A CFAR-guided PolSAR target detection method based on polarimetric characteristics and superpixel segmentation

Abstract

The present invention discloses a CPSAR target detection method based on polarization characteristics and superpixel segmentation guided by CFAR. The method comprises the following steps: rotating a polarization scattering matrix along the radar line of sight to obtain a rotated polarization scattering matrix, and obtaining a polarization correlation pattern based on the rotated polarization scattering matrix; obtaining two polarization correlation features with the largest signal-to-clutter ratio between the target and clutter based on training samples; obtaining a first binary detection result based on the one of the two polarization correlation features with the smaller signal-to-clutter ratio; performing superpixel segmentation on the PolSAR image using a superpixel segmentation algorithm to obtain a superpixel segmentation result; detecting an intensity image corresponding to the PolSAR image using dual-parameter CFAR to obtain a CFAR binary detection result; calculating the features of each superpixel region in the superpixel segmentation result using the CFAR binary detection result to obtain a superpixel feature map; and obtaining a final binary detection result based on the first binary detection result and the superpixel feature map. The present invention improves the performance of PolSAR ground vehicle target detection.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a CFAR-guided PolSAR vehicle target detection method based on polarization characteristics and superpixel segmentation. Background Art

[0002] Space remote sensing systems include optical remote sensing systems (visible light remote sensing, thermal infrared remote sensing, etc.) and microwave remote sensing systems (such as Synthetic Aperture Radar (SAR) systems and microwave radiometers). Traditional optical remote sensing systems operate only in the presence of sunlight and are easily affected by meteorological conditions such as clouds, rain, and fog. SAR systems, on the other hand, are active microwave imaging systems capable of all-day, ground-to-ground observation. Furthermore, SAR systems operate in the microwave band, allowing the radar beam to effectively penetrate clouds and dense fog, enabling all-weather, ground-to-ground observation. The electromagnetic waves in the bands used by SAR also have a certain degree of penetration through vegetation and artificial camouflage, allowing them to detect echoes from targets hidden beneath vegetation or shallowly beneath the surface. Therefore, SAR offers unique advantages in military applications. Compared to single-polarization SAR systems, fully polarized SAR systems can not only obtain the target's radar cross-section (RCS) but also the complete electromagnetic scattering characteristics of the observed target (i.e., the polarization scattering matrix), enriching the dimensionality of the information obtained. Using the polarization scattering matrix, researchers can more deeply analyze and extract the physical parameters and structural information of the target. Therefore, the study of PolSAR (Polarimetric Synthetic Aperture Radar) images has important theoretical significance and practical value.

[0003] Research on detection algorithms based on polarimetric SAR data began in the late 1980s, initiated by researchers at the Lincoln Laboratory. Leveraging their expertise in developing advanced PolSAR systems, the laboratory rapidly acquired a wealth of valuable polarimetric SAR data. Researchers such as Novak, Burl, Irving, and Chaney systematically investigated PolSAR detection algorithms, proposing high-performance PolSAR detectors such as the span detector (span), the optimal polarimetric detector (OPD), the polarimetric matched filter detector (PMF), and the polarimetric whitening filter detector (PWF). Marino studied polarimetric detection algorithms for both single and distributed targets. He subsequently proposed the polarimetric notch filter (PNF), leveraging the differences in polarimetric scattering characteristics between ships and sea clutter, and successfully implemented the detection of ships on the sea surface. Similarly, Wang, Nunziata, and Chen et al. proposed using reflection symmetry (RS) and polarimetric cross-entropy (PCE) to detect ship targets in polarimetric SAR images. More recently, Chen et al. proposed a polarization rotation domain theory and a polarization-related feature for PolSAR data. The core idea is to expand the original polarization coherence matrix into the rotation domain to obtain more hidden information, and they have successfully applied this to the task of detecting ship targets in PolSAR.

[0004] Currently, the main research directions for PolSAR imagery include ship target detection and ground feature classification. However, research on ground target detection is relatively limited, and existing detection algorithms still have some problems. Since the backscatter of radar targets is very sensitive to the relative geometric relationship between the target pose and the radar line of sight, the scattering characteristics of the same target (e.g., a vehicle) can vary significantly when its pose relative to the radar line of sight is different. Current mainstream schemes (e.g., PWF and RS) ignore this relative geometric relationship. Although PolSAR target detection algorithms based on the rotation domain take these issues into account, they still rely on pixel-level feature differences and are therefore susceptible to speckle noise. Furthermore, as the resolution of PolSAR images increases, the structure and texture of the target region become finer. Pixel-based detection methods only consider the features of individual pixels and fail to fully utilize the information in the pixel region, resulting in a high false alarm rate for PolSAR ground target detection. Furthermore, most of the detection methods described above only consider polarization scattering characteristics and ignore their integration with intensity image-based detection algorithms.

[0005] In summary, the performance of current detection algorithms for PolSAR ground vehicle target detection is still not good enough. Summary of the Invention

[0006] To address the above-mentioned problems in the prior art, the present invention provides a CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] In one embodiment of the present invention, a CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation is provided. The PolSAR target detection method includes:

[0008] Obtaining a polarization scattering matrix constructed from four channels of a PolSAR image, and rotating the polarization scattering matrix by a preset angle along the radar line of sight to obtain a rotated polarization scattering matrix, so as to obtain four final polarization-correlated patterns based on the rotated polarization scattering matrix;

[0009] Obtaining polarization-related features based on the four final polarization-related patterns, and selecting two polarization-related features with the largest signal-to-clutter ratios by calculating the signal-to-clutter ratios of the target and the clutter background in the training samples;

[0010] A first binary detection result is obtained according to the one with the smaller signal-to-noise ratio among the two polarization-related features;

[0011] Performing superpixel segmentation on the PolSAR image using a superpixel segmentation algorithm to obtain a superpixel segmentation result, wherein the superpixel segmentation result includes a plurality of superpixel regions;

[0012] Using dual-parameter CFAR to detect the intensity image corresponding to the PolSAR image, and obtain a CFAR binary detection result;

[0013] Calculating the features of each superpixel region in the superpixel segmentation result using the CFAR binary detection result, so as to obtain a superpixel feature map based on the features of all the superpixel regions;

[0014] A final binary detection result is obtained based on the first binary detection result and the superpixel feature map.

[0015] In one embodiment of the present invention, a polarization scattering matrix constructed from four channels of a PolSAR image is obtained, and the polarization scattering matrix is ​​rotated at a preset angle along the radar line of sight to obtain a rotated polarization scattering matrix, so as to obtain four final polarization correlation patterns based on the rotated polarization scattering matrix, including:

[0016] Obtain raw PolSAR image data;

[0017] A polarization scattering matrix is ​​constructed based on the four channels of the original PolSAR image. The polarization scattering matrix is ​​expressed as:

[0018]

[0019] Where S is the polarization scattering matrix, S HH is the SAR image corresponding to the case where electromagnetic waves are sent and received with horizontal polarization, S HV is the SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization, S VH is the SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with horizontal polarization, S VV The SAR image corresponding to the case where electromagnetic waves are transmitted with vertical polarization and electromagnetic waves are received with vertical polarization;

[0020] The polarization scattering matrix is ​​rotated along the radar line of sight at a preset angle using a rotation matrix to obtain a rotated polarization scattering matrix. The rotated polarization scattering matrix is ​​expressed as:

[0021]

[0022]

[0023] Where S(θ) is the polarization scattering matrix after rotation, θ is the preset angle, T is the transpose, and R2(θ) is the rotation matrix;

[0024] The four components of the rotated polarization scattering matrix are expressed as:

[0025] S HH (θ)=S HH cos 2 θ+S HV cosθsinθ+S VH cosθsinθ+S VV sin 2 θ

[0026] S HV (θ)=-S HH cosθsinθ+S HV cos 2 θ-S VH sin 2 θ+S VV cosθsinθ

[0027] S VH (θ)=-S HH cosθsinθ-S HV sin 2 θ+S VH cos 2 θ+S VV cosθsinθ

[0028] S VV (θ)=S HH sin 2 θ-S HV cosθsinθ-S VH cosθsinθ+S VV cos 2 θ

[0029] Among them, S HH (θ) is the rotated SAR image corresponding to the horizontal polarization when electromagnetic waves are sent and received, S HV (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization, S VH (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with horizontal polarization, S VV (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are transmitted with vertical polarization and received with vertical polarization;

[0030] Based on the Pauli eigenvectors and the dictionary eigenvectors, four final polarization-related modes are obtained according to the rotated polarization scattering matrix.

