Radar target scattering center region segmentation method based on local density clustering

Abstract

The radar target scattering center region segmentation method based on local density clustering first carries out Frost filtering to the original radar image to realize global noise suppression of the image; secondly, the image segmentation is carried out to the Frost filtered image based on the level set method, and the suspected target region of interest is preliminarily obtained; then, the area filtering is adopted to extract the maximum connected domain, and the target region of interest is obtained; finally, the scattering center is detected and the region segmentation is carried out by using the local density clustering algorithm. The radar target scattering center region segmentation method based on local density clustering has high accuracy, and can effectively improve the technical problem that the SAR image is difficult to have high enough resolution in the actual situation, so that the traditional image segmentation algorithm is difficult to segment the region containing only one scattering center.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target scattering center detection, and particularly relates to a radar target scattering center region segmentation method based on local density clustering. BACKGROUND

[0002] The backscattering echo of a radar target working in high frequency region can be considered as the result of coherent combination of a limited number of strong scattering sources, which are usually equivalent to the scattering centers of the target. As an effective feature to describe the high frequency scattering mechanism of the target, the scattering center parameter extraction has wide research and application value in the military fields such as radar target recognition, RCS extrapolation fitting, and target three-dimensional reconstruction.

[0003] At present, people mainly use the Approximate Maximum Likelihood (AML) method and the like to estimate the scattering center parameters. These parameter estimation methods have high precision and the idea is intuitive and easy to understand, but they will face the mixed parameter optimization problem of high dimension and nonlinearity, and the calculation complexity is extremely large. On the high-resolution radar image, the scattering centers of the target correspond to block-shaped regions with concentrated energy and mutual separation, which makes the extraction of the scattering center information of the target into the processing of various block-shaped regions. First, a number of Region Of Interest (ROI) regions with high peak values are obtained by using an image segmentation algorithm, and then the scattering center parameter estimation is performed on each ROI region. The idea of "region decoupling" can greatly improve the speed of the scattering center parameter estimation. However, in actual situations, the SAR image is difficult to have high enough resolution, and therefore the traditional image segmentation algorithm is difficult to segment the region containing only one scattering center, and the accuracy of the radar target scattering center region segmentation needs to be improved. SUMMARY

[0004] In view of the technical problem that in actual situations, the SAR image is difficult to have high enough resolution, and the traditional image segmentation algorithm is difficult to segment the region containing only one scattering center, the present application provides a radar target scattering center region segmentation method based on local density clustering, which can effectively realize the radar target scattering center region segmentation, has high automation degree and high speed, and has the advantages of high accuracy.

[0005] The technical scheme adopted by the present application is as follows:

[0006] The radar target scattering center region segmentation method based on local density clustering comprises the following steps:

[0007] Step 1: Frost filtering is performed on the original SAR radar image to realize global noise suppression of the image;

[0008] Step 2: image segmentation based on LSM, preliminary acquisition of suspected target ROI region;

[0009] Step 3: area filtering is adopted to retain the maximum connected domain to remove interference targets and obtain the target ROI region;

[0010] Step 4: the LDC algorithm is used to detect the scattering center and perform region segmentation.

[0011] The step 1 is specifically as follows:

[0012] It is assumed that the SAR radar image is a stationary process and the speckle noise is multiplicative noise, and the mathematical model is as follows:

[0013] S(i,j)=S0(i,j)·N0(i,j);

[0014] In the formula, S(i,j) is the observed image; S0(i,j) is the image without noise; N0(i,j) represents the speckle noise.

[0015] When the size of the filtering window S kl is (2N+1)×(2N+1), the Frost filtering result is as follows:

[0016]

[0017] D kl is the Euclidean distance between the pixels (k,l) and (i,j) in the window:

[0018] U i,j represents the mean value of the gray scale of all pixels in the filtering window S kl

[0019] V i,j represents the variance of the gray scale of all pixels in the filtering window S kl

[0020] Wherein, (2N+1) is the length or width of the filtering window, and p is the fine tuning factor; k, l, i, j represent the indexes of the image pixels respectively.

