A method and system for joint noise image filtering and target enhancement based on MIMO-SAR

Abstract

The application provides a joint noise image filtering and target enhancement method and system based on MIMO-SAR, relates to the technical field of radar image processing, and comprises collecting echo data, remodeling the data matrix to obtain structured data, performing coarse focusing after obtaining slice data matrix through target region signal preprocessing and preliminary extraction, jointly using CA-CFAR algorithm and DBSCAN algorithm to perform information extraction on a target plane after wavelet shrinkage denoising, performing multiple iteration screening on data points corresponding to the target, performing index remapping, rearrangement and focusing on selected target pixel points, and obtaining an accurate target image; constructing and training an image enhancement network based on a CycleGAN model, and performing image enhancement through the image enhancement network. Noise filtering and target image enhancement of the MIMO-SAR image are realized from two stages of algorithm processing and image processing, the average reduction of the target image entropy is 3.42 bits, the average improvement of the structural similarity index of the processed image is 0.10, and the average improvement of the peak signal-to-noise ratio is 1.87 dB.

Description

Technical Field

[0001] This application relates to the field of radar image processing technology, and more specifically, to a method and system for joint noise image filtering and target enhancement based on MIMO-SAR. Background Technology

[0002] Due to factors such as channel random noise, obstruction, and complex backgrounds, SAR radar imaging suffers from reduced target saliency and image aliasing / blurring. Furthermore, issues such as lead screw jitter and mechanical delays in mechanical data acquisition platforms further introduce more channel random phase noise. In recent years, neural network-based noise filtering techniques have significantly improved noise processing capabilities, such as:

[0003] (1) Separating the feature frequency domain and suppressing high-frequency noise: Chinese invention patent with publication number CN120374604A discloses an optically guided SAR target detection method based on frequency domain enhancement and dynamic masking, including acquiring SAR images and corresponding optical images, using a pre-trained optical detection model as the teacher model and the SAR model to be trained as the student model; the optical and SAR images are respectively input into the corresponding models to generate feature maps, and the SAR features are decomposed into low-frequency global and high-frequency detail components through wavelet transform, and noise is suppressed and target features are enhanced through multi-scale convolution and self-attention mechanism.

[0004] (2) Noise Removal Scheme Based on Feature Clustering: Chinese Invention Patent CN117934856A discloses a method, device, storage medium and electronic device for noise removal of point cloud data. By constructing a front view, the three-dimensional initial point cloud data transmitted by the lidar is converted into a two-dimensional front view, overcoming the problems of large computational load, low efficiency and incomplete noise removal. The front view is divided into multiple initial sets, and each initial set is denoised to obtain the statistical number of pixels in each pixel value range. The peak range is determined based on the statistical number. The effective area in the initial set is determined based on the peak range and the width matched by the peak range. Pixels outside the effective area are regarded as noise points in the initial set and removed to obtain the feature point set. The peak range is determined according to the statistical number, and then the coordinate point corresponding to the detection object is determined based on the peak range to complete the denoising.

[0005] However, existing research still has two limitations: First, due to the large number of environmental noise factors and the non-fixed surface features of the target, if neural networks are used directly for image processing, the network's understanding of the target under different aliased states is prone to discrepancies, which increases the burden on network construction and training. Second, the random jitter of the system's motion model may cause the target image to be blurred and the target reflectivity to be inconsistent, resulting in incomplete target images after processing. Summary of the Invention

[0006] To address the aforementioned problems, this application employs a joint noise image filtering and target enhancement method based on MIMO-SAR, comprising the following steps:

[0007] Algorithm processing stage: Echo data is acquired, and the data matrix is ​​reshaped to obtain structured data. Then, after the target area signal preprocessing and preliminary extraction process, a slice data matrix is ​​obtained. The slice data matrix is ​​coarsely focused to obtain the pixel matrix corresponding to the target plane. Then, wavelet shrinkage denoising is performed. After wavelet shrinkage denoising, the CA-CFAR algorithm and DBSCAN algorithm are used together to extract information from the target plane. The data points corresponding to the target are iteratively filtered multiple times. Finally, the selected target pixels are indexed and remapped. The data at these positions are rearranged and focused to obtain an accurate target image.

[0008] Image processing stage: An image enhancement network is built and trained based on the CycleGAN model, and the image enhancement network is used to enhance the accurate target image obtained in the algorithm processing stage.

[0009] Optionally, the target area signal preprocessing and preliminary extraction process includes virtual channel calibration of structured data, then reorganizing the calibrated multi-channel data according to spatial location and time dimensions to construct a planar aperture data matrix, performing a two-dimensional fast Fourier transform on the planar aperture data matrix to obtain a data cube, and then sequentially performing complex gain compensation, fast time index filtering, and target data slicing to obtain a slice data matrix.

[0010] Optionally, the joint use of the CA-CFAR algorithm and the DBSCAN algorithm to extract information from the target plane includes: using the wavelet shrinkage and denoising data as the input of the CA-CFAR algorithm to obtain the initial screening result of the target pixels, using the initial screening result of the target pixels as the input of the DBSCAN algorithm to obtain the target cluster and the noise cluster, and performing a secondary screening of the target cluster based on the target initial screening and the target initial screening bounding box range obtained by morphological analysis.

[0011] Optionally, the CA-CFAR algorithm includes taking the wavelet-shrinkable and denoised data as input, setting the units adjacent to the unit to be detected as protection units, calculating the average power of reference units outside the protection units around the unit to be detected to determine the initial screening threshold, and determining whether the unit to be detected is the target unit based on the initial screening threshold to obtain the initial screening result of the target pixel.