[0031] In one embodiment of the present invention, based on the Pauli eigenvector and the dictionary eigenvector, four final polarization-related patterns are obtained according to the rotated polarization scattering matrix, including:

[0032] In the case of satisfying the reciprocity theorem S HV =S VH Under the condition of Pauli characteristic vector and dictionary feature vector Six polarization-related modes are obtained, wherein the six polarization-related modes include a first polarization-related mode |γ HH-VV (θ)|, the second polarization-related mode|γ HH-HV (θ)|, the third polarization-related mode|γ VV-HV (θ)|, the fourth polarization-related mode|γ (HH+VV)-(HH-VV) (θ)|, the fifth polarization-related mode|γ (HH+VV)-HV | and the sixth polarization-related mode |γ (HH-VV)-(HV) (θ)|;

[0033] Based on |γ (HH+VV)-(HH-VV) (θ)|=|γ (HH+VV)-HV (θ+π / 4)| and |γ HH-VV (θ)|=|γ VV-HV (θ+π / 2)|, the four final polarization-related modes are obtained, and the four final polarization-related modes include |γ HH-VV (θ)|、|γ HH-HV (θ)|、|γ (HH+VV)-(HH-VV) (θ)| and |γ (HH-VV)-(HV) (θ)|, wherein the four final polarization-related modes are expressed as:

[0034]

[0035]

[0036]

[0037]

[0038] Among them, S HH (θ) is the rotated SAR image corresponding to the horizontal polarization when electromagnetic waves are sent and received, S VV (θ) is the rotated SAR image corresponding to the vertical polarization when electromagnetic waves are sent and received, S HV(θ) is the rotated SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization, S (HH+VV) (θ) is S HH (θ) and S VV The sum of (θ), S (HH-VV) (θ) is S HH (θ) and S VV The difference of (θ), ·* represents conjugate, and <·> represents the sample geometric mean.

[0039] In one embodiment of the present invention, polarization correlation features are obtained based on the four final polarization correlation patterns, and two polarization correlation features with the largest signal-to-clutter ratios are selected by calculating the signal-to-clutter ratios of the target and the clutter background in the training sample, including:

[0040] For each polarization correlation mode, calculating the signal-to-clutter ratio of each polarization correlation feature corresponding to the target area and the clutter background area in the training sample;

[0041] The two polarization-related features with the largest signal-to-noise ratio are selected.

[0042] In one embodiment of the present invention, obtaining a first binary detection result according to the one of the two polarization-related features having a smaller signal-to-noise ratio includes:

[0043] The one with the smaller signal-to-noise ratio among the two polarization-related features with the largest signal-to-noise ratio is selected as the first polarization screening feature;

[0044] Compare the first feature map corresponding to the first polarization screening feature with the threshold T1, set the pixel points in the first feature map that are greater than or equal to the threshold T1 to 1, and set the pixel points in the first feature map that are less than the threshold T1 to 0, so as to obtain the first binary detection result.

[0045] In one embodiment of the present invention, superpixel segmentation is performed on the PolSAR image using a superpixel segmentation algorithm to obtain a superpixel segmentation result. The superpixel segmentation result includes several superpixel regions, including:

[0046] Step 4.1, dividing the two-dimensional image space of the PolSAR image into equally spaced grids according to a preset superpixel segmentation scale to obtain a divided PolSAR image, and using the position of each superpixel segmentation scale in the divided PolSAR image as the initial position of a cluster center;

[0047] Step 4.2: Taking the initial position of the cluster center as the center, obtain its 3*3 neighborhood as a preset area, calculate the edge strength of each position in the preset area, and use the position with the lowest edge strength in the preset area as the new cluster center position;

[0048] Step 4.3, calculating the similarity of the polarization coherence matrix between the pixel point at the new cluster center position and the pixels to be clustered at other positions by using the Wishart distance, and calculating the similarity of the spatial position between the pixel point at the new cluster center position and the pixels to be clustered at other positions by using the Euclidean distance;

[0049] Step 4.4: Calculate the total distance metric between the pixel at the new cluster center position and the pixels to be clustered at other positions based on the similarity of the polarization coherence matrix and the similarity of the spatial position. The total distance metric is expressed as:

[0050] d=d W +m×(d xy / K)

[0051] Among them, d is the total distance metric, d W is the similarity of the polarization coherence matrix, d xy is the similarity of spatial position, m is the weight coefficient, and K is the preset superpixel segmentation scale;

[0052] Step 4.5: Search for the new cluster center position within a 2K×2K area with the pixel to be clustered as the center, measure the total distance between the pixel to be clustered and all pixels at the new cluster center position within the 2K×2K area, classify the pixel to be clustered into the cluster center closest to the total distance measurement, and obtain the corresponding category label;

[0053] Step 4.6: After completing one iteration, calculate the average value of the position coordinates of all pixels in the superpixel area corresponding to each cluster, and use the position corresponding to the average value as the new cluster center position for the next iteration;

[0054] Step 4.7: Repeat steps 4.3 to 4.6 until clustering reaches a preset number of iterations to obtain the superpixel segmentation result.

[0055] In one embodiment of the present invention, the intensity image corresponding to the PolSAR image is detected using dual-parameter CFAR to obtain a CFAR binary detection result, including:

[0056] Step 5.1, based on the PolSAR image, calculate the intensity map using the polarization scattering matrix;

[0057] Step 5.2, selecting a number of clutter background areas in the intensity map;

[0058] Step 5.3: Obtain a corresponding histogram according to each piece of the clutter background area, and based on the histogram, select different distribution models to fit the distribution of the histogram to obtain fitting distribution results corresponding to different distribution models;

[0059] Step 5.4: obtaining a final fitting distribution result according to the fitting distribution result corresponding to the different distribution models that is closest to the change trend of the histogram;

[0060] Step 5.5: Obtain the detection statistic of the pixel to be detected based on the maximum likelihood estimation of the mean and standard deviation corresponding to the final fitting distribution result. When the final fitting distribution result is a lognormal distribution, the detection statistic of the pixel to be detected is expressed as:

[0061]

[0062] Among them, D i is the pixel x to be detected i The test statistic, is the maximum likelihood estimate of the mean corresponding to the final fitting distribution result, is the maximum likelihood estimate of the standard deviation corresponding to the final fitting distribution result;

[0063] Step 5.6: Compare the detection statistic of the pixel to be detected with a threshold value T2. If the detection statistic of the pixel to be detected is greater than or equal to the threshold value T2, the pixel to be detected is determined to be a target, and the pixel to be detected is set to 1. If the detection statistic of the pixel to be detected is less than the threshold value T2, the pixel to be detected is determined to be background, and the pixel to be detected is set to 0, thereby obtaining a CFAR binary detection result for the pixel to be detected.

[0064] Step 5.7: Repeat steps 5.5 to 5.6 to obtain the CFAR binary detection result corresponding to each pixel in the intensity map corresponding to the PolSAR image, thereby obtaining the CFAR binary detection result corresponding to the PolSAR image.

