PolSAR target detection method based on super-pixel-level saliency improvement and contrast enhancement

Abstract

The invention discloses a PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement, and the method comprises the steps: obtaining a polarization scattering matrix and a matrix obtained after the polarization scattering matrix is transformed, and obtaining a plurality of polarization features; superpixel segmentation results of different scales are obtained; by using the segmentation results of the superpixels of different scales as indexes, calculating an average value of partial polarization characteristics randomly selected from the plurality of polarization characteristics to obtain polarization characteristic patterns of different scales of each polarization characteristic; constructing a multi-layer hierarchical saliency map, and performing normalization processing to obtain a feature saliency map; adding the feature saliency maps corresponding to each polarization feature in the partial polarization features to obtain a comprehensive feature saliency map; and performing contrast enhancement on the comprehensive feature saliency map, comparing with a threshold value to obtain a saliency result binary image, adding, and comparing with a voting threshold value to obtain a final detection result map. According to the invention, the detection performance of the PolSAR image vehicle target in a complex scene can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar image target detection, and particularly relates to a PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement. Background Art

[0002] Polarimetric Synthetic Aperture Radar (PolSAR) is an advanced remote sensing technology that uses the polarization characteristics of radar signals for target detection. Compared with traditional single-polarization SAR, polarimetric SAR can provide richer target information by measuring and analyzing different polarization channels of radar echo signals. This technology can better identify and distinguish different targets in complex environments, such as ground objects, buildings, water bodies, vegetation, etc.

[0003] With the development of remote sensing technology, the application of polarimetric SAR in the fields of environmental monitoring, agriculture, urban management, etc. is becoming more and more extensive. For the military field, target detection based on polarimetric SAR can effectively identify important targets; in disaster monitoring and environmental change analysis, polarimetric SAR can help identify ground disasters, building damage, etc.; in agriculture and vegetation monitoring, polarimetric SAR can provide analysis data for crop growth status, soil humidity, etc. Therefore, the polarimetric SAR feature extraction technology based on scattering mechanism not only has important theoretical value in the scientific research field, but also has broad prospects in practical applications.

[0004] By deeply exploring the scattering characteristics in polarimetric SAR images, the accuracy and reliability of target detection can be improved, which is of great significance for promoting the application and development of remote sensing technology. With the continuous progress of polarimetric SAR technology, this research will provide more possibilities for various target detection tasks, especially in complex scenarios.

[0005] The research on PolSAR-based target detection emerged in the late 1980s. According to the scale of the detection unit, PolSAR target detection algorithms can be divided into two categories: pixel-level target detection algorithms and pixel-block-level target detection algorithms. Among them, pixel-level target detection algorithms include the following categories: 1. Detection methods based on optimization techniques. In 1989, Novak et al. proposed the Optimal Polarization Detector (OPD). This method can achieve optimal detection according to the Neyman-Pearson criterion when the means and correlation matrices of the target and clutter are known. Subsequently, Chaney et al. proposed the Identity Likelihood Ratio Detector (ILRT), which simplifies the calculation by substituting the covariance matrix in the OPD. In 1990, Novak and Burl designed the Polarization Whitening Filter Detector (PWF). This method does not require prior information about the target and its performance is close to that of the OPD. 2. Detection methods based on statistical distribution models. This type of method designs target detection algorithms by modeling the statistical distribution of the polarization covariance matrix. For example, Song et al. developed a ship detector based on the assumption of a multivariate complex Gaussian distribution using the variational Bayesian method and achieved good detection results. 3. Detection methods based on physical scattering mechanism information. This method utilizes the scattering mechanism information in the polarization covariance matrix and conducts target detection by converting it into physical features. Nunziata et al. proposed the Reflection Symmetry (RS) detector, which measures the reflection symmetry of the target through the product of the co-polarization and cross-polarization channels. Wang et al. further developed this theory and achieved effective threshold segmentation by combining it with the adaptive CFAR detector. 4. Detection methods based on polarization decomposition. Polarization decomposition methods analyze the target by decomposing the polarization covariance matrix into matrices representing different scattering mechanisms. The basic scattering mechanisms include single, double, volume scattering, and helical scattering, etc. Polarization decomposition methods can be divided into two categories: coherent decomposition (such as Pauli, Krogager, Cameron decompositions) and incoherent decomposition (such as Yamaguchi, Freeman-Durden, Cloude decompositions).

