Real-aperture scanning radar content adaptive super-resolution imaging method based on k-means image segmentation

Abstract

The application discloses a kind of real aperture scanning radar content self-adaptive super-resolution imaging method based on K-means image segmentation, for the problem that traditional regularization super-resolution method is difficult to consider artifact suppression and detail protection, the method introduces a kind of image content self-adaptive regularization model, significantly improves the super-resolution imaging effect of super-resolution method to complex scene.The method is first segmented by K-means algorithm, according to intensity, the image is divided into several regions with different scattering coefficients, then the constraint strength of different regions is adjusted according to the prior knowledge of local image structure, and finally the objective function is solved by using gradient descent method to obtain the super-resolution result.The method of the application completes noise filtering, region segmentation based on image intensity and constructs parameter transition region between each region, and then realizes content self-adaptive super-resolution imaging.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to an adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation. Background Technology

[0002] Forward-looking radar imaging has a wide range of applications, such as aircraft landing, terrain avoidance, and airport surveillance. However, in the forward-looking direction, the Doppler bandwidth of the radar echo drops sharply, and there is also severe left-right Doppler blurring. This characteristic makes it difficult to directly apply traditional synthetic aperture radar imaging techniques and methods such as Doppler beam sharpening to forward-looking imaging.

[0003] Convolutional inversion technology performs super-resolution processing on the azimuth echo of real-aperture scanning radar from a signal processing perspective. This overcomes the limitations of radar's real-aperture azimuth resolution, improving overall imaging performance and practical application capabilities. Its advantages include simple imaging methods, compatibility with existing radar systems, and the ability to distinguish multiple targets within the same beam. It has gradually become a major research hotspot in radar forward-looking imaging.

[0004] However, due to the ill-conditioned nature of deconvolution, many deconvolution methods are sensitive to noise. Regularization methods can effectively alleviate the ill-conditioned nature of convolution inversion. They construct different regularization penalty terms based on known prior information to constrain the target solution, thus addressing the problem that convolution inversion solutions are sensitive to noise and discontinuously dependent on observation data. For example, the paper "A hybrid norm regularization approach for radar forward-looking angle super-resolution imaging" (Tuo X, Zhang Y, Huang Y et al.. 2021 IEEE RadarConference, 2021, pp. 1-5.) uses a combined L1+L2 regularization algorithm for super-resolution imaging of the target. The L1 regularization term can impose sparse constraints on the solution, thereby improving angular resolution, but it can damage background contour information when the number of iterations is too high. The L2 regularization term can ensure the smoothness of the solution, thereby reducing noise, but its super-resolution performance is limited due to excessive smoothing. The paper "A super-resolution imaging method for forward-looking scanning radar based on improved total variation" (Shen J, Mao D, Zhang Y, et al.. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, 2024, pp. 10471-1047) utilizes an improved total variation regularized super-resolution algorithm to reduce artifacts while preserving scene contours relatively well. However, due to the limitations of globally uniform parameters, the above methods all assume that the entire image shares the same prior information. In imaging tasks, each region has different structural features and energy distributions, and scenes are usually complex. Therefore, adaptive super-resolution methods still need to be explored and utilized. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation. This method overcomes the problem that traditional regularized super-resolution methods, which use globally uniform parameters, are unable to simultaneously address artifact suppression and detail preservation, thereby achieving content-adaptive super-resolution imaging.

[0006] The technical solution to achieve the purpose of this invention is: a content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation, comprising:

[0007] The image is segmented using the K-means algorithm, dividing it into several regions with different scattering coefficients based on intensity.

[0008] The constraint strength of different regions is adjusted based on prior knowledge of the local image structure, and parameter transition areas are constructed between regions to smooth the region boundaries;

[0009] The objective function is solved using the gradient descent method to obtain the super-resolution results.

[0010] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0011] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.