[0012] Optionally, the DBSCAN algorithm includes taking the initial screening results of the target pixels as input, and performing density clustering on the target data point cloud sequentially based on the preset target radius and threshold to obtain target clusters and noise clusters.

[0013] Optionally, obtaining the target radius specifically includes:

[0014] S1: Given a dataset P={p(i);i=0,1,…,n}, for any point P(i), calculate the distance between point P(i) and all points in set P. The distances are sorted in ascending order, and the sorted distance set is D={d(1),d(2),…,d(k-1),d(k),d(k+1),…,d(n)}. Then d(k) is called the k-distance, which is the k-th closest distance between point p(i) and all other points except p(i). Calculate the k-distance for each point p(i) in the cluster set, and finally obtain the k-distance set of all points E={e(1),e(2),…,e(n)}.

[0015] S2: Based on the obtained set E of k-distances to all points, sort the set E in ascending order to obtain the set E' of k-distances. Fit the change curve of k-distances in set E' and determine the value of k-distance corresponding to the position with the largest slope as the value of the target radius.

[0016] S3: Repeat the S1-S2 process iteratively to test the optimal preset parameters for the k value;

[0017] The specific steps for obtaining the filtering threshold are as follows: After DBSCAN clustering, obtain the number of points in each cluster and sort them from largest to smallest; draw a scatter plot with the X-axis representing the cluster ranking and the Y-axis representing the size of the corresponding cluster, fit the scatter plot into a curve, select the position with the largest downward slope of the curve, and take the cluster size value of that position as the filtering threshold. Clusters below the threshold are defined as noise clusters, and clusters above or equal to the threshold are defined as target clusters.

[0018] Optionally, the input for target screening and morphological analysis is the wavelet shrinkage and denoising data. Target screening and morphological analysis include threshold segmentation, connected component analysis and target classification. The output is the target screening bounding box range.

[0019] Optionally, the image enhancement network includes two pairs of generative adversarial links, including two generators G1 and G2 and two discriminators D1 and D2. G1 and G2 have the same architecture, and D1 and D2 have the same architecture. Based on the cycle consistency loss constraining the training direction, the image enhancement network is trained to obtain a generator G1 that can transform a blurry and incomplete image into a clear image for use in the image enhancement stage of image processing.

[0020] Among them, G1 and G2 are built based on residual networks and include an initial convolutional layer, a downsampling layer, a residual block, an upsampling layer and a final convolutional layer. D1 and D2 are composed of one-dimensional convolutional layers, which are used to extract features from the image and determine the category of the feature, and finally output a one-dimensional discrimination result.

[0021] Optionally, let the feature domain of the incomplete / blurred image be... Features The enhanced, clearer image feature domain is Features Generators are used in domains and Establish a mapping between the two generator mappings. and Two discriminators mapped and , right and To distinguish; right and To differentiate, and These are the features generated by generators G1 and G2, respectively. The objective function of the generators includes adversarial loss, cycle consistency loss, and identity loss.

[0022] This application also provides a joint noise image filtering and target enhancement system based on MIMO-SAR, applicable to any of the aforementioned joint noise image filtering and target enhancement methods based on MIMO-SAR, including:

[0023] Algorithm processing module: Acquire echo data, reshape the data matrix to obtain structured data, then perform target area signal preprocessing and preliminary extraction to obtain slice data matrix, coarsely focus the slice data matrix to obtain the pixel matrix corresponding to the target plane, then perform wavelet shrinkage denoising, after wavelet shrinkage denoising, the CA-CFAR algorithm and DBSCAN algorithm are used together to extract information from the target plane, and the data points corresponding to the target are iteratively filtered multiple times. Finally, the selected target pixels are indexed and remapped, and the data at these positions are rearranged and focused to obtain an accurate target image;

[0024] Image processing module: Used to build and train an image enhancement network based on the CycleGAN model, and to enhance the accurate target image obtained in the algorithm processing stage through the image enhancement network.

[0025] The beneficial effects of the joint noise image filtering and target enhancement method and system based on MIMO-SAR provided in this application are as follows:

[0026] (1) This application realizes MIMO-SAR image noise filtering and target image enhancement in two stages: algorithm processing and image processing. In the algorithm processing stage, the channel is first processed using the wavelet shrinkage algorithm to ensure the accuracy of range-oriented target selection; then, CFAR and DBSCAN are used to handle the target pixel initial screening and pixel cluster sorting modules respectively, and the two modules are linked to realize three-iteration joint noise image filtering based on the coarse focusing pixel matrix, and the target pixels are indexed, remapped and refocused. In terms of image processing, an image enhancement network is designed based on the CycleGAN model, and large model learning is used to complete the transformation from the blurred and incomplete image feature domain to the clear feature domain.

[0027] (2) According to the test, the joint screening of CFAR-DBSCAN provided in this application reduced the entropy of the reconstructed target image by an average of 3.42 bits, making the target features more concentrated. The MIMO-SAR image enhancement network subsequently constructed improved the structural similarity index of the processed image by an average of 0.10 and the peak signal-to-noise ratio by an average of 1.87dB compared with the unprocessed result. The processed image quality is more complete and closer to the clear target image. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0029] Figure 1 The overall system block diagram of a joint noise image filtering and target enhancement method based on MIMO-SAR provided in the embodiments of this application;

[0030] Figure 2 This is a flowchart of wavelet denoising provided in an embodiment of this application;

[0031] Figure 3 This is a general flowchart of CFAR detection provided in the embodiments of this application;

[0032] Figure 4 This is a schematic diagram of the 2D-CFAR detection principle provided in the embodiments of this application;

[0033] Figure 5 This is a schematic diagram of cluster partitioning for the DBSCAN algorithm provided in the embodiments of this application;

[0034] Figure 6 A flowchart of the joint noise removal algorithm based on CFAR and DBSCAN provided for embodiments of this application;

[0035] Figure 7 This is a flowchart illustrating the construction of the joint noise removal filter provided in an embodiment of this application.