[0065] In one embodiment of the present invention, obtaining a final fitting distribution result according to a fitting distribution result that is closest to the change trend of the histogram among the fitting distribution results corresponding to the different distribution models includes:

[0066] For each of the clutter background regions, obtaining a fitting distribution result closest to the change trend of the histogram as the fitting distribution result to be used;

[0067] Determine whether the fitting distribution results to be used corresponding to all the clutter background areas are consistent. If they are consistent, the fitting distribution results to be used are used as the final fitting distribution results. If they are inconsistent, the fitting distribution results to be used corresponding to the main clutter background area are used as the final fitting distribution results. The main clutter background area is the clutter background area with the largest area in the experimental data scene.

[0068] In one embodiment of the present invention, the feature of each superpixel region in the superpixel segmentation result is calculated using the CFAR binary detection result to obtain a superpixel feature map based on the features of all the superpixel regions, including:

[0069] Step 6.1: Select the other polarization-related feature of the two polarization-related features with the largest signal-to-noise ratio as the second polarization screening feature;

[0070] Step 6.2: Determine the value of the pixel at the corresponding position in the CFAR binary detection result of the pixel in the superpixel area belonging to the same cluster in the superpixel segmentation result. If the value of the pixel at the corresponding position in the CFAR binary detection result is 1, obtain the characteristic value of the pixel at the corresponding position in the CFAR binary detection result as 1 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel in the superpixel area by the weight w1 to obtain a new characteristic value of the pixel. If the value of the pixel at the corresponding position in the CFAR binary detection result is 0, obtain the characteristic value of the pixel at the corresponding position in the CFAR binary detection result as 0 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel in the superpixel area by the weight w2 to obtain a new characteristic value of the pixel, wherein w1>1, w2<1;

[0071] Step 6.3: Repeat step 6.2 until a new feature value of each pixel in the superpixel area is obtained;

[0072] Step 6.4, adding the new eigenvalues ​​corresponding to all pixels in the superpixel area to obtain an addition result;

[0073] Step 6.5: Divide the summed result by the total number of pixels in the superpixel region to obtain the feature of the superpixel region;

[0074] Step 6.5, repeat steps 6.2-6.5. When the number of repetitions is equal to the total number of superpixel regions in the superpixel segmentation result, the features of all the superpixel regions are obtained, thereby obtaining the superpixel feature maps corresponding to the features of all the superpixel regions.

[0075] In one embodiment of the present invention, obtaining a final detection result according to the first binary detection result and the superpixel feature map includes:

[0076] Step 7.1: Find a pixel with a value of 1 in the first binary detection result, find the feature value of the corresponding position of the pixel in the superpixel feature map, and set the value of the pixel in the first binary detection result as the feature value of the corresponding position in the superpixel feature map;

[0077] Step 7.2, repeating step 7.1, setting the values ​​of the pixels whose values ​​are 1 in the first binary detection result to the feature values ​​of the corresponding positions in the superpixel feature map, to obtain a permuted feature map;

[0078] Step 7.3: Compare the characteristic value of each pixel in the replacement feature map with the threshold T3. If it is greater than or equal to the threshold T3, the value of the pixel in the replacement feature map is set to 1. If it is less than the threshold T3, the value of the pixel in the replacement feature map is set to 0 to obtain the final binary detection result.

[0079] Beneficial effects of the present invention:

[0080] The present invention addresses the shortcomings of existing PolSAR target detection algorithms and proposes a PolSAR target detection method guided by CFAR based on polarization characteristics and superpixel segmentation. The method selects a two-stage PolSAR target detection framework (the two stages include obtaining a first binary detection result and obtaining a final detection result based on a superpixel feature map). The two-stage PolSAR target detection framework is used for the detection task of ground vehicle targets in complex scenes. The present invention first performs preliminary detection and screening on the image by calculating relevant features of the polarization rotation domain, thereby obtaining a first binary detection result, completing the first stage of detection. Since there are many false alarms in the first stage, the original PolSAR image is then segmented into superpixel regions using a superpixel segmentation algorithm. The features of each superpixel region in the superpixel segmentation result are then calculated based on the dual-parameter CFAR detection results as the feature values ​​of the pixels in the superpixel region, and finally a superpixel feature map is obtained to complete the second stage of detection. Finally, based on the first stage detection results, a second screening is performed using the second stage results to remove most of the false alarms and obtain the final detection results. The method of the present invention deeply studies the combination of polarization scattering characteristics, intensity image clutter statistical characteristics and superpixel characteristics, which can significantly improve the detection performance of PolSAR image vehicle target detection tasks in complex ground scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1This is a flow chart of a CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation provided by an embodiment of the present invention;

[0082] Figure 2 This is a flow chart of another CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation provided by an embodiment of the present invention;

[0083] Figure 3 is an input PolSAR intensity image provided by an embodiment of the present invention;

[0084] Figure 4 is an extracted superpixel feature map provided by an embodiment of the present invention;

[0085] Figure 5 This is a final detection result diagram provided by an embodiment of the present invention;

[0086] Figure 6 This is a detection result diagram obtained according to the PWF comparison algorithm provided by an embodiment of the present invention;

[0087] Figure 7 is a detection result diagram obtained according to the comparison algorithm RS provided in an embodiment of the present invention;

[0088] Figure 8 This is a detection result diagram obtained according to the comparison algorithm PCP provided in an embodiment of the present invention; DETAILED DESCRIPTION

[0089] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0090] Example 1

[0091] Synthetic aperture radar (SAR) is a powerful active microwave imaging sensor. With the continuous deepening of SAR theoretical research and the continuous advancement of system development, SAR systems are rapidly developing towards multi-band, multi-view, and multi-polarization capabilities. In particular, polarimetric synthetic aperture radar (PolSAR), with its full polarimetric measurement capabilities, has attracted widespread attention worldwide due to its superior features and has become one of the mainstream research directions of scholars today.

[0092] Currently, the main research directions for PolSAR imagery include ship target detection and ground feature classification. However, research on ground target detection is relatively limited. Existing detection algorithms still have some problems. Since the backscatter of radar targets is very sensitive to the relative geometric relationship between the target pose and the radar line of sight, the scattering characteristics of the same target (e.g., a vehicle) can vary significantly when its pose relative to the radar line of sight is different. Current mainstream schemes (e.g., PWF and RS) ignore this relative geometric relationship. Although PolSAR target detection algorithms based on the rotation domain take these issues into account, they still rely on pixel-level feature differences and are therefore susceptible to speckle noise. Furthermore, as the resolution of PolSAR images increases, the structure and texture of the target region become finer. Pixel-based detection methods only consider the features of individual pixels and fail to fully utilize the information in the pixel region, resulting in a high false alarm rate for PolSAR ground target detection. Furthermore, most existing detection methods only consider polarization scattering characteristics and ignore their integration with intensity image-based detection algorithms.

[0093] Based on the above reasons, the present invention provides a CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation, see Figure 1 and Figure 2 The PolSAR target detection method provided by the present invention specifically includes:

[0094] Step 1: Obtain a polarization scattering matrix constructed from four channels of the PolSAR image, and rotate the polarization scattering matrix along the radar line of sight at a preset angle to obtain a rotated polarization scattering matrix, so as to obtain four final polarization-related patterns based on the rotated polarization scattering matrix.

[0095] Here, for the original PolSAR image, each independent polarization correlation pattern is first calculated based on the polarization scattering matrix extended to the polarization rotation domain, so that the corresponding polarization correlation feature is calculated for each polarization correlation pattern.

[0096] In a specific embodiment, step 1 may specifically include:

[0097] Step 1.1: Obtain raw PolSAR image data.

[0098] Here, the four channels of PolSAR image are S HH 、S HV 、S VH and S VV , S HH is the SAR image corresponding to the case where electromagnetic waves are sent and received with horizontal polarization, S HVis the SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization, S VH is the SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with horizontal polarization, S VV It is the SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with vertical polarization.

