it is merged into the adjacent superpixels.

[0093] Comparative Example 1:

[0094] This comparative example presents a superpixel segmentation method, namely the SLIC method (doi:10.1109 / TPAMI.2012.120). The general framework of the above embodiment is the same as that of this comparative example, but the specific steps are different. This method does not utilize edge information, so the specific steps involving edges are different from those in the embodiment. Furthermore, this method performs hard clustering when clustering local regions, and the superpixel to which a pixel belongs is determined during iteration.

[0095] Comparative Example 2:

[0096] This comparative example presents a superpixel segmentation algorithm for synthetic aperture radar (SAR) images that incorporates edge information (doi:10.3969 / j.issn.1671-1815.2020.24.038), denoted as SLIC-EDGE. This algorithm is based on SLIC, but its specific steps differ from the embodiments described above. The algorithm first preprocesses the SAR image with Gaussian smoothing, then constructs an improved similarity measurement parameter based on the ratio of exponential weighted average (ROEWA) operator, and finally performs local hard clustering using a hexagonal initialization cluster center and circular region search method to obtain the final superpixel segmentation result.

[0097] Comparative Example 3:

[0098] This comparative example presents a superpixel segmentation algorithm for SAR images. The algorithm in this comparative example uses the SLIC-GLR method (doi:10.1109 / LGRS.2018.2805714) to perform superpixel segmentation on SAR images. The SLIC-GLR method is based on SLIC and improves the metric method according to the characteristics of SAR images.

[0099] Comparative Example 4:

[0100] This comparative example presents a superpixel segmentation algorithm for SAR images. The algorithm in this comparative example uses the SREP method (doi:10.1049 / el.2016.0020) to perform superpixel segmentation on SAR images.

[0101] Simulation experimental conditions and data:

[0102] The specific simulation environment was as follows: Python 3.8, Intel(R) Core(TM) i5-6600 CPU@3.30GHz and 8-GB RAM, Windows 10 Professional Edition. The simulation data consisted of two SAR images: the first was a three-view simulated SAR image (Figure 2(a)) with a size of 400×300; the second was a SAR image selected from the publicly available HRSID dataset from the University of Electronic Science and Technology of China (Figure 3(a)). Specific information is shown in Table 1.

[0103] Table 1. Specific information of the selected SAR images

[0104] sensor TerraSAR-X model SM Angle of incidence (°) 20～45 Resolution (m) 3 polarization mode HH Image size 800x800

[0105] Comparison of superpixel segmentation results:

[0106] A comparative analysis was performed on the embodiments and comparative examples 1, 2, 3, and 4. In simulation 1 (… Figures 2(a) to 2(f) In simulation 2), when performing superpixel segmentation of a three-view simulated SAR image with a superpixel size of 40, only the fuzzy iterative SAR image superpixel segmentation method that fuses edge information in this embodiment did not integrate small targets into large superpixels, thus maintaining the independence of small targets. Therefore, it can be seen that when the target sizes differ significantly, this invention can effectively preserve the boundaries of small targets. Similarly, in simulation 2... Figures 3(a) to 3(f) In this embodiment, when performing superpixel segmentation with a superpixel size of 45 on the selected real SAR image, the fuzzy iterative SAR image superpixel segmentation method that fuses edge information yields superpixel segmentation results with good boundary preservation and pixel identity. Furthermore, through quantitative comparison (Table 2), it can be seen that the fuzzy iterative SAR image superpixel segmentation method that fuses edge information in this embodiment achieves higher BR (boundary recall, the larger the better) and lower UE (undersegmentation error, the smaller the better) than the other three superpixel methods on both three-view simulated SAR images and real SAR images.

[0107] Table 2 Evaluation of Superpixel Segmentation Results

[0108]

[0109] Note: "Invention" in Table 2 refers to the fuzzy iterative SAR image superpixel segmentation method that fuses edge information according to an embodiment of the present invention.

Claims

1. A method for superpixel segmentation of blurred iterated SAR image fusing edge information, characterized in that, The method comprises the following steps: Step 1, edge detection is performed on the SAR image by using a Gauss-Gamma-shaped double-window-based edge detector to obtain an edge-enhanced image ESM and an edge image EM; Step 2, according to the set superpixel size , using the edge enhancement map ESM and the edge map EM, initializing the cluster centers; Step 3, in the neighborhood of the cluster center, compute the metric distance between the cluster center and each pixel point ;​ Step 4, metric distance comparison update, each pixel point and the surrounding cluster center have metric distance, and the smaller one of each pixel point and the surrounding cluster center is reserved metric distance In step 4, the distance comparison updating method is as follows: The search range of the cluster center is The pixel point is in the search range of multiple cluster centers, so each pixel point has multiple metric distances, only the metric distance of the pixel point is reserved, when the number of metric distances between the pixel point and the surrounding cluster centers is greater than , the new metric distance is compared with the saved maximum metric distance, if it is smaller, it is replaced. Step 5, Compute membership and cluster centers ; Step 6, steps 3, 4 and 5 are repeated until a set iteration number is reached; Step 7, a cluster center class with the maximum pixel membership degree is selected as the superpixel class of the pixel to obtain an initial superpixel segmentation result; Step 8, an edge information-based post-processing step is performed on the edge image EM to obtain a final superpixel segmentation result.

2. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 1, wherein, In step 2, the cluster center initialization method is as follows: Step 201, according to the superpixel size Within the SAR image, the cluster centers are evenly distributed on a regular grid of pixels; Step 202, contour information is obtained by searching for contours on the edge image EM; Step 203, according to the target contour information, in turn, determine whether each target contour contains clustering center, if contains, skip, if not, the target contour in two-dimensional plane axis and the pixel point at the average coordinate value of the maximum and minimum value on the axis as the new clustering center; Step 204, in order to avoid the cluster center from falling on an edge or a noise pixel, the cluster center is moved to a position with minimum edge intensity in a 3*3 neighborhood according to the edge-enhanced image ESM information.

3. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 1, characterized in that, In step 3, the metric distance between the cluster center and each pixel point is calculated The method is as follows: Metric distance The formula for the calculation of the metric distance is as follows: In the formula: is the cluster center; pixel point; measure the distance of the cluster center from the intensity of the pixel point; is the spatial distance between the cluster center and the pixel point; A balance factor balances the weight of the strength metric distance and the spatial distance. is the edge information coefficient.

4. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 3, characterized in that, In step 3: In the formula: , respectively denote the average intensity value of the cluster center , pixel point and its 8-neighborhood. The number of pixels for taking the average intensity value is 9 here. is a logarithm operation.

5. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 3, characterized in that, In step 3: In the formula: coordinate values of the cluster center coordinate values of the cluster center coordinate value of a pixel point coordinate value of a pixel point 6. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 3, wherein, In step 3: In the formula: A is an amplification factor; is the edge information factor; to cluster center to pixel linear path; is a pixel point on the linear path; To be; arbitrary; For the edge map EM in the value of the point.

7. The fuzzy iterative SAR image superpixel segmentation method of fusing edge information according to claim 1, wherein, In step 8, the post-processing step is as follows: Many isolated pixels and small pixel regions are contained in the initial superpixel segmentation result, and in the post-processing step, the isolated pixels and small pixel regions are merged into adjacent superpixels, but in the merging of small pixel regions, the edge image EM is used to check whether the small pixel region contains edge information, if yes, the small pixel region is considered as a small target region and is taken as a superpixel alone without being merged into adjacent superpixels, and if no, the small pixel region is merged into adjacent superpixels.

Images