[0036] Figure 8 To generate a basic structural diagram of the adversarial link network;

[0037] Figure 9 This is a generative adversarial link in the image enhancement network provided in the embodiments of this application;

[0038] Figure 10 This is a structural diagram of the image enhancement network generator provided in the embodiments of this application;

[0039] Figure 11 This is a planar target noise simulation scene provided in the embodiments of this application - target and scanning planar view;

[0040] Figure 12 This is a side view of a planar target noise simulation scene provided in an embodiment of this application.

[0041] Figure 13 This is a schematic diagram illustrating the generation of the noise matrix and the noise-added data provided in the embodiments of this application;

[0042] Figure 14 This is an image of the undenoised imaging result provided in an embodiment of this application;

[0043] Figure 15 The image shows the single-channel result of wavelet shrinkage processing provided in the embodiments of this application, where (a) is the original signal spectrum; (b) is the signal spectrum with noise; and (c) is the signal spectrum after denoising.

[0044] Figure 16 This is a preliminary imaging result diagram provided for an embodiment of this application;

[0045] Figure 17 This is a diagram showing the morphological bounding box result of the joint noise filtering based on CFAR-DBSCAN provided in the embodiments of this application;

[0046] Figure 18 The image shows the 2D-CFAR result of joint noise filtering based on CFAR-DBSCAN provided in the embodiments of this application.

[0047] Figure 19 This is a diagram showing the results of joint noise filtering based on CFAR-DBSCAN and DBSCAN cluster classification, provided in an embodiment of this application.

[0048] Figure 20 The image shows the results of the combined noise filtering based on CFAR-DBSCAN and the secondary screening by DBSCAN, as provided in the embodiments of this application.

[0049] Figure 21 The image shows the result of denoising using the algorithm in this paper after joint noise filtering based on CFAR-DBSCAN, as provided in the embodiments of this application.

[0050] Figure 22 The target echo image after joint noise filtering and denoising based on CFAR-DBSCAN provided in the embodiments of this application;

[0051] Figure 23 The following images show partial results of the corner code feature dataset provided in this application embodiment: (a) is the image after parameter processing for experimental group 1; (b) is the image after parameter processing for experimental group 2; (c) is the image after parameter processing for experimental group 3; (d) is the image after parameter processing for experimental group 4; (e) is the image after parameter processing for experimental group 5; and (f) is the image after parameter processing for experimental group 6.

[0052] Figure 24 A partial diagram showing the structure of the MIMO-SAR image enhancement network training dataset provided in the embodiments of this application.

[0053] Figure 25 The images shown are test results of the MIMO-SAR image enhancement network provided in this application embodiment, wherein (a) is a clear image of knife 1; (b) is a clear image of knife 2; (c) is a clear image of simulation 1; (d) is a clear image of simulation 2; (e) is a clear image of the model gun; (f) is an image of knife 1 before processing; (g) is an image of knife 2 before processing; (h) is an image of simulation 1 before processing; (i) is an image of simulation 2 before processing; (j) is an image of the model gun before processing; (k) is an image of knife 1 after processing; (l) is an image of knife 2 after processing; (m) is an image of simulation 1 after processing; (n) is an image of simulation 2 after processing; and (o) is an image of the model gun after processing. Detailed Implementation

[0054] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0055] like Figure 1 As shown, this application provides a method for joint noise image filtering and target enhancement based on MIMO-SAR, including the following steps:

[0056] Algorithm processing stage: Echo data is acquired, and the data matrix is ​​reshaped to obtain structured data. Then, after the target area signal preprocessing and preliminary extraction process, a slice data matrix is ​​obtained. The slice data matrix is ​​coarsely focused to obtain the pixel matrix corresponding to the target plane. Then, wavelet shrinkage denoising is performed. After wavelet shrinkage denoising, the CA-CFAR algorithm and DBSCAN algorithm are used together to extract information from the target plane. The data points corresponding to the target are iteratively filtered multiple times. Finally, the selected target pixels are indexed and remapped. The data at these positions are rearranged and focused to obtain an accurate target image.

[0057] Image processing stage: An image enhancement network is built and trained based on the CycleGAN model, and the image enhancement network is used to enhance the accurate target image obtained in the algorithm processing stage.

[0058] Noise removal process as follows Figure 2 As shown, firstly, wavelet decomposition transform is performed on the observed data. Then, the wavelet coefficients are thresholded using a soft threshold and Bayesian estimation is used to optimize the threshold. Finally, the wavelet coefficients after the above processing are subjected to inverse transform to reconstruct the signal, ultimately obtaining the denoised form of the contaminated sampled signal.

[0059] Wavelet contraction denoising utilizes the characteristics of orthogonal wavelet transform and Gaussian white noise to threshold-quantize the wavelet decomposition coefficients of the contaminated signal, detecting the true signal from the additive white Gaussian noise and performing distortion-free reconstruction. Assuming the noise model of a single channel is additive white Gaussian noise and the signal length is N, the contaminated received signal can be represented as:

[0060] ;

[0061] In the formula, The sampled signal contains noise. For an ideal sampling signal, It is independent and identically distributed Gaussian white noise and has , Noise level.

[0062] By utilizing the differences in signal and noise characteristics under wavelet transform, the wavelet decomposition coefficients are processed to distinguish the target signal components from the noise signal components, thus recovering the true signal from the noise-contaminated signal. Taking a three-level decomposition as an example (i=1,2,3), the wavelet decomposition coefficients are processed... The decomposition can be represented as:

[0063] ;

[0064] In the formula, The approximate signal part of the decomposition, The detailed signal portion is decomposed. The noise portion is included. , , By using a reasonable threshold to filter the wavelet coefficients, the purpose of denoising and reconstructing the signal can be achieved.