[0012] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0013] Compared with existing technologies, the advantages of this invention are as follows: The innovation of this invention lies in segmenting the image using the K-means clustering algorithm, dividing the image into several regions with different scattering coefficients based on intensity. By combining prior information about the image and the desired effect, different regularization parameters are set for different regions, ultimately obtaining super-resolution imaging results, thus solving the problem of traditional super-resolution methods assuming all regions share the same prior information. The advantage of this invention is that by adaptively adjusting the constraint intensity of different regions of the image through image segmentation, it fully utilizes the prior information of the image, thereby enabling better recovery of details in complex scenes. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method provided by the present invention.

[0015] Figure 2 This is a spatial relationship diagram of motion scanning of a real aperture scanning radar used in a specific embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of the simulated imaging scene used in a specific embodiment of the present invention.

[0017] Figure 4 The signal is an echo signal superimposed with Gaussian white noise in a specific embodiment of the present invention, with a signal-to-noise ratio of 15dB.

[0018] Figure 5 This is the preliminary region segmentation result after segmenting the image using the K-means algorithm (K=4).

[0019] Figure 6 It is the final region division result constructed after region merging and parameter transition region construction.

[0020] Figure 7 It is the content-adaptive super-resolution imaging result after processing using the method of this invention. Detailed Implementation

[0021] This invention proposes a content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation, introducing a content-adaptive regularization model. The method first segments the image using the K-means algorithm, dividing it into regions with different scattering coefficients based on intensity. Then, it adjusts the constraint intensity of different regions based on prior knowledge of the local image structure and constructs parameter transition zones between regions to smooth the region boundaries. Finally, it uses gradient descent to solve the objective function, ultimately achieving content-adaptive super-resolution imaging.

[0022] To facilitate the description of the present invention, the following terms will first be explained:

[0023] Term 1: Real Aperture Scanning Radar (RASR)

[0024] Real Aperture Scanning Radar (RASR) is a radar system relative to Synthetic Aperture Radar (SAR). Unlike SAR, which uses synthetic aperture to improve resolution in the azimuth direction, RASR does not perform any signal processing in the azimuth direction and only uses real beams to distinguish targets.

[0025] Term 2: K-means clustering

[0026] K-means is an unsupervised clustering algorithm that aims to divide a dataset into K disjoint clusters, such that data points within the same cluster are as similar as possible, and data points between different clusters are as different as possible.

[0027] Combination Figure 1 The present invention proposes a content-adaptive super-resolution imaging method for real aperture scanning radar, comprising the following steps:

[0028] Step 1: Constructing the azimuth convolution signal model

[0029] By using an airborne real-aperture scanning radar to image the target area and obtain a real-beam image, and focusing only on the amplitude of the azimuth signal, the azimuth signal can be written as...

[0030]

[0031] in, Here are the azimuth coordinates of the target. The scattering intensity of a point target. For convolution operations, This is the antenna radiation pattern. Noise is taken into consideration. The influence of the azimuth signal is expressed as follows:

[0032]

[0033] Based on the above model, construct an azimuth convolution signal vector model:

[0034]

[0035] in, and These represent the echo data and target scattering coefficients within a range gate, respectively. For noise vectors, The number of elements in the vector. Antenna radiation pattern The constructed convolution matrix.

[0036] Step 2: Construct the imaging objective function

[0037] To address the noise sensitivity issue in deconvolution operations, regularization is typically employed. The standard regularization equation for solving this problem is:

[0038]

[0039] in, For data fidelity items, For regularization terms, The constraints used to balance data fidelity and regularization terms can be determined by the L-curve method.

[0040] Choose a weighted combination of the L1 norm, L2 norm, and total variation regularization term as the joint regularization term:

[0041]

[0042] in, , , These are the coefficients of the L1 regularization term, the L2 regularization term, and the total variation regularization term, respectively. For L1 regularization terms, For L2 regularization terms, For the total variation regularization term, its formula is:

[0043]

[0044] Absolute value penalty promotes sparsity of solutions, penalizing many subtle gradients to zero. Squared penalty uniformly and smoothly suppresses the overall gradient of the image, tending to produce smoother solutions. Total variation regularization penalizes the sum of the image gradient magnitudes, minimizing the overall variation of the image while allowing for discontinuities (such as edges).