[0006] There are two types of object detection algorithms at the pixel block level: 1. Detection methods based on superpixels. Siwei Chen et al. constructed an automatic detection algorithm at the superpixel scale based on the polarization rotation invariant features of objects. By measuring the local contrast of superpixel blocks, some significant patches in polarimetric SAR images were highlighted, and it was demonstrated on RADARSAT-2 data that this method can still achieve good detection performance under low signal-to-clutter ratio conditions. 2. Detection methods based on saliency. The team led by Professor Wang Wei of the National University of Defense Technology proposed to obtain a superpixel-level corrected polarimetric coherence measure from the perspective of physical scattering mechanisms. This measure can distinguish the echoes of small ships with low saliency and strong sea clutter with high saliency. By utilizing physical scattering mechanisms and polarimetric coherence measures, small ships can be effectively distinguished and identified, while suppressing strong sea clutter from the sea surface, thereby improving the accuracy of target detection.

[0007] Due to the particularity of PolSAR data, there is currently little research on ground target detection, and there are also certain difficulties that need to be solved:

[0008] 1) In the feature extraction of PolSAR images, the data contains complex, non-linear features, and multi-dimensional information. Since the dimension of PolSAR data is relatively high, and different polarization channels reflect different scattering characteristics, how to effectively reduce the dimension or select features of it, avoid information redundancy and retain useful features, is a challenge in feature extraction. At the same time, the extracted features need to solve subsequent problems through feature fusion methods, and how to reasonably fuse different features is also a challenge in feature extraction.

[0009] 2) The problem of diverse target scales is often encountered in target detection. For example, small targets and large targets may appear simultaneously in an image, which requires the algorithm to be able to adapt to targets of different scales. At the same time, the problem of target occlusion may be encountered. Target occlusion means that the target is covered by other objects or occluders in the image, such as trees, houses, etc., resulting in the difficulty of fully displaying target information. Due to these problems, it is required that the target detection algorithm has a certain scale adaptability, that is, it can accurately detect targets at different scales, and has a high ability to resist occlusion and other interferences, and can detect targets in complex environments.

[0010] In summary, the performance of current detection algorithms in PolSAR target detection still has certain deficiencies. Summary of the Invention

[0011] To solve the above problems existing in the prior art, the present invention provides a PolSAR target detection method based on superpixel-level saliency enhancement and contrast enhancement. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0012] In a first aspect, the present invention provides a PolSAR target detection method based on superpixel-level saliency enhancement and contrast enhancement, including:

[0013] S1. Obtain the original image, obtain the polarization scattering matrix according to the data of the original image, and perform transformation on the polarization scattering matrix to obtain the transformed polarization scattering matrix;

[0014] S2. Obtain multiple polarization features according to the polarization scattering matrix and the transformed polarization scattering matrix;

[0015] S3. Use a preset superpixel segmentation algorithm to perform superpixel segmentation on the original image to obtain superpixel segmentation results of different scales;

[0016] S4. Use the superpixel segmentation results of different scales as indexes, calculate the average values of some polarization features randomly selected from multiple polarization features, and obtain polarization feature maps of different scales for each polarization feature in the some polarization features;

[0017] S5. Obtain multi-layer hierarchical saliency maps according to the polarization feature maps of different scales for each polarization feature, perform normalization processing on the multi-layer hierarchical saliency maps to obtain feature saliency maps; add the feature saliency maps corresponding to each polarization feature in the some polarization features to obtain a comprehensive feature saliency map;

[0018] S6. Perform contrast enhancement on the comprehensive feature saliency map, and compare it with a threshold to obtain a binary saliency result map;

[0019] S7. Repeat S4 to S6 to obtain multiple binary saliency result maps, add the multiple binary saliency result maps, and compare the result with a threshold to obtain a final detection result map.