[0099] Step 1.2: Construct the polarization scattering matrix based on the four channels of the original PolSAR image. In the horizontal and vertical polarization bases (H, V), the polarization scattering matrix is ​​expressed as:

[0100]

[0101] Where S is the polarization scattering matrix;

[0102] Step 1.3: Use the rotation matrix to rotate the polarization scattering matrix at a preset angle along the radar line of sight to obtain the rotated polarization scattering matrix. The rotated polarization scattering matrix is ​​expressed as:

[0103]

[0104]

[0105] Where S(θ) is the rotated polarization scattering matrix, θ is the preset angle, θ∈[-π,π), T is the transpose, and R2(θ) is the rotation matrix;

[0106] The four components of the rotated polarization scattering matrix are expressed as:

[0107] S HH (θ)=S HH cos 2 θ+S HV cosθsinθ+S VH cosθsinθ+S VV sin 2 θ

[0108] S HV (θ)=-S HH cosθsinθ+S HV cos 2 θ-S VH sin 2 θ+S VV cosθsinθ

[0109] S VH (θ)=-S HH cosθsinθ-S HV sin 2 θ+S VH cos2 θ+S VV cosθsinθ

[0110] S VV (θ)=S HH sin 2 θ-S HV cosθsinθ-S VH cosθsinθ+S VV cos 2 θ

[0111] Among them, S HH (θ) is the rotated SAR image corresponding to the horizontal polarization when electromagnetic waves are sent and received, S HV (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization, S VH (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with horizontal polarization, S VV (θ) is the rotated SAR image corresponding to the case where electromagnetic waves are transmitted with vertical polarization and received with vertical polarization;

[0112] Step 1.4: Based on the Pauli eigenvectors and the dictionary eigenvectors, four final polarization-related patterns are obtained according to the rotated polarization scattering matrix.

[0113] In this embodiment, step 1.4 may specifically include:

[0114] Step 1.41: When the reciprocity theorem S is satisfied HV =S VH Under the condition of Pauli characteristic vector and dictionary feature vector Six polarization-related modes are obtained, including the first polarization-related mode |γ HH-VV (θ)|, the second polarization-related mode|γ HH-HV (θ)|, the third polarization-related mode|γ VV-HV (θ)|, the fourth polarization-related mode|γ (HH+VV)-(HH-VV) (θ)|, the fifth polarization-related mode|γ (HH+VV)-HV | and the sixth polarization-related mode |γ (HH-VV)-(HV) (θ)|.

[0115] Specifically, for two polarization channels S1 and S2, the polarization correlation is defined as follows:

[0116]

[0117] in, Indicates the conjugate of S2, and <·> represents the sample geometric mean. In PolSAR images, the correlation between the two polarization channels contains rich information. By extending the original fixed-angle polarization correlation between the two channels to the rotation domain along the radar line of sight, the polarization correlation pattern obtained is:

[0118]

[0119] in, represents the conjugate of S2(θ).

[0120] Therefore, under the premise of satisfying the reciprocity theorem, that is, S HV =S VH , based on the Pauli characteristic vector and dictionary feature vector Six typical polarization correlation modes can be obtained, namely |γ HH-VV |、|γ HH-HV |、|γ VV-HV |、|γ (HH+VV)-(HH-VV) |、|γ (HH+VV)-HV | and |γ (HH-VV)-HV |.

[0121] Step 1.42, based on |γ (HH+VV)-(HH-VV) (θ)|=|γ (HH+VV)-HV (θ+π / 4)| and |γ HH-VV (θ)|=|γ VV-HV (θ+π / 2)|, and obtain four final polarization-related modes. The four final polarization-related modes include |γ HH-VV (θ)|、|γ HH-HV (θ)|、|γ (HH+VV)-(HH-VV) (θ)| and |γ (HH-VV)-(HV) (θ)|.

[0122] Here, due to:

[0123]

[0124]

[0125] The following conclusions can be drawn:

[0126] |γ (HH+VV)-(HH-VV) (θ)|=|γ (HH+VV)-HV (θ+π / 4)|

[0127] |γ HH-VV (θ)|=|γ VV-HV (θ+π / 2)|

[0128] Therefore, this means that |γ (HH+VV)-(HH-VV) | and |γ (HH+VV)-HV |、|γHH-VV | and |γ VV-HV |These two pairs of polarization correlations are equivalent. Therefore, we finally get four independent polarization correlation patterns |γ HH-VV (θ)|、|γ HH-HV (θ)|、|γ (HH+VV)-(HH-VV) (θ)| and |γ (HH-VV)-(HV) (θ)|, the specific calculation formula is as follows:

[0129]

[0130]

[0131]

[0132]

[0133] Among them, S (HH+VV) (θ) is S HH (θ) and S VV The sum of (θ), S (HH-VV) (θ) is S HH (θ) and S VV The difference of (θ).

[0134] Therefore, four polarization correlation modes are calculated according to the above formula, where θ takes values ​​from -π to π at intervals of 0.05, and the polarization correlation modes corresponding to different angles are calculated cyclically.

[0135] Step 2: Obtain polarization-correlation features based on the four final polarization-correlation patterns. By calculating the signal-to-clutter ratio of the target and the clutter background in the training sample, select the two polarization-correlation features with the largest signal-to-clutter ratio. The polarization-correlation features are used to characterize the information hidden in the polarization-correlation pattern of a given polarization channel.

[0136] Specifically, the exploration of hidden features in the rotation domain provides a lot of valuable information for understanding and studying PolSAR images. In order to more accurately quantitatively characterize the polarization-related patterns, ten polarization-related features can be calculated for each polarization-related pattern. These features fully describe the hidden information in the polarization-related pattern of a given polarization channel. Therefore, features with a large gap between the target and background areas can be selected for the first and second stage detection tasks.

[0137] In a specific embodiment, step 2 may specifically include:

[0138] Step 2.1: For each polarization correlation pattern, calculate the signal-to-clutter ratio of each corresponding polarization correlation feature in the target area and background area of ​​the training sample.

[0139] In this embodiment, in order to more accurately and quantitatively characterize the polarization-related pattern, a set of polarization-related features is used. By calculating these polarization-related features, the hidden information in the rotation domain is further mined. These polarization-related features are defined as follows:

[0140] 1) Original correlation feature γ -org : This feature is the polarization-related feature commonly used without any rotation processing, representing the target decorrelation effect of the two polarization channels under the original imaging geometry.

[0141] 2) Correlation mean γ -mean : It is the average correlation value in the rotation domain and indicates the general correlation level. The average correlation value is a measure of the average decorrelation effect of the target in the rotation domain. The higher the average correlation value, the lower the decorrelation phenomenon.

[0142] 3) Correlation standard deviation γ -std : This is the standard deviation of the correlation in the rotational domain. It can identify the diversity of the target's scattering directions. Generally, the larger the correlation standard deviation, the more obvious the diversity of the target's scattering directions in the rotational domain. For rotationally invariant scatterers, the coherence fluctuation will approach zero.

[0143] 4) Maximum correlation γ -max : Maximum value of correlation in the rotation domain, upper limit of polarization correlation between two polarization channels at different angles.

[0144] 5) Minimum correlation γ -min : Minimum correlation in the rotation domain, the lower limit of the polarization correlation between two polarization channels at different angles.

[0145] 6) Correlation contrast γ -contrast =γ -max -γ -min : It reflects the absolute contrast in the rotation domain. In addition, the correlation contrast can also reflect the diversity of the target scattering direction in the rotation domain. For rotation-invariant targets with no direction dependence, the correlation contrast will become 0.

[0146] 7) Correlated anisotropy γ -A =(γ -max -γ -min ) / (γ -max +γ -min ): reflects the relative contrast in the rotation domain. Correlation anisotropy is a complementary feature to correlation contrast. For rotated targets with relatively low correlation, correlation anisotropy can further enhance the contrast and achieve better discrimination performance.