[0065] "Using a reasonable threshold" refers to wavelet coefficient shrinkage based on soft thresholding, building upon the sqtwolog threshold. The specific calculation process is as follows:

[0066] In the formula, The noise standard deviation is obtained from the median absolute deviation (MAD); N is the total number of data points.

[0067] Let the wavelet coefficients be w and the threshold be T:

[0068] The processing formula is as follows: .

[0069] As a feasible implementation, wavelet coefficient shrinkage can also be achieved through hard thresholding. Coefficients with absolute values ​​less than a given threshold are set to zero, while coefficients with absolute values ​​greater than or equal to the threshold remain unchanged. The processing formula is as follows: .

[0070] The target area signal preprocessing and preliminary extraction process includes virtual channel calibration of structured data, followed by reorganizing the calibrated multi-channel data according to spatial location and time dimensions to construct a planar aperture data matrix. The planar aperture data matrix is ​​then subjected to a two-dimensional fast Fourier transform to obtain a data cube, which is then subjected to complex gain compensation, fast time index filtering, and target data slicing to obtain a slice data matrix.

[0071] The method of extracting information from the target plane by jointly using the CA-CFAR algorithm and the DBSCAN algorithm includes: using the wavelet shrinkage and denoising data as the input of the CA-CFAR algorithm to obtain the initial screening result of the target pixels; using the initial screening result of the target pixels as the input of the DBSCAN algorithm to obtain the target cluster and the noise cluster; and performing a secondary screening of the target cluster based on the target screening bounding box range obtained by the initial screening and morphological analysis.

[0072] This module, based on the preprocessing results of the wavelet denoising channel, jointly utilizes the CA-CFAR and DBSCAN algorithms to extract useful information from the target plane. It then performs three iterative screenings on the data points corresponding to the target, gradually distinguishing the target pixels from background clutter pixels that are not part of the target source. Finally, the selected target pixels are indexed and remapped, and the data at these locations is rearranged and focused to obtain the final noise-filtered image. The specific algorithm flow is as follows: Figure 6 As shown.

[0073] The input to the CFAR target pixel screening module is the wavelet shrinkage and denoising data. The presence of a target at a detection point is determined by comparing the power of the detection point with the average power of the background. The defined protection area isolates the detection from the influence of other target points.

[0074] To prevent a decrease in edge detection rate, the original pixel matrix is ​​padded with zeros to fill adjacent pixels. The increased number of units improves edge detection performance. This paper selects a detection threshold... The detection unit is set as follows:

[0075] ;

[0076] The protection unit is set as follows:

[0077] ;

[0078] In the formula, This represents the detection unit parameters in the x-direction. This represents the detection unit parameters in the y-direction. This represents the feature quantity of the preprocessed data in the x-direction. This represents the feature quantity of the preprocessed data in the y-direction. Indicates the protection unit parameters in the x-direction. This represents the protection unit parameter in the y-direction. Using a mask, data that is judged as background is set to zero. The output is a new data matrix that retains only the elements of the initially screened target, filtering out non-target interference sources while preserving the complete pixels of the target.

[0079] The DBSCAN pixel cluster sorting module takes as input the initial screening results of target pixels, with the data index positions corresponding to the 2D sampling plane. Utilizing the characteristics of continuous target distribution and random noise distribution, it separates and removes noise data points in the data matrix, outputting a cluster sorting matrix of targets and noise.

[0080] The process of obtaining the target radius and threshold of DBSCAN is as follows:

[0081] S1: Given a dataset P={p(i);i=0,1,…,n}, for any point P(i), calculate the distance between point P(i) and all points in set P. The distances are sorted in ascending order, and the sorted distance set is D={d(1),d(2),…,d(k-1),d(k),d(k+1),…,d(n)}. Then d(k) is called the k-distance, which is the k-th closest distance between point p(i) and all points (except p(i). Calculate the k-distance for each point p(i) in the cluster set, and finally obtain the k-distance set of all points E={e(1),e(2),…,e(n)}.

[0082] S2: Based on the obtained set E of k-distances to all points, sort the set E in ascending order to obtain the set E' of k-distances. Fit the change curve of k-distances in set E' and determine the value of k-distance corresponding to the position with the largest slope as the value of the target radius.

[0083] S3: Repeat the S1-S2 process to iteratively test the optimal preset parameters for the k value.

[0084] The specific steps for obtaining the filtering threshold are as follows: After DBSCAN clustering, obtain the number of points in each cluster and sort them from largest to smallest; draw a scatter plot with the X-axis representing the cluster ranking and the Y-axis representing the size of the corresponding cluster, fit the scatter plot into a curve, select the position with the largest downward slope of the curve, and take the cluster size value of that position as the filtering threshold. Clusters below the threshold are defined as noise clusters, and clusters above or equal to the threshold are defined as target clusters.

[0085] Target radius in this embodiment , The number of target points within a certain range is defined as the spatial density of that target point, with a preset threshold. Density clustering is performed on the target data point cloud sequentially to filter out noise clusters and real target clusters. Clusters with spatial density below a threshold are identified as noise clusters, while clusters with spatial density above a threshold are identified as target clusters.

[0086] The cluster classification results are then subjected to a second screening, classifying clusters whose spatial locations are outside the initial target selection frame as noise. Finally, clusters identified as noise in the classification results are removed, completing the removal of background noise. This process can be viewed as the construction of a denoising filter, where the filter's system function is the calibrated target index, such as... Figure 7 As shown.