[0021] In the step 2, the LSM is applied to image segmentation, which regards the curve function of a two-dimensional image as a zero-equal surface of a higher one-dimensional surface, and tracks the evolution of the curve by using the evolution of the implicit equation. The LSM is specifically as follows:

[0022] ​​The image is divided into two types of target region and background region, which are smooth and homogeneous regions. Let F(x, y) ∈ Ω, the closed contour curve C in the processing region Ω divides Ω into target region Ω in (C's interior) and background region Ω out (C's exterior), the average gray values of target region Ω in and background region Ω out are μ in and μ out respectively, then the energy function is:

[0023]

[0024] Where x, y are the indexes of image pixels; α, β > 0, γ ≥ 0 are the weighted coefficients; L C is the length of the closed contour curve C; the first two terms of the energy function are fidelity terms, which can get the minimum value when the closed contour curve C converges to the target boundary and realizes the separation of the target and the background; the third term is a regularization constraint term, which ensures the smoothness of the evolving curve.

[0025] The ROI region is the region of interest (Region Of Interest). In the field of image processing, the region of interest is selected from the image as the key region for image analysis. The region is circled for further image processing, which can reduce the processing time and increase the accuracy.

[0026] In step 3, the area filtering is the connected component extraction, which is to find the connected components in the binary image and assign a unique identifier to each connected component, then retain the largest connected component and set the rest to zero.

[0027] The gray-scale threshold L is used to binarize the gray-scale image T1 to obtain the binary image T2, and the calculation formula is as follows:

[0028]

[0029] Where T1(i, j) represents the pixel value of image T1; T2(i, j) represents the pixel value of image T2; i, j represent the indexes of image pixels; max(T1(i, j)) represents the maximum value of T1(i, j), and min(T1(i, j)) represents the maximum value of T1(i, j).

[0030] The connected region is a pixel set composed of adjacent pixels with the same pixel value, and the connected component extraction is to find the connected regions in the binary image and assign a unique identifier to each connected component, then retain the required connected components according to the requirements and set the rest to zero.

[0031] In step 4, the local density clustering is as follows:

[0032] Step 4.1: First, calculate three parameters for each point: local density ρ, distance σ, and gradient.

[0033] For the i-th data point p i Local density ρ i Defined as:

[0034] For the i-th data point p i distance σ i Defined as:

[0035] For the i-th data point p i gradient Defined as:

[0036] Where, d c d represents the cutoff distance. ij Represents the i-th data point p i and the j-th data point p j Euclidean distance between them, I j Represents data point p j The strength at that location.

[0037] Step 4.2: As Figures 1(a) to 1(c) As shown, after calculating the three parameters, the relationship between distance σ and density ρ is plotted as a decision graph. Figure 1(a) shows simulation data with 10 scattering centers, which are generated by the Omega-K radar imaging algorithm with Gaussian noise as the background. Figure 1(b) shows the detected scattering centers, with each scattering center and its corresponding region marked by red dots and blue ellipses, respectively. Figure 1(c) is the decision graph, with the detected scattering centers marked by circles. σ0 and ρ0 are hyperparameters of the LDC algorithm, where σ0 represents the minimum distance between two scattering centers, and ρ0 represents the minimum peak intensity value of the candidate scattering center region.

[0038] The LDC algorithm posits that cluster centers should possess two characteristics: first, they should be surrounded by data points with lower local density; second, they should be significantly distanced from other data points with higher local density. Therefore, only data points with high density ρ and large distance σ can be considered cluster centers. The decision graph visually reflects the relationship between distance σ and density ρ, thus allowing the determination of the location and number of cluster centers.

[0039] Step 4.3: Then, based on the density ρ and gradient... It is possible to assign non-cluster centers to the nearest cluster centers, thereby determining a single scattering center C. k The area.

[0040]

[0041] in: Represents the scattering center region C k The j-th point in ρ; j p represents the j-th data point j The density; It is the scattering center region C k The average density; p represents the j-th data point j The gradient of n; k It is the scattering center region C k The total number of data points in the middle; These are the hyperparameters of the LDC algorithm.