[0147] 8) Correlation beam width γ -bw0.95 : That is, the correlation value is not less than 0.95×γ -maxThe smaller the value, the greater the decorrelation effect and the stronger the direction dependence.

[0148] 9) Maximum rotation angle θ γ-max : Defined as the rotation angle that produces the maximum correlation within the main range.

[0149] 10) Minimum rotation angle θ γ-min : Defined as the rotation angle that produces the minimum correlation in the main range.

[0150] According to the definition of polarization-related features, by calculating the ten polarization-related features corresponding to each polarization-related pattern, a total of 40 polarization-related features can be obtained. These polarization-related features have a certain degree of redundancy.

[0151] Step 2.2: Select the two polarization-related features with the largest signal-to-noise ratio.

[0152] Specifically, for each of the four polarization-related modes, the ten polarization-related features described above can be obtained. Therefore, it is necessary to select features with large differences between the target and the background. Different selections will be made for different data sets. In order to obtain more feature information, different polarization-related features are generally selected in the first and second stage feature selection.

[0153] Therefore, in order to select features with large differences between the target and the background, this embodiment calculates the signal-to-clutter ratio of each polarization-related feature corresponding to the target and the clutter background in the training sample, and thus selects the two polarization-related features with the largest signal-to-clutter ratio for the first and second stage processing.

[0154] Step 3: Obtain a first binary detection result based on the one with the smaller signal-to-noise ratio among the two polarization correlation features with the largest signal-to-noise ratio, thereby completing the first stage of detection.

[0155] In a specific embodiment, step 3 may include:

[0156] Step 3.1: Select the polarization-related feature with the smallest signal-to-noise ratio among the two polarization-related features with the largest signal-to-noise ratio as the first polarization screening feature.

[0157] Here, the first polarization screening feature is the one with the smaller signal-to-noise ratio among the two polarization-related features with the largest signal-to-noise ratio.

[0158] Step 3.2: Compare the first feature map corresponding to the first polarization screening feature with threshold T1, set pixels in the first feature map greater than or equal to threshold T1 to 1, and set pixels in the first feature map less than threshold T1 to 0, to obtain a first binary detection result. Here, the first polarization screening feature is the first feature map.

[0159] It should be noted that the threshold T1 can be set according to actual conditions, and this embodiment does not limit this. For example, the threshold T1 is 30.

[0160] Step 4: Use a superpixel segmentation algorithm to perform superpixel segmentation on the PolSAR image to obtain a superpixel segmentation result, wherein the superpixel segmentation result includes a plurality of superpixel regions.

[0161] In this embodiment, the original SLIC (simple linear iterative clustering) algorithm is a superpixel segmentation algorithm developed based on optical images. To make it applicable to PolSAR image data, this embodiment uses an improved SLIC algorithm from the literature for experiments. The main differences compared to the original algorithm are: the original color vector is replaced by the elements of the polarization coherence matrix; the original function for measuring the distance of the color vector is replaced by the Wishart distance; and the edge strength calculation is changed to using the likelihood ratio to test whether the two average coherence matrices in the first and last rows of the preset area are equal.

[0162] In a specific embodiment, step 4 may specifically include:

[0163] Step 4.1: Divide the two-dimensional image space of the PolSAR image into equally spaced grids according to the preset superpixel segmentation scale to obtain the divided PolSAR image, and use the position of every preset superpixel segmentation scale in the divided PolSAR image as the initial position of a cluster center.

[0164] Specifically, a preset superpixel segmentation scale is first set. This parameter is set based on the size of the target. For example, in this embodiment, it is set to 40. As a result, the two-dimensional image space of the PolSAR image can be divided into equally spaced grids according to the preset superpixel segmentation scale to obtain a divided PolSAR image. Then, in order to obtain the initial position of the cluster center, in this embodiment, the position of each preset superpixel segmentation scale in the divided PolSAR image is used as the initial position of a cluster center, that is, the divided position is used as the initial position of the cluster center.

[0165] Step 4.2: Obtain a preset area with the initial position of the cluster center as the center, calculate the edge strength of each position in the preset area, and use the position with the lowest edge strength in the preset area as the new cluster center position.

[0166] Specifically, this embodiment divides the preset area into the initial position of the cluster center as the center. For example, the preset area is the area corresponding to the 3*3 neighborhood near the center. Then, the edge strength of each position in the preset area is calculated using the calculation formula of horizontal edge strength and vertical edge strength, and the position with the lowest edge strength is used as the new cluster center position. The polarization coherence matrix and coordinates constitute the eigenvector of the cluster center.

[0167] Here, the calculation of the horizontal edge strength uses the likelihood ratio to test whether the two average coherence matrices of the first and last rows in the 3*3 neighborhood near each position pixel in the preset area are equal, then the horizontal edge strength E h The calculation formula is:

[0168]

[0169] Among them, x0 and y0 are the initial horizontal and vertical position coordinates of the cluster center, is the polarization coherence matrix corresponding to the pixel point (x, y0-1).

[0170] Vertical edge strength E v The calculation method of is similar, as shown below.

[0171]

[0172] The final edge strength of the center pixel E=max([E h , E v ]).

[0173] Step 4.3: Calculate the similarity of the polarization coherence matrix between the pixel point at the new cluster center position and the pixel points to be clustered at other positions using the Wishart distance, and calculate the similarity of the spatial position between the pixel point at the new cluster center position and the pixel points to be clustered at other positions using the Euclidean distance.

[0174] Here, the similarity d of the polarization coherence matrices of two pixels is W Calculated by Wishart distance, the specific calculation formula is as follows:

[0175] d W =ln(|V|)+tr(V -1 T)

[0176] Where T is the coherence matrix of a pixel, V is the coherence matrix of a cluster center, and tr(·) represents the trace of a matrix.

[0177] The similarity d between the spatial positions of two pixels xy Calculated by Euclidean distance, the calculation formula is as follows:

[0178]

[0179] Among them, (x1, y1) and (x2, y2) are the coordinates of two pixel points respectively.

[0180] Step 4.4: Calculate the total distance metric between the pixel at the new cluster center and the pixels to be clustered at other locations based on the similarity of the polarization coherence matrix and the similarity of the spatial position. The total distance metric is expressed as:

[0181] d=d W +m×(d xy / K)

[0182] Among them, d is the total distance metric, d W is the similarity of the polarization coherence matrix, d xy is the similarity of spatial position, m is the weight coefficient, which is used to control the compactness of superpixels. The larger m is, the more compact superpixels will be produced. K is the preset superpixel segmentation scale.

[0183] Step 4.5: With the pixel to be clustered as the center, search for a new cluster center position in the 2K×2K area, measure the total distance metric between the pixel to be clustered and all pixels at the new cluster center positions in the 2K×2K area, divide the pixel to be clustered into the cluster center closest to the total distance metric (that is, divide the pixel to be clustered into the cluster center with the smallest value in the total distance metric d), and obtain the corresponding category label (that is, label the category of each cluster).

[0184] Step 4.6: After completing one iteration, calculate the average value of the position coordinates of all pixels in the superpixel area corresponding to each cluster, and use the position corresponding to the average value as the new cluster center position for the next iteration, where each superpixel area is the area corresponding to a cluster.

[0185] Step 4.7: Repeat steps 4.3 to 4.6 until the clustering reaches a preset number of iterations to obtain the superpixel segmentation result corresponding to the original PolSAR image.

[0186] It should be noted that the number of iterations can be set according to actual needs and is not specifically limited in this embodiment.

[0187] Step 5: Use the dual-parameter CFAR (Constant False-Alarm Rate, CFAR) to detect the intensity image corresponding to the PolSAR image to obtain a CFAR binary image.

[0188] Specifically, the intensity image is calculated using full polarization data. The statistical characteristics of the clutter region in the intensity image are analyzed, and the distribution of the clutter region in the image is fitted to obtain the distribution of the clutter data. The calculation method of the mean and variance in the two-parameter CFAR is modified according to this distribution. The intensity image is detected using the modified two-parameter CFAR detector to obtain the detection result.