[0020] Advantages of the present invention:

[0021] In view of the problems of high false alarms and poor detection performance in existing PolSAR image target detection algorithms, the present invention provides a PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement. First, feature extraction is performed through the polarization scattering matrix and the transformation matrix of the polarization scattering matrix, and then different-scale superpixel segmentation results are obtained using the superpixel segmentation algorithm; different-scale polarization feature maps are calculated from some of the polarization features randomly selected from multiple polarization features, and then a feature saliency map is obtained; then the feature saliency maps are added together to obtain a comprehensive feature saliency map; considering that the feature saliency map has a high false alarm rate, the comprehensive feature saliency map needs to be enhanced in contrast and compared with a threshold to obtain a binary saliency result map; based on multiple selections of features, multiple binary saliency result maps are obtained, and after adding the multiple binary saliency result maps and comparing with the threshold, a final detection result map is obtained. The method provided by the present invention deeply studies polarization scattering characteristics, feature fusion methods, and superpixel-level characteristics, and can significantly improve the detection performance of vehicle targets in PolSAR images in complex scenarios.

[0022] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0023] Figure 1 is a flowchart of a PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement provided by an embodiment of the present invention;

[0024] Figure 2 is another flowchart of a PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement provided by an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of a segmentation result map with the selected scale S = 10 for superpixel scale segmentation provided by an embodiment of the present invention;

[0026] Figure 4 is a flowchart of obtaining a feature saliency map provided by an embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of an input PolSAR intensity image provided by an embodiment of the present invention;

[0028] Figure 6 is a schematic diagram of a final detection result map of 500 sampling votes provided by an embodiment of the present invention;

[0029] Figure 7 is a schematic diagram of a detection result map obtained according to the comparison algorithm PNF provided by an embodiment of the present invention;

[0030] Figure 8It is a schematic diagram of the detection result graph obtained according to the comparison algorithm PWF provided by the embodiment of the present invention;

[0031] Figure 9 It is a schematic diagram of the detection result graph obtained according to the comparison algorithm RS provided by the embodiment of the present invention. Specific embodiments

[0032] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0033] Please refer to Figure 1 and Figure 2 , Figure 1 It is a flowchart of a PolSAR target detection method based on superpixel-level saliency enhancement and contrast enhancement provided by the embodiment of the present invention, Figure 2 It is another flowchart of a PolSAR target detection method based on superpixel-level saliency enhancement and contrast enhancement provided by the embodiment of the present invention. A PolSAR target detection method based on superpixel-level saliency enhancement and contrast enhancement provided by the present invention includes:

[0034] S1. Obtain the original image, obtain the polarization scattering matrix according to the data of the original image, and perform a transformation on the polarization scattering matrix to obtain the transformed polarization scattering matrix.

[0035] Specifically, in this embodiment, first obtain the original PolSAR image and perform basic matrix transformation.

[0036] Construct a polarization scattering matrix according to the four channels of the original PolSAR image. In the horizontal and vertical polarization bases (H, V), the polarization scattering matrix is expressed as:

[0037]

[0038] In the polarization scattering matrix S (H,V) , the meanings of each channel are: S HH represents the scattering coefficient of the horizontally polarized incident wave that is scattered into a horizontally polarized wave; S HV represents the scattering coefficient of the horizontally polarized incident wave that is scattered into a vertically polarized wave (Vertical polarization, V); S VH represents the scattering coefficient of the vertically polarized incident wave that is scattered into a horizontally polarized wave; S VV represents the scattering coefficient of the vertically polarized incident wave that is scattered into a vertically polarized wave.

[0039] According to the principle of optimal power reception of polarization radar, obtain the Graves power scattering matrix of the polarization radar:

[0040]

[0041] Among them, represents the conjugate transpose of the polarization scattering matrix.

[0042] Under the condition that the reciprocity theorem S HV = S VH is satisfied, the expression of the Pauli basis polarization scattering vector is:

[0043]

[0044] The expression of the lexicographic basis polarization scattering vector is:

[0045]

[0046] The expression of the polarization coherence matrix is:

[0047]

[0048] The expression of the polarization covariance matrix is:

[0049]

[0050] Using the rotation matrix, the polarization scattering matrix S is transformed along the radar line of sight at the polarization orientation angle to obtain the rotated polarization scattering matrix, and the rotated polarization scattering matrix is expressed as:

[0051]

[0052] Among them, S(θ) is the rotated polarization scattering matrix, θ is the preset angle, and R2(θ) is the rotation matrix.

[0053] S2. According to the polarization scattering matrix and the transformed polarization scattering matrix, multiple polarization features are obtained.

[0054] S201. According to the polarization scattering matrix and the transformed polarization scattering matrix, multiple polarization features to be processed are obtained.

[0055] First, polarization features are extracted based on the polarization scattering matrix.