[0087] The target preliminary screening bounding box range is obtained based on the target preliminary screening and morphological analysis. The input of the target preliminary screening and morphological analysis is the data after wavelet shrinkage and denoising. The target preliminary screening and morphological analysis includes threshold segmentation, connected component analysis and target classification. The output is the target preliminary screening bounding box range.

[0088] The above process combines CFAR and DBSCAN methods to gradually separate and filter targets to achieve accurate target reconstruction. At the same time, morphological closing operation is added to constrain the noise judgment during DBSCAN processing. By combining 2D-CFAR, morphological bounding box selection and clustering methods, the accurate reconstruction of the target is finally achieved.

[0089] The CA-CFAR algorithm includes taking a coarsely focused pixel signal matrix as input, setting the cells immediately adjacent to the cell to be detected as guard cells, calculating the average power of reference cells outside the guard cells around the cell to be detected to determine the initial screening threshold, and determining whether the cell to be detected is the target cell based on the initial screening threshold to obtain the initial screening result of the target pixel.

[0090] like Figure 3 and Figure 4 As shown, this module uses Cell-Averaging Constant False Alarm Rate (CA-CFAR) for target detection. CA-CFAR determines the threshold by calculating the average power of reference cells surrounding the cell under test (CUT). Specifically, it uses the average of the sum of the power values ​​of the reference cells surrounding each CUT as an estimate of the noise power. To avoid the influence of the CUT's own power, the cells immediately adjacent to the CUT are set as guard cells. When the power value of a CUT cell exceeds the average power value of all reference cells on both sides, the CUT is determined to be the target cell to be detected. The detection threshold of CA-CFAR is... The calculation formula is as follows:

[0091] ;

[0092] In the formula This represents the threshold factor (typically between 0.1 and 0.5). Let P(i,j) represent the total number of detection units, where P(i,j) represents the power value of the reference unit in the i-th row and j-th column of the two-dimensional plane.

[0093] The DBSCAN algorithm involves taking the initial screening results of target pixels as input, and performing density clustering on the target data point cloud sequentially based on the preset target radius and threshold to obtain target clusters and noise clusters.

[0094] like Figure 5As shown, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm that assumes the clusters are related to the density of the sample distribution. It identifies clusters through dense regions of sample points and is highly robust to noisy points. Point relationships in DBSCAN include density connectivity, density reachability, and density direct reachability. These relationships define how to traverse from a core point to the entire cluster, progressively sorting the input point set into sets of points with different functions.

[0095] Neighborhood radius and minimum number of points These are two important parameters of DBSCAN. The objects are mainly divided into core points, boundary points, and noise points. Core points are those whose neighborhood radius contains more than or equal to the minimum number of points; boundary points are those adjacent to core points but do not themselves meet the core point criteria; points outside the core and boundary points are noise points. The DBSCAN algorithm classifies and filters samples through category analysis and processing. The processing flow is shown in Table 1.

[0096] Table 1. Pseudocode of the DBSCAN algorithm

[0097]

[0098] This module takes two-dimensional pixel data as input, sorts the target into different categories through pixel clustering, and finally outputs the clustering results.

[0099] like Figure 7 As shown, the input for target screening and morphological analysis is a coarsely focused pixel signal matrix. Target screening and morphological analysis include threshold segmentation, connected component analysis, and target classification. The output is the target screening bounding box range.

[0100] like Figure 8 As shown, generator G generates images, taking random noise as input and outputting fake images. Discriminator D's task is to distinguish whether G's output is a real or fake image, outputting the probability that the image is real. In this cyclical game, G aims to generate the most realistic images possible to "deceive" D, while D continuously optimizes its discrimination ability by analyzing a large number of real and fake samples. If D identifies the image generated by G as fake, it will "penalize" G through a loss function. G adjusts its network parameters based on D's feedback to generate more realistic images. Both are trained simultaneously, and when Nash equilibrium is reached, the data distribution generated by G completely overlaps with the real data distribution, and D cannot distinguish between generated and real samples. At this point, both G and D have reached conditional optimality.

[0101] like Figure 9 As shown, the image augmentation network provided in this application is a CycleGAN (Cycle-Consistent GAN), consisting of two pairs of generative adversarial links, including two generators (G1, G2) and two discriminators (D1, D2). G1 and G2 have identical architectures and are used to learn and transform features from two image domains; D1 and D2 have identical architectures and focus on local image features to determine whether the image is real or generated. To ensure consistency in the transformation between G1 and G2, a cycle consistency loss is introduced to constrain the training direction. The goal of this section is to obtain generator G1, which can transform blurry and incomplete images into clear images, thus completing the augmentation of MIMO-SAR images.

[0102] like Figure 10 As shown, G1 and G2 are both built on ResNet, containing an initial convolutional layer, a downsampling layer, a residual block, an upsampling layer, and a final convolutional layer. The structures of the two Ds consist of one-dimensional convolutional layers, used to extract features from the image and determine the category of those features, ultimately outputting a one-dimensional discrimination result.

[0103] Let the feature domain of the incomplete / blurred image be . Features The enhanced, clearer image feature domain is Features The main purpose of G is in the domain. and Establish a mapping between the two generator mappings. and Two discriminators mapped and , right and To distinguish; right and The objective function of G includes adversarial loss, cycle consistency loss, and identity loss.