[0189] In a specific embodiment, step 5 may specifically include:

[0190] Step 5.1: Based on the PolSAR image, calculate the intensity map using the polarization scattering matrix.

[0191] Here, the intensity map is:

[0192] span=|S HH | 2 +|S HV | 2 +|S VH | 2 +|S VV | 2

[0193] Among them, span represents the intensity map.

[0194] Step 5.2: Select several clutter background areas in the intensity map.

[0195] For example, select two cluttered background areas.

[0196] Step 5.3: Obtain a corresponding histogram according to each clutter background area, and based on the histogram, select different distribution models to fit the distribution of the histogram to obtain fitting distribution results corresponding to different distribution models.

[0197] Specifically, several clutter background areas are selected in the intensity map, a histogram is drawn based on the clutter background data, and different distribution models are selected based on the histogram to fit the distribution of the clutter background. There are five common clutter distribution models: Gaussian distribution, logarithmic Gaussian distribution, Weibull distribution, Rayleigh distribution, and Gamma distribution.

[0198] Step 5.4: Obtain a final fitting distribution result based on the fitting distribution result that is closest to the change trend of the histogram among the fitting distribution results corresponding to different distribution models.

[0199] Step 5.41: For each clutter background area, obtain the fitting distribution result that is closest to the change trend of the histogram as the fitting distribution result to be used.

[0200] Specifically, when each distribution model is used to fit the selected clutter background area, a fitting distribution result will be obtained. From all the fitting distribution results corresponding to all distribution models, the fitting distribution result that is closest to the change trend of the histogram is selected, and the fitting distribution result is used as the fitting distribution result to be used.

[0201] Step 5.42: Determine whether the to-be-used fitting distribution results corresponding to all clutter background areas are consistent. If they are consistent, the to-be-used fitting distribution results are used as the final fitting distribution results. If they are inconsistent, the to-be-used fitting distribution results corresponding to the main clutter background areas are used as the final fitting distribution results.

[0202] In this embodiment, it is found experimentally that the clutter distribution of the data used satisfies the lognormal distribution, so the lognormal distribution is selected as the final fitting distribution result. The probability density function of the lognormal distribution is:

[0203]

[0204] Where x i is the intensity value of the i-th pixel in the clutter background area, μ and σ are the mean and standard deviation respectively, and N is the total number of pixels in the clutter background area.

[0205] When the distribution models used by the to-be-used fitting distribution results corresponding to all clutter background areas are inconsistent, the to-be-used fitting distribution result corresponding to the main clutter background area is selected as the final fitting distribution result. The main clutter background area is the clutter background area with the largest area in the experimental data scene. For example, if the forest area accounts for the largest proportion in the background, the main clutter background area is the forest.

[0206] Step 5.5: Based on the maximum likelihood estimation of the mean and standard deviation corresponding to the final fitting distribution result, the detection statistic of the pixel to be detected is obtained. When the final fitting distribution result is a lognormal distribution, the detection statistic of the pixel to be detected is expressed as:

[0207]

[0208] Among them, D i is the pixel x to be detected i The test statistic, is the maximum likelihood estimate of the mean corresponding to the final fitting distribution result, is the maximum likelihood estimate of the standard deviation corresponding to the final fitting distribution result, and is the pixel x i The pixels in the surrounding clutter area are calculated as follows:

[0209]

[0210] Step 5.6: Compare the detection statistic of the pixel to be detected with the threshold T2. If the detection statistic of the pixel to be detected is greater than or equal to the threshold T2, the pixel to be detected is judged as the target, and the pixel to be detected is set to 1. If the detection statistic of the pixel to be detected is less than the threshold T2, the pixel to be detected is judged as the background, and the pixel to be detected is set to 0 to obtain the CFAR binary detection result of the pixel to be detected.

[0211] It should be noted that the threshold T2 can be set according to actual conditions, and this embodiment does not limit this. For example, the threshold T2 is 3.09.

[0212] Step 5.7: Repeat steps 5.5 to 5.6 to obtain the CFAR binary detection result corresponding to each pixel in the intensity map corresponding to the PolSAR image, thereby obtaining the CFAR binary image corresponding to the PolSAR image.

[0213] Step 6: Use the CFAR binary image to calculate the features of each superpixel region in the superpixel segmentation result, so as to obtain a superpixel feature map based on the features of all superpixel regions, thereby completing the second stage of detection.

[0214] Specifically, the recall rate of the dual-parameter CFAR detection results is high, indicating that all targets can be basically detected, while most false alarms are strong scattering points. The features of the superpixel area are calculated based on the CFAR detection results. Specifically, the eigenvalues ​​of the pixels detected by CFAR in the superpixel area have higher weights, while the pixels not detected by CFAR are given lower weights. The weighted eigenvalues ​​in the superpixel block are summed and averaged to obtain the features corresponding to the superpixel block.

[0215] In a specific embodiment, step 6 may specifically include:

[0216] Step 6.1, selecting the other polarization-related feature with the largest signal-to-noise ratio from the two polarization-related features as the second polarization screening feature;

[0217] Step 6.2: Determine the value of the pixel point at the corresponding position in the CFAR binary detection result of the pixel point in the superpixel area belonging to the same cluster in the superpixel segmentation result. If the value of the pixel point at the corresponding position in the CFAR binary detection result is 1, obtain the characteristic value of the pixel point at the corresponding position in the CFAR binary detection result as 1 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel point in the superpixel area by the weight w1 to obtain a new characteristic value of the pixel point. If the value of the pixel point at the corresponding position in the CFAR binary detection result is 0, obtain the characteristic value of the pixel point at the corresponding position in the CFAR binary detection result as 0 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel point in the superpixel area by the weight w2 to obtain a new characteristic value of the pixel point, wherein w1>1, w2<1.

[0218] Here, the second polarization screening feature is the second feature map, and each pixel point in the superpixel area corresponds to a pixel point in the second feature map. Therefore, the feature value of each pixel point in the superpixel area can be determined by the feature value of the corresponding pixel point in the second feature map.

[0219] Step 6.3: Repeat step 6.2 until a new feature value is obtained for each pixel in the superpixel area.

[0220] Step 6.4, add the new eigenvalues ​​corresponding to all pixels in the superpixel area to obtain the addition result;

[0221] Step 6.5: Divide the sum by the total number of pixels in the superpixel region to obtain the features of the superpixel region.

[0222] Step 6.5: Repeat steps 6.2 to 6.5. When the number of repetitions is equal to the total number of superpixel regions in the superpixel segmentation result, the features of all superpixel regions are obtained, thereby obtaining the superpixel feature maps corresponding to the features of all superpixel regions.

[0223] Step 7: Obtain a final binary detection result based on the first binary detection result and the superpixel feature map.

[0224] Specifically, the detection result obtained in the first stage is used as an index, and the features of the superpixel area obtained in the second stage are used to obtain a new feature map, and the threshold of the feature map is taken to obtain the final detection result.

[0225] In a specific embodiment, step 7 may specifically include:

[0226] Step 7.1. Use the detection result of the first stage as an index, find a pixel point with a value of 1 in the first binary detection result, and find the eigenvalue of the corresponding position of the pixel point in the superpixel feature map, and set the value of the pixel point in the first binary detection result as the eigenvalue of the corresponding position in the superpixel feature map.

[0227] Step 7.2: Repeat step 7.1 so that the values ​​of the pixels with a value of 1 in the first binary detection result are all set as the feature values ​​of the corresponding positions in the superpixel feature map to obtain a replacement feature map.

[0228] Step 7.3: Compare the feature value of each pixel in the replacement feature map with the threshold T3. If it is greater than or equal to the threshold T3, the value of the pixel in the replacement feature map is set to 1. If it is less than the threshold T3, the value of the pixel in the replacement feature map is set to 0 to obtain the final binary detection result.