[0056] 1. SPAN = |S HH | 2 + |S HV | 2 + |S VH | 2 + |S VV | 2 , which represents the ability of the target to intercept and reflect the incident wave.

[0057] 2. Reciprocity = S HV -S VH , representing the reciprocity of the scattering mechanism.

[0058] 3. Where, λ1(S) and λ2(S) represent the eigenvalues of the scattering matrix S.

[0059] 4. Tr(S) = λ1(S) + λ2(S) = S HH +S VV , the first element of the Pauli vector, representing even-order scattering.

[0060] 5. Det(S) = λ1(S)λ2(S) = S HH S VV -S HV S VH , to a certain extent, represents the physical scale of the target. When the horizontal direction of the target is greater than the vertical direction, Det(S) > 1. When the horizontal direction of the target is less than the vertical direction, Det(S) < 1.

[0061] 6. Where, max[λ i (S H S)] represents the larger eigenvalue of the matrix S H S, which is the characterization of the maximum reflection power density of the target to the incident wave.

[0062] 7. Similar to SPAN, it represents the ability of the target to intercept and reflect the incident wave.

[0063] 8. m = |λ1(S)|, representing the maximum radar cross-section.

[0064] 9. Representation of multiple scattering. For odd-order scattering ν = 0, for even-order scattering

[0065] 10. Represents the target polarization sensitivity.

[0066] 11. Reflects the number of target scattering centers.

[0067] 12. Reflects the relative direction between the antenna and the eigen-polarization ellipse axis.

[0068] 13. Reflects the target symmetry.

[0069] In the circular polarization basis, the polarization scattering matrix is rewritten through a basis transformation as:

[0070]

[0071] The conversion of the scattering matrix is expressed as:

[0072] S (L,R) = TST H ;

[0073] where S represents the scattering matrix and T represents the circular polarization basis transformation matrix, which is defined as: T H represents the conjugate transpose.

[0074]

[0075] k s 、k D and k H are derived in the circular polarization basis as follows:

[0076]

[0077] 14、k S = |S RL |, representing the spherical scattering component coefficient.

[0078] 15、k D = min(|S LL |,|S RR |), representing the dihedral angle scattering component coefficient.

[0079] 16、k H = ||S RR |-|S LL ||, representing the helical scattering component coefficient.

[0080] Second, extract polarization features based on the power scattering matrix.

[0081] 1、 Characterize the maximum reflected power density of the target for the incident wave.

[0082] 2、Tr(G) = λ1(G) + λ2(G) = G 11 + G 22 represents the ability of the target to intercept and reflect the incident wave.

[0083] 3、Det(G) = |λ1(G)λ2(G)| = |G 11 G 22 - G 12 G 21 |, the determinant of the power scattering matrix.

[0084] 4、 represents the maximum scattering power of the target.

[0085] 5、 The Frobenius norm of the power scattering matrix.

[0086] Third, extract polarization features based on the polarization coherence matrix and the polarization covariance matrix.

[0087] The basic idea of the Huynen target decomposition theory is to decompose the input data into a single scattering target component and a residual component (referred to as the "N target") with the aim of extracting the desired target component from the clutter background.

[0088] Generally, the polarization coherence matrix T is used to describe a distributed target with nine independent parameters:

[0089]

[0090] The distributed target can be decomposed into an equivalent single target T0 with five independent parameters and a residual component target T containing the remaining four parameters N , as shown in the following equation:

[0091] T = T0 + T N ;

[0092]

[0093] 1、 Represents the total scattering power of the regular, smooth, and protruding parts of the scatterer.

[0094] 2、 Represents the total scattering power of the depolarizing components from the irregular, rough, and non-protruding parts of the scatterer.

[0095] 3、F = Im[T 23 , which represents the depolarizing component caused by the overall deformation of the scatterer.

[0096] 4、 Represents the total scattering power of the depolarizing components of the N target that is irregular, rough, and non-convex.

[0097] 5、 Represents the depolarizing component in the helicity of the asymmetric N target.

[0098] 6、 Represents the total scattering power of the depolarizing components of the irregular, rough, and non-convex single scatterer.

[0099] 7、 Represents the depolarizing component of the helicity of the asymmetric single scatterer target.

[0100] Fourth, extract features based on the rotation of the polarization coherence matrix.