[0042] This invention provides a radar target scattering center region segmentation method based on local density clustering, which has the following beneficial effects:

[0043] 1) The method of the present invention performs local density clustering on radar targets and uses processing methods such as Frost filtering, LSM image segmentation, and area filtering to effectively segment the radar target scattering center region.

[0044] 2) The method of the present invention can achieve batch segmentation of the radar target scattering center region, with a high degree of automation and high speed.

[0045] 3) The method of the present invention has a high accuracy in segmenting the radar target scattering center region, and both the false detection rate and the false detection rate are relatively low. Attached Figure Description

[0046] Figure 1(a) shows the simulation data, which is a SAR image generated by the Omega-K radar imaging algorithm. The background is Gaussian noise with 10 scattering centers.

[0047] Figure 1(b) shows the detected scattering centers, with each scattering center and corresponding region marked by a red dot and a blue ellipse, respectively.

[0048] Figure 1(c) is a decision map, with the detected scattering centers marked by circles.

[0049] Figure 2 This is a flowchart of the radar target scattering center region segmentation algorithm proposed in this invention.

[0050] Figure 3(a) shows the original radar image W1.

[0051] Figure 3(b) shows the result image W2 after W1 is filtered by Frost.

[0052] Figure 3(c) shows the result image W3 after LSM image segmentation of W2.

[0053] Fig. 3(d) is the result image W4 of W3 after area filtering.

[0054] Fig. 3(e) is the result image W5 of W4 after local density clustering, and the scattering center information is marked on the image W4 with red dots.

[0055] Fig. 3(f) is the result image W6 of W4 after local density clustering, and the scattering center and region segmentation information are marked on the image W1 with red dots and blue ellipses respectively.

[0056] Fig. 4(a) is simulation data, which is a SAR image simulated by Omega-K radar imaging algorithm.

[0057] Fig. 4(b) is a noise background extracted from a real database (MSTAR data set).

[0058] Fig. 4(c) is synthetic data, which is a real SAR image with noise simulated by adding a noise background to the simulation data.

[0059] Fig. 5(a) is a correct case of the algorithm in real data detection.

[0060] Fig. 5(b) is a case of obvious missed detection of the algorithm in real data.

[0061] Fig. 5(c) is a case of obvious false detection of the algorithm in real data. DETAILED DESCRIPTION

[0062] The present application proposes a radar target scattering center region segmentation method combining machine vision and digital image processing technology, which adopts a local density clustering (LDC) algorithm to detect scattering centers and perform region segmentation. First, the original radar image is subjected to Frost filtering to realize global noise suppression of the image; second, the image segmentation is performed on the Frost filtered image based on a level set method (LSM) to preliminarily obtain a suspected target region of interest (ROI); then, the area filtering is adopted to extract the maximum connected domain to obtain the target ROI region; finally, the LDC algorithm is used to detect scattering centers and perform region segmentation. The specific implementation is as follows:

[0063] On high-resolution SAR images, the scattering centers of targets correspond to blocky regions with high energy and separated from each other, which makes the extraction of target scattering center information transform into the processing of various blocky regions. First, the ROI region with high energy of single scattering center response of a target can be obtained by using an image segmentation algorithm, and then, the scattering center parameter estimation is performed on each ROI region respectively. The radar imaging system is based on the coherent principle, and therefore, in the radar echo signal, the gray value of adjacent pixels will produce some random changes around a certain mean value due to the coherence, which makes the noise in the SAR image inevitable. Therefore, before the scattering center region segmentation, a series of image preprocessing is required on the original image, and the whole algorithm flow chart is shown in Figure 2 The radar target scattering center region segmentation method based on local density clustering is provided, and the specific steps are as follows:

[0064] Step 1: Frost filtering

[0065] The Frost filtering is performed on the original observation image W1 (MSTAR official data set), global noise suppression of the image is realized, and the image W2 is obtained. The original observation image W1 is shown in Fig. 3(a), and the result image W2 of the Frost filtering is shown in Fig. 3(b). As shown in Fig. 3(a) and Fig. 3(b), it can be seen that the global noise of the image after the Frost filtering is obviously suppressed.