[0229] It should be noted that the threshold T3 can be set according to actual conditions, and this embodiment does not limit this. For example, the threshold T3 is 289.

[0230] The present invention selects a set of public data sets. The experimental data is obtained by a synthetic aperture radar onboard an unmanned aerial vehicle. The system operates in the X-band with a maximum resolution of 0.15m and adopts full polarization imaging.

[0231] Table 1 below shows the detection results of the method of the present invention on the above dataset, and compares them with the existing PolSAR image target detection algorithms: polarization whitening filter detector PWF (Polarimetric Whitening Filter Detector, from the paper "Optimal polarimetric processing for enhanced target detection [J]. IEEE Transactions on Aerospace and Electronic Systems, 1993, 29 (1): 234-244"), target detector based on reflection symmetry RS (Reflection Symmetry, from the paper "Reflection symmetry for polarimetric observation of man-made metallic targets at sea [J]. IEEE Journal of Oceanic Engineering, 2012, 37 (3): 384-394") and target detection algorithm based on polarization correlation pattern PCP (Polarimetric Correlation Pattern, from the paper "PolSAR ship detection based on polarimetric correlation pattern [J]. IEEE Geoscience and Remote Sensing Letters, 2020, 18 (3): 471-475"). The target level detection rate P in Table 1 is 0.01, which is 0.01. d It represents the ratio of the number of correctly detected targets to the total number of real targets, and the pixel-level false alarm rate P f Indicates the ratio of the number of detected error pixels to the total number of background pixels; Figure 3 is the input raw PolSAR intensity image data, Figure 4 is the extracted superpixel feature map, Figure 5 The final test result diagram is shown in Figure 2. Figure 6 、 Figure 7 、 Figure 8 The results of the comparison algorithm are shown in FIG. From the results, it can be seen that the method proposed by the present invention has fewer false alarms and better detection performance.

[0232] Table 1 Comparison of detection performance between the method of the present invention and the existing method

[0233]

[0234]

[0235] To address the problem that targets have different scattering characteristics depending on their posture relative to the radar line of sight, the present invention expands the target polarization scattering matrix obtained under specific geometric relationships to the rotation domain around the radar line of sight in the first stage. Based on the analytical expression of the polarization scattering matrix in the rotation domain, the characteristics of the polarization-related pattern are adopted to fully exploit the hidden features in the polarization rotation domain.

[0236] The present invention addresses the impact of inherent speckle noise in SAR images and studies the spatial relationship and structural similarity between pixels and their surrounding neighborhoods in PolSAR images. In the second stage, regional-level features are obtained by combining the results of superpixel segmentation, which reduces a large number of noise false alarms compared to pixel-level detection results.

[0237] The present invention analyzes the statistical characteristics of the background clutter area in the PolSAR intensity image, uses a histogram to fit the clutter distribution, modifies the dual-parameter CFAR detector to obtain more accurate CFAR detection results, and uses the CFAR detection results to weight the superpixel features, further enhancing the feature differences between the target and the background, thereby significantly improving the target detection performance.

[0238] Aiming at the problem of ground vehicle target detection in PolSAR images, this paper proposes a two-stage detection framework based on CFAR guidance, which effectively improves the indicators of target detection tasks.

[0239] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, those skilled in the art may combine and integrate the different embodiments or examples described in this specification. Although the present application is described herein with reference to various embodiments, those skilled in the art may understand and implement other variations of the disclosed embodiments by examining the drawings, the disclosure, and the appended claims in the course of implementing the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit may perform several of the functions recited in a claim. The fact that certain measures are recited in different dependent claims does not mean that these measures cannot be combined to produce advantageous results.

[0240] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A CFAR-guided PolSAR target detection method based on polarization characteristics and superpixel segmentation, characterized in that: The PolSAR target detection method comprises: Obtaining a polarization scattering matrix constructed from four channels of a PolSAR image, and rotating the polarization scattering matrix by a preset angle along the radar line of sight to obtain a rotated polarization scattering matrix, so as to obtain four final polarization-correlated patterns based on the rotated polarization scattering matrix; Obtaining polarization-related features based on the four final polarization-related patterns, and selecting two polarization-related features with the largest signal-to-clutter ratios by calculating the signal-to-clutter ratios of the target and the clutter background in the training samples; A first binary detection result is obtained according to the one with the smaller signal-to-noise ratio among the two polarization-related features; Performing superpixel segmentation on the PolSAR image using a superpixel segmentation algorithm to obtain a superpixel segmentation result, wherein the superpixel segmentation result includes a plurality of superpixel regions; Using dual-parameter CFAR to detect the intensity image corresponding to the PolSAR image, and obtain a CFAR binary detection result; Calculating the features of each superpixel region in the superpixel segmentation result using the CFAR binary detection result, so as to obtain a superpixel feature map based on the features of all the superpixel regions; Obtaining a final binary detection result according to the first binary detection result and the superpixel feature map; Calculating the features of each superpixel region in the superpixel segmentation result using the CFAR binary detection result to obtain a superpixel feature map based on the features of all the superpixel regions, including: Step 6.1: Select the other polarization-related feature of the two polarization-related features with the largest signal-to-noise ratio as the second polarization screening feature; Step 6.2: Determine the value of the pixel at the corresponding position in the CFAR binary detection result of the pixel in the superpixel area belonging to the same cluster in the superpixel segmentation result. If the value of the pixel at the corresponding position in the CFAR binary detection result is 1, obtain the characteristic value of the pixel at the corresponding position in the CFAR binary detection result as 1 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel in the superpixel area by the weight w1 to obtain a new characteristic value of the pixel. If the value of the pixel at the corresponding position in the CFAR binary detection result is 0, obtain the characteristic value of the pixel at the corresponding position in the CFAR binary detection result as 0 according to the second feature map corresponding to the second polarization screening feature, and multiply the characteristic value of the pixel in the superpixel area by the weight w2 to obtain a new characteristic value of the pixel, wherein w1>1, w2<1; Step 6.3: Repeat step 6.2 until a new feature value of each pixel in the superpixel area is obtained; Step 6.4, adding the new eigenvalues ​​corresponding to all pixels in the superpixel area to obtain an addition result; Step 6.5: Divide the summed result by the total number of pixels in the superpixel region to obtain the feature of the superpixel region; Step 6.5: Repeat steps 6.2 to 6.

5. When the number of repetitions is equal to the total number of superpixel regions in the superpixel segmentation result, the features of all the superpixel regions are obtained, thereby obtaining the superpixel feature maps corresponding to the features of all the superpixel regions. Obtaining a final detection result according to the first binary detection result and the superpixel feature map, including: Step 7.1: Find a pixel with a value of 1 in the first binary detection result, find the feature value of the corresponding position of the pixel in the superpixel feature map, and set the value of the pixel in the first binary detection result as the feature value of the corresponding position in the superpixel feature map; Step 7.2, repeating step 7.1, setting the values ​​of the pixels whose values ​​are 1 in the first binary detection result to the feature values ​​of the corresponding positions in the superpixel feature map, to obtain a permuted feature map; Step 7.3: Compare the characteristic value of each pixel in the replacement feature map with the threshold T3. If it is greater than or equal to the threshold T3, the value of the pixel in the replacement feature map is set to 1. If it is less than the threshold T3, the value of the pixel in the replacement feature map is set to 0 to obtain the final binary detection result.