[0101] Similar to the rotation polarization scattering matrix, using the rotation matrix, the polarization coherence matrix is transformed only by the polarization orientation angle along the radar line of sight to obtain the rotated polarization coherence matrix, and the rotated polarization coherence matrix is expressed as:

[0102]

[0103] where \(T\) represents the polarization coherence matrix and \(R_3(\theta)\) represents the three-dimensional rotation matrix.

[0104]

[0105] The data in part \(T(\theta)\) are expressed as:

[0106] \(T\) 11 (\theta)=T 11 ;

[0107] \(T\) 12 (\theta)=T 12 \cos2\theta + T 13 \sin2\theta;

[0108] \(T\) 13 (\theta)= - T 12 \sin2\theta + T 13 \cos2\theta;

[0109]

[0110] \(T\) 22 (\theta)=T 22 \cos 2 2\theta + T 33 \sin 2 2\theta + \text{Re}[T 23 \sin4\theta;

[0111] \(T\) 33 (\theta)=T 22 \sin 2 2\theta + T 33 \cos 2 2\theta - \text{Re}[T 23 \sin4\theta;

[0112] The square of the modulus of the off-diagonal terms of the coherence matrix also has polarization coherence and is expressed as:

[0113]

[0114] All elements of the rotated polarization coherence matrix can be expressed as sine functions:

[0115] f(\theta)=A\sin[\omega(\theta + \theta_0)] + B;

[0116] Among them, A represents the amplitude square, B represents the oscillation center, ω represents the angular frequency, and θ0 represents the initial phase.

[0117] 1. A_Re[T 12 = Re 2 [T 12 + Re 2 [T 13 , representing the amplitude square of Re[T 12 .

[0118] 2. A_Im[T 12 = Im 2 [T 12 + Im 2 [T 13 , representing the amplitude square of Im[T 12 .

[0119] 3. Represents the amplitude square of T 22 .

[0120] 4. Represents the amplitude square of |T 12 | 2 .

[0121] 5. Represents the oscillation center of T 22 .

[0122] 6. Represents the oscillation center of |T 12 | 2 .

[0123] 7. Represents the oscillation center of |T 23 | 2 .

[0124] In the circular polarization basis, the dictionary feature vector of the target can be written as:

[0125]

[0126] The polarization covariance matrix in the circular polarization basis can be expressed as:

[0127]

[0128] 8. Represents the LL co-polarization component.

[0129] 9. Represents the LR cross-polarization component.

[0130] 10. Represents the RR co-polarization component.

[0131] 11、 Represents the polarization correlation value between LL and LR.

[0132] 12、 Represents the polarization correlation value between LL and RR.

[0133] 13、 Represents the polarization correlation value between LR and RR.

[0134] 14、 Represents the odd-order scattering component.

[0135] 15、 Represents the sum of the even-order scattering component and the cross polarization.

[0136] 16、 Represents the helical scattering component.

[0137] 17、

[0138] 18、

[0139] 19、

[0140] 20、

[0141] 21、

[0142] 22、 Represents the degree of polarization, which is the proportion of the polarization component in the total intensity of the radar signal and is used to describe the polarization characteristics of the signal.

[0143] 23、 Represents the angular parameter based on the non-model three-component decomposition.

[0144] 24、 Represents the even-order scattering power based on the non-model three-component decomposition.

[0145] 25、 Represents the odd-order scattering power based on the non-model three-component decomposition.

[0146] 26、P v,fp =Tr(C)(1 - DOP), represents the volume scattering power based on the non-model three-component decomposition.

[0147] 27、 Symmetry parameter, used to describe the symmetry of the target.

[0148] 28、 Represents the complex correlation coefficient between the HH and VV channels.

[0149] 29. To a certain extent, it reflects the difference between man-made targets and natural backgrounds.

[0150] Fifth, extract features based on the 2D polarization coherence pattern.

[0151] Based on the Pauli eigenvector and the Lexicographic eigenvector Six typical polarization correlation patterns 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 |. For two polarization channels S1 and S2, the definition of polarization correlation is as follows:

[0152]

[0153] Here, since:

[0154]

[0155] The following conclusions can be obtained:

[0156] |γ (HH+VV)-(HH-VV) (θ)| = |γ (HH+VV)-HV (θ + π / 4)|;

[0157] |γ HH-HV (θ)| = |γ VV-HV (θ + π / 2)|;

[0158] Therefore, 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)| are finally obtained, which are respectively expressed as:

[0159]

[0160] Seven polarization coherence features can be extracted based on each polarization coherence pattern, which are expressed as follows:

[0161] 1. |γ 1-2 (θ)| org = |γ 1-2 (θ = 0)|, which is the polarization correlation feature usually used without any rotation processing, representing the target decorrelation effect of two polarization channels in the original imaging geometry.