[0066] Step 2: LSM image segmentation

[0067] The LSM image segmentation is performed on the image W2 based on the level set method LSM, and the suspected target ROI region is preliminarily obtained. The result image W3 of the LSM image segmentation is shown in Fig. 3(c). As shown in Fig. 3(c), it can be seen that the LSM image segmentation can preliminarily obtain the suspected target ROI region, but there are still some interference regions, which are caused by local noise extreme points.

[0068] Step 3: Area filtering

[0069] The area filtering is performed on the image W3, the maximum connected domain is retained to remove the interference target, and the target ROI region is obtained. The result image W4 of the area filtering is shown in Fig. 3(d). As shown in Fig. 3(d), it can be seen that after the area filtering, the interference region is filtered out, and the remaining region is the target ROI region.

[0070] Step 4: Local density clustering

[0071] Finally, the scattering center in the image W4 is detected by using the LDC algorithm and region segmentation is performed. The scattering center information is marked on the image W4 to obtain the image W5 as shown in Fig. 3(e), and the scattering center and region segmentation information are marked on the image W1 to obtain the image W6 as shown in Fig. 3(f). As can be seen from Fig. 3(e) and Fig. 3(f), the LDC algorithm can accurately detect the position of the scattering center point and the surrounding influence area in the target ROI region obtained after area filtering, and realize the region segmentation of the scattering center.

[0072] Step 5: Algorithm verification

[0073] In order to prove the effectiveness of the method for radar target scattering center region segmentation, 200 SAR images containing label information are simulated in batches by using the Omega-K radar imaging algorithm as simulation data samples. Then, the noise background is extracted from the real database (MSTAR dataset) and added to the simulation data to obtain the synthetic data. A typical case is selected as a demonstration: the simulation data is shown in Fig. 4(a), the noise background is shown in Fig. 4(b), and the synthetic data is shown in Fig. 4(c). As can be seen from Fig. 4(a), each scattering center point corresponds to a block-shaped region with concentrated energy and is separated from each other in the figure, which is consistent with the real situation. Fig. 4(b) is the noise background extracted from the real database (MSTAR dataset), and Fig. 4(c) is the noise background added to the simulation data to simulate a real SAR image with noise.

[0074] Batch testing is performed on the synthetic data, and the test results are compared with the label information for verification. The accuracy and recall rate of each image are calculated, and the average value is obtained to obtain the algorithm evaluation index. The results of the LDC algorithm are compared with the results of the watershed algorithm (traditional image segmentation algorithm), as shown in Table 1. As can be seen from Table 1, compared with the watershed algorithm, the method of the present application has higher accuracy and recall rate for synthetic data, and can more effectively detect the scattering center of the radar target and perform region segmentation.

[0075] Table 1: Comparison of algorithm performance of synthetic data

[0076]

[0077] In order to fully test the algorithm performance, the T-72 A04 variant real test data of the MSTAR official dataset is taken as an example, and batch tests are performed on the LDC algorithm and the watershed algorithm.

[0078] Since the real data does not have a standard test result for verification, only subjective judgment by the human eye can be made, and the test results are divided into three categories: correct detection, obvious missed detection, and obvious false detection. The results are shown in Table 2.

[0079] Table 2: Comparison of algorithm performance of real data

[0080]

[0081] From Table 2, it can be seen that the correct detection rate of LDC algorithm on real data reaches 90.97%; the correct detection rate of watershed algorithm is only 81.94%. It can be seen that compared with the watershed algorithm, the detection ability of LDC algorithm in real data also has higher superiority.

[0082] From the above, the typical cases in the detection results of LDC algorithm are selected as display. The cases of correct detection are shown in Fig. 5(a), from which it can be seen that the marking results are basically in the ROI region. The cases of obvious missed detection are shown in Fig. 5(b), from which it can be seen that there is a part of ROI region that is not detected, which causes this phenomenon because in the image preprocessing, the LSM image segmentation algorithm divides the ROI region into two parts, and then the area filtering removes the part with smaller area, finally leading to the obvious missed detection of LDC detection. The cases of obvious false detection are shown in Fig. 5(c), from which it can be seen that part of the noise background is detected as the target, which causes this phenomenon because the sample signal-to-noise ratio is too low, the filtering is not complete, so that the subsequent LSM image segmentation judges part of the noise background as the ROI region, finally leading to the obvious false detection of LDC detection.