[0045] Substituting equations (5) and (6) into equation (4), the objective function for super-resolution imaging of RASR can be written as:

[0046]

[0047] Step 3: Calculate the gradient of the objective function

[0048] Due to the nondifferentiability of the L1 norm, a minimal constant is introduced. Formula (13) can be written as:

[0049]

[0050] in, Because the image has high dimensionality, directly solving formula (8) is very difficult. Therefore, the gradient descent method is used to solve for the minimum value. The gradient of the objective function is obtained as follows:

[0051]

[0052] in, It is a diagonal matrix. .

[0053] Step 4: Adaptive Image Segmentation

[0054] Next, image segmentation is performed based on the image content. To avoid the influence of noise on image segmentation and to smooth region edges, a Gaussian filter is first applied to the image, defined as having a size of [value missing]. Gaussian kernel:

[0055]

[0056] in, This indicates the offset of the current pixel relative to the center of the filter. It is the standard deviation of the Gaussian distribution, controlling the degree of smoothness. For the input image Output image The value of each pixel is a weighted average of the values ​​of its neighboring pixels:

[0057]

[0058] Next, the K-means algorithm is used to segment the image, with the goal of minimizing the sum of squared errors within clusters:

[0059]

[0060] in, It is the grayscale value of each pixel in the image. It is the center point of each cluster. It is the division of each cluster. yes To its cluster center The Euclidean distance. Sort all pixels in the image according to their grayscale values ​​from low to high, and initialize K cluster centers, with the value of each cluster center being:

[0061]

[0062] in, Given the total number of pixels in the image. Calculate the distance from the grayscale value of each pixel to all cluster centers, and assign the pixel to the nearest cluster:

[0063]

[0064] This refers to the partitioning of each cluster in the t-th iteration (loop); In the t-th iteration, The center point of the cluster to which it belongs; In the t-th iteration, except The center point of any cluster other than the one it belongs to.

[0065] After classification is completed, the average gray value of all pixels in each class is calculated, and the result is updated as the new cluster center for that class.

[0066]

[0067] The K-means algorithm clustering steps are iterated continuously until the cluster centers no longer change, at which point the loop terminates. Based on prior knowledge of the image, the segmented regions are merged to obtain the final region segmentation result.

[0068] Step 5: Construct the parameter transition region

[0069] Parameter variations between different regions can cause artifacts, therefore, it is necessary to construct parameter transition regions between regions with drastic parameter changes. The regularization parameter for the transition region between region A and region B is given by the following formula:

[0070]

[0071] in, Point Distance to region A, This represents the maximum distance from all points within the transition area to region A. and These represent the regularization parameters for regions A and B, respectively. The closer a point in the transition region is to a certain region, the closer its regularization parameter is to that region.

[0072] Step 6: Construct the full graph regularization coefficient matrix

[0073] Based on prior knowledge of the local image, different regularization parameters are set for each region divided in step four, and a full-image regularization coefficient matrix is ​​constructed.

[0074] Step 7: Solve for the objective function

[0075] Substitute the regularization coefficient matrix set in step six into the gradient of the objective function in step three, and use gradient descent to solve the objective function until convergence or the maximum number of iterations is reached. Randomly select a starting point. Typically, the input image is taken, and the gradient at the current point is calculated. And adjust the parameters along the negative gradient direction:

[0076]

[0077] in, It is the pixel value updated at step k+1. The step size controls the magnitude of each update.

[0078] After each iteration, the objective function value decreases. Through gradual convergence, the objective function value eventually stabilizes near a local minimum, resulting in the reconstructed image.

[0079] The following figures and embodiments illustrate the invention in detail.

[0080] Example

[0081] This invention is primarily verified using simulation experiments, and all steps and conclusions have been verified correctly using Matlab 2024a. The following provides a further detailed description of the specific implementation methods.

[0082] Step 1: The airborne forward-looking radar scans and transmits pulse signals simultaneously to image the target area. A schematic diagram of the scanning radar imaging in this embodiment is shown below. Figure 2 As shown in Table 1, the parameters required for the simulation are as follows.