2. The PolSAR target detection method according to claim 1, wherein: Obtain a polarization scattering matrix constructed from four channels of a PolSAR image, and rotate the polarization scattering matrix along the radar line of sight by a preset angle to obtain a rotated polarization scattering matrix, so as to obtain four final polarization-related patterns based on the rotated polarization scattering matrix, including: Obtain raw PolSAR image data; A polarization scattering matrix is ​​constructed based on the four channels of the original PolSAR image. The polarization scattering matrix is ​​expressed as: in, is the polarization scattering matrix, The SAR images corresponding to the horizontal polarization when electromagnetic waves are sent and received are: The SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with vertical polarization is shown in the figure. The SAR image corresponding to the case where electromagnetic waves are sent with vertical polarization and received with horizontal polarization is shown. The SAR image corresponding to the case where electromagnetic waves are transmitted with vertical polarization and electromagnetic waves are received with vertical polarization; The polarization scattering matrix is ​​rotated along the radar line of sight at a preset angle using a rotation matrix to obtain a rotated polarization scattering matrix. The rotated polarization scattering matrix is ​​expressed as: in, is the rotated polarization scattering matrix, is the preset angle, is the transpose, is the rotation matrix; The four components of the rotated polarization scattering matrix are expressed as: in, is the rotated SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with horizontal polarization, The SAR image is rotated when electromagnetic waves are sent with horizontal polarization and received with vertical polarization. The SAR image is rotated when electromagnetic waves are sent with vertical polarization and received with horizontal polarization. The rotated SAR image corresponding to the case where electromagnetic waves are transmitted with vertical polarization and received with vertical polarization; Based on the Pauli eigenvectors and the dictionary eigenvectors, four final polarization-related modes are obtained according to the rotated polarization scattering matrix.

3. The PolSAR target detection method according to claim 2, wherein: Based on the Pauli eigenvector and the dictionary eigenvector, four final polarization-related modes are obtained according to the rotated polarization scattering matrix, including: In the case of satisfying the reciprocity theorem Under the condition of Pauli characteristic vector and dictionary feature vector Six polarization-related modes are obtained, wherein the six polarization-related modes include a first polarization-related mode , the second polarization correlation mode , the third polarization correlation mode , the fourth polarization correlation mode , the fifth polarization-related mode and the sixth polarization-related mode ; based on and , the four final polarization-related modes are obtained, and the four final polarization-related modes include 、 、 and , wherein the four final polarization-related modes are respectively expressed as: in, is the rotated SAR image corresponding to the case where electromagnetic waves are sent with horizontal polarization and received with horizontal polarization, The corresponding rotated SAR image when electromagnetic waves are sent with vertical polarization and received with vertical polarization is: The SAR image is rotated when electromagnetic waves are sent with horizontal polarization and received with vertical polarization. for and of and, for and The difference, represents conjugation, represents the sample geometric mean.

4. The PolSAR target detection method according to claim 1, wherein: Polarization-related features are obtained based on the four final polarization-related patterns. By calculating the signal-to-clutter ratio of the target and the clutter background in the training sample, two polarization-related features with the largest signal-to-clutter ratio are selected, including: For each polarization correlation mode, calculating the signal-to-clutter ratio of each polarization correlation feature corresponding to the target area and the clutter background area in the training sample; The two polarization-related features with the largest signal-to-noise ratio are selected.

5. The PolSAR target detection method according to claim 1, wherein: Obtaining a first binary detection result according to the one of the two polarization-related features having a smaller signal-to-noise ratio, including: The one with the smaller signal-to-noise ratio among the two polarization-related features with the largest signal-to-noise ratio is selected as the first polarization screening feature; Compare the first feature map corresponding to the first polarization screening feature with the threshold T1, set the pixel points in the first feature map that are greater than or equal to the threshold T1 to 1, and set the pixel points in the first feature map that are less than the threshold T1 to 0, so as to obtain the first binary detection result.

6. The PolSAR target detection method according to claim 1, wherein: The PolSAR image is subjected to superpixel segmentation using a superpixel segmentation algorithm to obtain a superpixel segmentation result. The superpixel segmentation result includes several superpixel regions, including: Step 4.1, dividing the two-dimensional image space of the PolSAR image into equally spaced grids according to a preset superpixel segmentation scale to obtain a divided PolSAR image, and using the position of each superpixel segmentation scale in the divided PolSAR image as the initial position of a cluster center; Step 4.2: Take the initial position of the cluster center as the center to obtain its The neighborhood is used as a preset area, the edge strength of each position in the preset area is calculated, and the position with the lowest edge strength in the preset area is used as the new cluster center position; Step 4.3, calculating the similarity of the polarization coherence matrix between the pixel point at the new cluster center position and the pixels to be clustered at other positions by using the Wishart distance, and calculating the similarity of the spatial position between the pixel point at the new cluster center position and the pixels to be clustered at other positions by using the Euclidean distance; Step 4.4: Calculate the total distance metric between the pixel at the new cluster center position and the pixels to be clustered at other positions based on the similarity of the polarization coherence matrix and the similarity of the spatial position. The total distance metric is expressed as: in, is the total distance metric, is the similarity of the polarization coherence matrix, is the similarity of spatial position, is the weight coefficient, is the preset superpixel segmentation scale; Step 4.5: Take the pixel to be clustered as the center and Search for the new cluster center position in the area, and measure the pixels to be clustered and the The total distance metric between all pixels at the new cluster center positions in the region, dividing the pixels to be clustered into the cluster center closest to the total distance metric, and obtaining the corresponding category label; Step 4.6: After completing one iteration, calculate the average value of the position coordinates of all pixels in the superpixel area corresponding to each cluster, and use the position corresponding to the average value as the new cluster center position for the next iteration; Step 4.7: Repeat steps 4.3 to 4.6 until clustering reaches a preset number of iterations to obtain the superpixel segmentation result.

7. The PolSAR target detection method according to claim 1, wherein: The intensity image corresponding to the PolSAR image is detected using dual-parameter CFAR to obtain a CFAR binary detection result, including: Step 5.1, based on the PolSAR image, calculate the intensity map using the polarization scattering matrix; Step 5.2, selecting a number of clutter background areas in the intensity map; Step 5.3: Obtain a corresponding histogram according to each piece of the clutter background area, and based on the histogram, select different distribution models to fit the distribution of the histogram to obtain fitting distribution results corresponding to different distribution models; Step 5.4: obtaining a final fitting distribution result according to the fitting distribution result corresponding to the different distribution models that is closest to the change trend of the histogram; Step 5.5: Obtain the detection statistic of the pixel to be detected based on the maximum likelihood estimation of the mean and standard deviation corresponding to the final fitting distribution result. When the final fitting distribution result is a lognormal distribution, the detection statistic of the pixel to be detected is expressed as: in, The pixel to be detected The test statistic, is the maximum likelihood estimate of the mean corresponding to the final fitting distribution result, is the maximum likelihood estimate of the standard deviation corresponding to the final fitting distribution result; Step 5.6: Compare the detection statistic of the pixel to be detected with a threshold value T2. If the detection statistic of the pixel to be detected is greater than or equal to the threshold value T2, the pixel to be detected is determined to be a target, and the pixel to be detected is set to 1. If the detection statistic of the pixel to be detected is less than the threshold value T2, the pixel to be detected is determined to be background, and the pixel to be detected is set to 0, thereby obtaining a CFAR binary detection result for the pixel to be detected. Step 5.7: Repeat steps 5.5 to 5.6 to obtain the CFAR binary detection result corresponding to each pixel in the intensity map corresponding to the PolSAR image, thereby obtaining the CFAR binary detection result corresponding to the PolSAR image.

8. The PolSAR target detection method according to claim 7, characterized in that: Obtaining a final fitting distribution result according to the fitting distribution result corresponding to the different distribution models that is closest to the change trend of the histogram, including: For each of the clutter background regions, obtaining a fitting distribution result closest to the change trend of the histogram as the fitting distribution result to be used; Determine whether the fitting distribution results to be used corresponding to all the clutter background areas are consistent. If they are consistent, the fitting distribution results to be used are used as the final fitting distribution results. If they are inconsistent, the fitting distribution results to be used corresponding to the main clutter background area are used as the final fitting distribution results. The main clutter background area is the clutter background area with the largest area in the experimental data scene.

Images