[0104] Adversarial loss is used to ensure that the generated image is indistinguishable from the target image in terms of distribution. For generator mapping... and its discriminator The objective function for combating loss is:

[0105] ;

[0106] For generators and its discriminator The objective function for combating loss is:

[0107] ;

[0108] In D, samples are generated for discrimination. or real sample After that, G will continue to generate with More similar domain samples The goal of G is to minimize the objective function value, while the goal of D is to maximize the objective function value. For the source domain... Image features After mapping Transform to target domain Then after mapping Convert back to source domain At that time, it should be able to restore the original image. During training, the periodicity of the mapping function should be guaranteed. Cyclic consistent loss is used to ensure the consistency of the transformations between two Gs, and its objective function is defined as:

[0109] ;

[0110] In the formula This represents the L1 norm, used to measure pixel-level differences between two images.

[0111] In generator mapping In the middle, through samples The generated result and Samples within the same domain are similar. Similarly, through samples... generated It should also be consistent with the original sample. Similar. Therefore, the original features of the input can be preserved through identity loss. The identity loss function has the following form:

[0112] ;

[0113] Based on the above definitions of the sub-network split loss function, it can be deduced that the overall network loss function consists of a forward-biased network and a backward-biased network. Therefore, the complete loss function is a weighted sum of the individual sub-loss functions. The overall loss function can be defined as:

[0114] ;

[0115] In the formula Indicates the weight of the cycle-consistent loss term. This represents the weight of the identity loss function term.

[0116] System testing and verification:

[0117] (1) Simulation test of noise filtering algorithm

[0118] First, the near-field planar target is constructed, and the simulation scene is as follows: Figures 11-12 As shown.

[0119] The target is a perforated metal plate with a reflectivity of 0.8 for the metal portion and a perforated pattern of uppercase English letters. The plate measures 130mm in size. 90mm. The radar scanning plane is parallel to the target plane and aligned at the center. Random point targets are used to simulate strong non-target interference between two consecutive slices of the target area. The reflectivity of different point targets is taken as random values ​​within a range. The designed planar target area size is 300mm. 300mm, the synthetic aperture of the radar is 400mm. 400mm, the distance between the target plane and the scanning plane is 800mm.

[0120] Based on a signal model with a Gaussian noise channel, and assuming a preset signal-to-noise ratio... The Gaussian white noise is designed based on the average power of each channel signal. The original data matrix is ​​iteratively calculated according to the channel's position on the sampling plane to obtain the full-channel noise matrix. The linear property of Gaussian noise is then used to add this noise matrix to the original data, resulting in a result consistent with the preset... Similar noisy data, detailed process as follows Figure 13 As shown.

[0121] Figure 13 The diagram shows channel 1, channel k, and channel n, where k represents any value between 1 and n, and n is the total number of channels. The total number is determined by the MIMO antenna configuration; in this embodiment, the radar measured 2 transmit and 4 receive channels, therefore there are 8 virtual channels.

[0122] Traditional near-field SAR imaging algorithms will not be elaborated upon in this paper. Preliminary imaging results under the influence of noise are as follows: Figure 14 As shown:

[0123] The wavelet shrinkage algorithm was used to denoise the virtual channel data. The simulation data was sampled in two channels. The experiment selected the sym8 wavelet basis, the decomposition level was 3, the threshold was a soft threshold, and Bayesian estimation was used to optimize the threshold.

[0124] (2) Image Augmentation Network Dataset Creation and Testing

[0125] Training the CycleGAN network requires two pairs of four datasets. The two pairs refer to the training set and the test set, where the ratio of images in the training set to the test set is approximately 4:1. The four sets refer to the data where images of different clarity classes are divided into two groups based on the proportions of the training and test sets, resulting in four datasets used for network training. Each image in the dataset has a pixel size of 256. 256. Name the folder containing the clear training set images "trainA", the folder containing the blurry training set images "trainB", the folder containing the clear test set images "testA", and the folder containing the blurry test set images "testB". Classify and organize the MIMO-SAR result images in the above manner to obtain the final dataset used for network training.

[0126] The construction of blurred and incomplete features in the results was achieved by increasing the sampling interval and the target filtering threshold. Taking target cheques as an example, two experimental scenarios were collected: one with occlusion interference and the other without. When there is occlusion interference, filtering out the noise image may lead to incomplete target images. Furthermore, using different sampling intervals in the two scenarios also causes different degrees of aliasing. The experimental parameters used for dataset feature construction are shown in Table 2.

[0127] Table 2 Experimental parameters for constructing the feature dataset

[0128]

[0129] To quantitatively evaluate the performance of the noise removal algorithm presented in this paper, the SSIM parameter is introduced to analyze the imaging results before and after processing. SSIM is a metric used to measure the similarity between two images, considering not only the difference in grayscale values ​​but also the structural information of the images. Its range is defined within... Between the two numbers, 1 represents that the two images are exactly the same, 0 represents that the two images are not similar, and -1 represents that the two images are completely opposite.

[0130] Let x and y be the two images before and after network processing, with mean values ​​of respectively. and The standard deviations are respectively and covariance is The luminance similarity index between the two images is: Its form is:

[0131] ;

[0132] In the formula, It is a constant, so avoid having a denominator of 0.

[0133] Let the contrast similarity index of the two images be denoted as . Its form is:

[0134] ;

[0135] In the formula, It is a constant, so avoid having a denominator of 0.

[0136] Let the structural similarity index between the two images be denoted as . Its form is:

[0137] (0-1)

[0138] In the formula, —A constant, to avoid a denominator of 0.

[0139] For constants The choice of L is usually defined with reference to the dynamic range L of the image, denoted as L. It is a small constant, usually taking the value of The reference values ​​for the three constants are as follows:

[0140] ;

[0141] remember If the weights are denoted as , then the final SSIM exponent is a weighted product of the three parts:

[0142] ;

[0143] This section defines the weight parameters. The definition of the SSIM exponent is:

[0144] ;

[0145] To better evaluate the performance of the MIMO-SAR image enhancement network, PSNR is added as an additional parameter in addition to SSIM. PSNR is a metric used to measure image quality, often used to assess the difference between the original and processed images. A higher PSNR value indicates higher image quality after processing. PSNR is calculated based on Mean Squared Error (MSE) and the maximum pixel value of the image. The two images before and after processing are considered... The image size is The MSE value is the average of the squared pixel differences between two images, and its form is:

[0146] ;

[0147] In the formula —Process the pixel values ​​corresponding to the two images before and after.