[0083] Table 1

[0084]

[0085] The antenna pattern used in this simulation is Gaussian, and its expression is:

[0086]

[0087] in, It's the azimuth. It is a 3dB beamwidth, when When the gain drops to its maximum value, Target scenario such as Figure 3 As shown, the azimuth range of the scene is -10° to 10°, and the range range is 3000m to 3500m. Linear convolution is performed row-by-row between the target scene and the antenna pattern. The convolution result is then centered and cropped to match the length of the target scene. Finally, Gaussian noise with a signal-to-noise ratio of 15dB is added to obtain a real beam echo with azimuth ambiguity, as shown. Figure 4 As shown.

[0088] Step 2: Construct the convolution matrix. Based on the antenna scanning speed... and pulse repetition frequency Calculate the scan angle for each pulse Then, the antenna scanning range Calculate the number of sampling points This yields the azimuth sampling number. From this, a convolution kernel is constructed to analyze the antenna pattern. Discretize the data by uniformly sampling 572 points within the range of -10° to 10° with a stride of 0.035° and normalize the results to obtain the convolution kernel. Convolution matrix By convolution kernel The arrangement is repeated along the distance direction, and its number of rows is the same as the number of rows in the target scene, both being 300. .

[0089] Step 3: Construct the objective function using regularized super-resolution methods:

[0090]

[0091] in, It is a convolution matrix. It is the echo matrix. The goal is to reconstruct the model. Since the L1 norm is not differentiable at zero, a minimal constant is introduced to facilitate differentiation. The objective function is rewritten as:

[0092]

[0093] Differentiating equation (20), we obtain the gradient of the objective function:

[0094]

[0095] Step 4: Image Segmentation. First, apply a 3×3 Gaussian kernel to the echo matrix using Gaussian filtering to reduce the impact of noise on image segmentation. Then, use the K-means algorithm to segment the filtered matrix, setting K=4, to obtain preliminary region division results, such as... Figure 5 As shown. Based on the expected classification results, the larger yellow area on the left (playground area) is retained, and the remaining yellow areas are merged into the green area (land area). A transition area of ​​approximately 25 pixels wide is set between the land area and the playground area to smooth the boundary, resulting in the final region division result, as shown. Figure 6 As shown.

[0096] Step 5: Apply different regularization parameters to each region based on prior knowledge of the local image. The regularization parameter table is shown in Table 2.

[0097] Table 2

[0098]

[0099] The parameters of the transition region are given by the following formula:

[0100]

[0101] in, Points in the transition region Distance to the playground area This represents the maximum distance from all points within the transition area to the playground area. and These represent the regularization parameters for the playground area and the land area, respectively.

[0102] Step Six: Based on the regularization parameters set in Step Five, solve the objective function from Step Three using gradient descent until it converges or reaches the maximum number of iterations. The maximum number of iterations is 100, and the convergence condition is that the mean square error between the results of each iteration does not exceed [a certain threshold]. The mean square error is defined as follows:

[0103]

[0104] in Indicates the first grayscale of a pixel, It is the total number of pixels. It represents the number of iterations.

[0105] Content-adaptive super-resolution imaging of a real aperture scanning radar was completed. The imaging results are as follows: Figure 7 As shown.

[0106] As can be seen from the specific embodiments of the present invention, the present invention can realize content-adaptive super-resolution imaging of real aperture scanning radar.

[0107] The specific embodiments of the present invention have been described in detail above. It should be noted that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation, characterized in that, include: The image is segmented using the K-means algorithm, dividing it into several regions with different scattering coefficients based on intensity. The constraint strength of different regions is adjusted based on prior knowledge of the local image structure, and parameter transition areas are constructed between regions to smooth the region boundaries; The objective function is solved using the gradient descent method to obtain the super-resolution results.

2. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 1, characterized in that, Before segmenting the image, we first construct an orientation convolution signal model: The target area is imaged using an airborne real-aperture scanning radar to obtain a real-beam image. The study focuses solely on the amplitude of the azimuth signal, which is denoted as: （1） in, Here are the azimuth coordinates of the target. The scattering intensity of a point target. For convolution operations, Antenna radiation pattern; considering noise The influence of the azimuth signal is expressed as follows: （2） Based on the above model, construct an azimuth convolution signal vector model: （3） in, and These represent the echo data and target scattering coefficients within a range gate, respectively. For noise vectors, The number of elements in the vector. Antenna radiation pattern The constructed convolution matrix.

3. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 2, characterized in that, The imaging objective function is constructed as follows: To address the noise sensitivity issue in deconvolution operations, regularization is employed. The standard regularization equation is: （4） in, For data fidelity items, For regularization terms, Constraints used to balance data fidelity and regularization terms; Choose a weighted combination of the L1 norm, L2 norm, and total variation regularization term as the joint regularization term: （5） in, , , These are the coefficients of the L1 regularization term, the L2 regularization term, and the total variation regularization term, respectively, used to balance the data fidelity term with the constraints of each regularization term; For L1 regularization terms, For L2 regularization terms, For the total variation regularization term, its formula is: （6） Substituting equations (5) and (6) into equation (4), the objective function for super-resolution imaging of RASR can be written as: （7）。 4. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 3, characterized in that, The gradient of the objective function is obtained as follows: Due to the nondifferentiability of the L1 norm, a minimal constant is introduced. Formula (13) can be written as: （8） in, The gradient descent method is used to solve for the minimum value, and the gradient of the objective function is obtained. （9） in, It is a diagonal matrix. .

5. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 4, characterized in that, The image is segmented using the K-means algorithm, specifically as follows: Apply Gaussian filtering to the image, defining a size of... Gaussian kernel: （10） in, This indicates the offset of the current pixel relative to the center of the filter. It is the standard deviation of the Gaussian distribution, controlling the degree of smoothing; for the input image Output image The value of each pixel is a weighted average of the values ​​of its neighboring pixels: （11） The goal of image segmentation using the K-means algorithm is to minimize the sum of squared errors within clusters. （12） in, It is the grayscale value of each pixel in the image. It is the center point of each cluster. It is the division of each cluster. yes To its cluster center The Euclidean distance; arrange all pixels in the image according to their grayscale values ​​from low to high, and initialize K cluster centers, with the value of each cluster center being: （13） in, Given the total number of pixels in the image; calculate the distance from the grayscale value of each pixel to all cluster centers, and assign the pixel to the nearest cluster: （14） This refers to the partitioning of each cluster in the t-th iteration; In the t-th iteration, The center point of the cluster to which it belongs; In the t-th iteration, except The center point of any cluster other than the one it belongs to; After classification is completed, the average gray level of all pixels in each class is calculated, and the result is used to update the new cluster centers for that class. （15） The K-means algorithm clustering steps are continuously iterated until the cluster centers no longer change, at which point the loop terminates. Based on prior knowledge of the image, the divided regions are merged to obtain the final region segmentation result.

6. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 5, characterized in that, To smooth the region boundaries, parameter transition zones are constructed between different regions, specifically as follows: The regularization parameter for the transition region between region A and region B is given by the following formula: （16） in, Point Distance to region A, This represents the maximum distance from all points within the transition area to region A. and These represent the regularization parameters for regions A and B, respectively; the closer a point in the transition region is to a certain region, the closer its regularization parameter is to that region.

7. The content-adaptive super-resolution imaging method for real aperture scanning radar based on K-means image segmentation according to claim 6, characterized in that, Based on prior knowledge of the local image, different regularization parameters are set for each divided region to construct a full-image regularization coefficient matrix; Substitute the regularization coefficient matrix into the gradient of the objective function, and use the gradient descent method to solve the objective function until convergence or the maximum number of iterations is reached; randomly select a starting point. Take the input image and calculate the gradient at the current point. And adjust the parameters along the negative gradient direction: （17） in, It is the pixel value updated at step k+1; The step size controls the magnitude of each update. After each iteration, the objective function value decreases. Through gradual convergence, the objective function value eventually stabilizes near a local minimum, resulting in the reconstructed image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images