Claims

1. A radar target scattering center region segmentation method based on local density clustering, characterized in that The method comprises the following steps: Step 1: Frost filtering is performed on a SAR radar image to realize global noise suppression of the image; Step 2: image segmentation is performed based on a level set method (LSM) to preliminarily obtain a suspected target ROI region; Step 3: area filtering is adopted to retain a maximum connected domain to remove interference targets and obtain a target ROI region; Step 4: a local density clustering (LDC) algorithm is used to detect scattering centers and perform region segmentation; The step 1 is specifically as follows: Suppose that the SAR radar image is a stationary process and the speckle noise is multiplicative noise, and the mathematical model is as follows: S(i,j) = S0(i,j) * N0(i,j); In the formula, S(i,j) is an observed image; S0(i,j) is an image without noise; N0(i,j) represents speckle noise; At the filter window S kl When the size is (2N+1) x (2N+1), the Frost filter result is: D kl is the Euclidean distance between the pixels (k, l) and (i, j) within the window: U i,j representative filter window S kl the mean value of the gray scale values of all pixels within: V i,j representative filter window S kl variance of all pixel gray values within: Wherein, (2N+1) is the length or width of the filtering window, and p is a fine tuning factor; k, l, i, and j represent the indexes of image pixels; In the step 2, the level set method (LSM) is specifically as follows: The image is divided into two types of target region and background region, which are smooth and homogeneous regions. Let F(x, y) ∈ Ω, the closed contour curve C in the processing region Ω divides Ω into target region Ω in and background region Ω out , and the average gray values of target region Ω in and background region Ω out are μ in and μ out respectively. The energy function is as follows: where x, y are the indices of image pixels; a, b > 0, g > 0 are the weighting coefficients; L C is the length of the closed contour curve C; the first two terms of the energy function are fidelity terms, which can obtain the minimum value when the closed contour curve C converges to the target boundary and achieves the separation of the target and the background; and the third term is a regularization constraint term, which ensures the smoothness of the evolving curve.

2. The method of claim 1, wherein the method further comprises: In the step 3, a gray threshold L is used to binarize the gray image T1 to obtain a binary image T2, and the calculation formula is as follows: Wherein, T1(i,j) represents the pixel value of the image T1; T2(i,j) represents the pixel value of the image T2; i and j represent the indexes of image pixels; max(T1(i,j)) represents the maximum value of T1(i,j), and min(T1(i,j)) represents the maximum value of T1(i,j); A connected region is a pixel set composed of adjacent pixels with the same pixel value, and connected domain extraction is to find a connected region in a binary image, and to assign a unique identifier to each connected domain, and then to retain the required connected domain according to the requirement, and to set the remaining part to zero.

3. The method of claim 1, wherein the method further comprises: The step 4 comprises the following steps: Step 4.1: First, compute three parameters for each point: local density p, distance s, and gradient For the ith data point p i The local density p i is defined as: The distance σ i for the i-th data point p i is defined as: For the ith data point p i the gradient is defined as: Where, d c d represents the cutoff distance. ij Represents the i-th data point p i and the j-th data point p j Euclidean distance between them, I j Represents data point p j Strength at the location; Step 4.2: After the three parameters are calculated, the relationship between the distance sigma and the density p is plotted into a graph, which is called a decision graph, the decision graph can intuitively reflect the relationship between the distance sigma and the density p, the data points with high density p and large distance sigma are clustering centers, therefore, the position and number of the clustering centers can be determined according to the decision graph; Step 4.3: Then, according to the density p and the gradient The non-cluster center points can be divided into the nearest cluster center to the point, so as to determine a single scattering center C k region; where: represents the jth point in the scatter center region C k ; p j represents the density of the jth data point p j ; is the average density of the scatter center region C k ; represents the gradient of the jth data point p j ; n k is the total number of data points in the scatter center region C k ; is a hyperparameter of the LDC algorithm.

Images