[0148] Let the maximum pixel value be MAX. Then PSNR is the logarithmic form of the ratio of the square of MAX to MSE(x,y), which is:

[0149] ;

[0150] In the formula, The number of binary bits representing the maximum value of the image pixel range.

[0151] Image quality is a key feature for evaluating visual effects and is a quantification of the observer's perception. Subjective evaluation relies on the judge's personal visual experience and is subject to variability. However, using the evaluation metrics SSIM and PSNR can effectively reflect the quality of images while increasing the objectivity of the judgment.

[0152] (3) Test results of noise filtering algorithm:

[0153] The wavelet shrinkage algorithm results are as follows: Figure 15 As shown, the wavelet shrinkage algorithm effectively highlights the target frequency. The imaging result obtained after initial denoising using the wavelet shrinkage algorithm following MIMO-SAR imaging is shown below. Figure 16 As shown:

[0154] At this point, there is still too much background clutter, requiring further processing. The following section will use the method provided in this application to further process the data, such as... Figures 17-22 As shown.

[0155] Figures 17-22 This paper details the intermediate processes and final results of the innovative noise removal algorithm presented in this paper. Figure 17 As can be seen, morphological bounding can locate the general position of a target, but interference from nearby noise severely affects the determination of boundaries when selecting the target area. Therefore, simply using morphological closing operations will absorb nearby stray noise, making it difficult to accurately capture the target. Figure 18 , Figure 19 , Figure 20 The three key steps of the innovative algorithm presented in this paper are demonstrated: 2D-CFAR, DBSCAN cluster classification, and cluster filtering. The process shows how the coarse focusing results are progressively filtered, classified, and organized. Ultimately, the data points representing the target are retained, while background and noise stray points are classified as noise and subsequently masked out. The results processed by the algorithm are shown below. Figure 21 As shown, the target has been accurately extracted, although the display area is within... Within the range, it is also clear that no extraneous stray points in the background are reconstructed into an image, and this is consistent with... Figure 17By comparison, it's easy to see that not only are excess noise effectively removed, but the edges of the target are also sharper. This is because the algorithm in this paper also achieves pixel-level removal of side lobes. Finally, combining... Figure 21 and Figure 22 As can be seen, the reconstructed image after processing by the innovative denoising algorithm in this paper has a clean background, proving that the algorithm can effectively extract data and achieve accurate imaging.

[0156] (4) Test results of image enhancement network

[0157] according to Figure 6 The process involves processing the echo signal according to different parameters as shown in Table 2, yielding the corresponding results as follows: Figure 23 As shown.

[0158] from Figure 23 The different target results with different characteristics can be clearly seen in the text. Figure 23 From left to right, images (a) to (f) show varying degrees of blurring and incompleteness of the target, consistent with expectations and meeting the requirements for the feature dataset. The resulting images used for training the MIMO-SAR image enhancement network consist of three parts: simulated data imaging results, dataset data imaging results, and measured data imaging results. The dataset is created by modifying image features as described above. Figure 24 As shown, Figure 24 In the "dataset" and Figure 23 The same angle iron imaging is the result of actual measurement data imaging.

[0159] The network was trained using the dataset shown in the image above, employing the Adam optimizer. The weight parameters in the overall loss function of the image augmentation network were configured as follows: While ensuring the transformation of the feature image domain, the original features of the target are preserved. The training parameters are shown in Table 3.

[0160] Table 3 Neural Network Training Parameters

[0161]

[0162] The trained network was tested using five sets of targets with different levels of blur. The target images and the processed result images are shown below. Figure 25 As shown.

[0163] Table 4 Performance metrics corresponding to network test results

[0164]

[0165] The results of network tests using five sets of targets are as follows: Figure 25 As shown, by processing the image before processing ( Figure 25(f), (g), (h), (i), (j)) and the processed image ( Figure 25 The (k), (l), (m), (n), and (o) in the image are respectively compared with the original clear image ( Figure 25 Comparing (a), (b), (c), (d), and (e) in the table, it can be seen that the target after image enhancement is more complete, closer to a clear target image, and has a higher similarity. Loose pixels are effectively removed, and incomplete pixels are effectively filled in. The quantitative analysis results of the test results using SSIM and PSNR are shown in Table 4. The results show that all five experimental groups showed positive improvements. The SSIM values ​​all increased, with an average increase of 0.10, indicating that the image after network processing is closer to a clear target image. The PSNR value increased by an average of 1.87 dB, indicating that the image quality after processing is better than before processing.

[0166] In summary, based on direct observation and quantitative analysis, the MIMO-SAR-based joint noise image filtering and target enhancement method and system provided in this application achieve positive optimization of image quality in the reconstructed image, demonstrating the effectiveness of the method provided in this application.

[0167] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for joint noise image filtering and target enhancement based on MIMO-SAR, characterized in that, Includes the following steps: Algorithm processing stage: Echo data is acquired, and the data matrix is ​​reshaped to obtain structured data. Then, after target region signal preprocessing and preliminary extraction, a slice data matrix is ​​obtained. The slice data matrix is ​​coarsely focused to obtain the pixel matrix corresponding to the target plane. Next, wavelet shrinkage denoising is performed. After wavelet shrinkage denoising, the CA-CFAR and DBSCAN algorithms are jointly used to extract information from the target plane, and three iterations of joint noise image filtering are performed. The method of jointly using the CA-CFAR algorithm and the DBSCAN algorithm to extract information from the target plane includes: using the wavelet shrinkage and denoising data as the input of the CA-CFAR algorithm to obtain the initial screening result of the target pixels, using the initial screening result of the target pixels as the input of the DBSCAN algorithm to obtain the target cluster and the noise cluster, and performing a secondary screening of the target cluster based on the target initial screening and the target initial screening bounding box range obtained by morphological analysis. The input to the target screening and morphological analysis is the wavelet shrinkage and denoising data. The target screening and morphological analysis includes threshold segmentation, connected component analysis and target classification. The output is the target screening bounding box range. Clusters whose spatial location is not within the target initial screening frame are classified as noise. Finally, clusters that are judged to be noise in the sorting results are removed to complete the removal of background noise. Finally, index remapping is performed on the selected target pixels, and the data at these positions is rearranged and focused to obtain an accurate target image. Image processing stage: An image enhancement network is built and trained based on the CycleGAN model, and the image enhancement network is used to enhance the accurate target image obtained in the algorithm processing stage.

2. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 1, characterized in that: The target area signal preprocessing and preliminary extraction process includes virtual channel calibration of the structured data, then reorganizing the calibrated multi-channel data according to spatial location and time dimension to construct a planar aperture data matrix, performing a two-dimensional fast Fourier transform on the planar aperture data matrix to obtain a data cube, and then sequentially performing complex gain compensation, fast time index filtering and target data slicing to obtain a slice data matrix.

3. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 1, characterized in that: The CA-CFAR algorithm includes taking the data after wavelet shrinkage and denoising as input, setting the units adjacent to the unit to be detected as protection units, calculating the average power of the reference units outside the protection units around the unit to be detected to determine the initial screening detection threshold, and determining whether the unit to be detected is the target unit based on the initial screening detection threshold to obtain the initial screening result of the target pixel.

4. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 1, characterized in that: The DBSCAN algorithm includes taking the initial screening results of target pixels as input, performing density clustering on the target data point cloud sequentially based on the preset target radius and threshold, and obtaining target clusters and noise clusters.

5. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 4, characterized in that: The acquisition of the target radius specifically includes: S1: Given a dataset P={p(i);i=0,1,…,n}, for any point P(i), calculate the distance between point P(i) and all points in set P. The distances are sorted in ascending order, and the sorted distance set is D={d(1),d(2),…,d(k-1),d(k),d(k+1),…,d(n)}. Then d(k) is called the k-distance, which is the k-th closest distance between point p(i) and all other points except p(i). Calculate the k-distance for each point p(i) in the cluster set, and finally obtain the k-distance set of all points E={e(1),e(2),…,e(n)}. S2: Based on the obtained set E of k-distances to all points, sort the set E in ascending order to obtain the set E' of k-distances. Fit the change curve of k-distances in set E' and determine the value of k-distance corresponding to the position with the largest slope as the value of the target radius. S3: Repeat the S1-S2 process iteratively to test the optimal preset parameters for the k value; The specific steps for obtaining the filtering threshold are as follows: After DBSCAN clustering, obtain the number of points in each cluster and sort them from largest to smallest; draw a scatter plot with the X-axis representing the cluster ranking and the Y-axis representing the size of the corresponding cluster, fit the scatter plot into a curve, select the position with the largest downward slope of the curve, and take the cluster size value of that position as the filtering threshold. Clusters below the threshold are defined as noise clusters, and clusters above or equal to the threshold are defined as target clusters.

6. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 1, characterized in that: The image enhancement network includes two pairs of generative adversarial links, including two generators G1 and G2 and two discriminators D1 and D2. G1 and G2 have the same architecture, and D1 and D2 have the same architecture. Based on the cycle consistency loss constraining the training direction, the image enhancement network is trained to obtain a generator G1 that can transform blurry and incomplete images into clear images for use in the image enhancement stage of image processing. Among them, G1 and G2 are built based on residual networks and include an initial convolutional layer, a downsampling layer, a residual block, an upsampling layer and a final convolutional layer. D1 and D2 are composed of one-dimensional convolutional layers, which are used to extract features from the image and determine the category of the feature, and finally output a one-dimensional discrimination result.

7. The method for joint noise image filtering and target enhancement based on MIMO-SAR according to claim 6, characterized in that: Let the feature domain of the incomplete / blurred image be... Features The enhanced, clearer image feature domain is Features Generators are used in domains and Establish a mapping between the two generator mappings. and Two discriminators mapped and , right and To distinguish; right and To differentiate, and These are the features generated by generators G1 and G2, respectively. The objective function of the generators includes adversarial loss, cycle consistency loss, and identity loss.

8. A joint noise image filtering and target enhancement system based on MIMO-SAR, characterized in that: The method applicable to the joint noise image filtering and target enhancement based on MIMO-SAR as described in any one of claims 1-7 includes: Algorithm processing module: Acquire echo data, reshape the data matrix to obtain structured data, then perform target area signal preprocessing and preliminary extraction to obtain slice data matrix, coarsely focus the slice data matrix to obtain the pixel matrix corresponding to the target plane, then perform wavelet shrinkage denoising, after wavelet shrinkage denoising, the CA-CFAR algorithm and DBSCAN algorithm are used together to extract information from the target plane, and the data points corresponding to the target are iteratively filtered multiple times. Finally, the selected target pixels are indexed and remapped, and the data at these positions are rearranged and focused to obtain an accurate target image; Image processing module: Used to build and train an image enhancement network based on the CycleGAN model, and to enhance the accurate target image obtained in the algorithm processing stage through the image enhancement network.

Images