[0162] 2. A measure representing the generalized coherence level and the target average decorrelation effect in the polarization rotation domain.

[0163] 3. Represents the standard deviation of the correlation in the rotation domain, which can identify the diversity of the target scattering direction.

[0164] 4.|γ 1-2 (θ)| max =max(|γ 1-2 (θ)|), which represents the upper limit of the polarization coherence of two polarization channels in the polarization rotation domain.

[0165] 5.|γ 1-2 (θ)| min =min(|γ 1-2 (θ)|), which represents the lower limit of the polarization coherence of the two polarization channels in the polarization rotation domain.

[0166] 6. |γ 1-2 (θ)| contrast =|γ 1-2 (θ)| max -|γ 1-2 (θ)| min , represents the coherence contrast in the polarization rotation domain.

[0167] 7. represents the relative coherence contrast in the polarization rotation domain.

[0168] According to the definition of polarization coherence mode, the above seven polarization coherence features corresponding to each polarization coherence mode are calculated, and a total of 28 polarization coherence features can be obtained.

[0169] According to step S201, 85 polarization features to be processed are obtained, but not all of these 85 polarization features to be processed can well distinguish targets and clutters. There is a certain redundancy, and it is necessary to perform preliminary screening of polarization features and eliminate some polarization features with too small signal-to-clutter ratios. Eliminating polarization features with too low signal-to-clutter ratios is to improve the effectiveness and reliability of features, avoid polarization features with too small signal-to-clutter ratios interfering with subsequent analysis, and effectively reduce data redundancy and improve calculation efficiency.

[0170] Since the above 85 features will be extracted from each pixel, a clutter area and a target area are selected respectively, and the average values ​​of the 85 features in each area are calculated, and the signal-to-clutter ratio is calculated. The specific steps are as follows.

[0171] Calculate the average of clutter and target area features.

[0172] First, a clutter area is selected in the image, which should represent the typical environment area as much as possible. Then a target area is selected to ensure that the target pixel features contained in the area can better reflect the actual characteristics of the target.

[0173] The 85 features of all pixels in each region are averaged to obtain the average performance of the two regions (clutter region and target region) in each feature.

[0174]

[0175] Among them, f i represents the i-th feature.

[0176] S202, calculating the signal-to-noise ratio of each polarization feature to be processed, wherein the signal-to-noise ratio expression of the i-th polarization feature to be processed is:

[0177]

[0178] Among them, SCR i represents the signal-to-noise ratio of the i-th polarization feature to be processed, f i represents the i-th polarization feature to be processed, i represents the index of the polarization feature to be processed, μ t (f i ) represents the average value of the i-th polarization feature to be processed in the target area, μ c (f i ) represents the average value of the i-th polarization feature to be processed in the clutter region.

[0179] S203, deleting the to-be-processed polarization features whose signal-to-noise ratio values ​​are less than the signal-to-noise ratio threshold, and obtaining a plurality of remaining polarization features.

[0180] The signal-to-clutter ratios of the 85 features calculated were sorted to select features with high signal-to-clutter ratios (i.e., features with strong ability to distinguish targets from clutter). A threshold for the signal-to-clutter ratio was set, where threshold = 5 was selected, features with SCR < 5 were eliminated, and the remaining features with high signal-to-clutter ratios were retained for subsequent experiments.

[0181] Through the above steps, features with strong ability to distinguish targets and clutter can be screened out from 85 features, reducing the interference of features with too small signal-to-clutter ratio in distinguishing clutter from targets, thereby providing more reliable input for subsequent target detection.

[0182] S3. Use a preset superpixel segmentation algorithm to perform superpixel segmentation on the original image to obtain superpixel segmentation results of different scales.

[0183] Using the improved SLIC algorithm for polarimetric SAR images, which is from the paper Wang Y, Liu H. PolSARShip Detection Based on Superpixel-Level Scattering Mechanism Distribution Features[J]. IEEE Geoscience and Remote Sensing Letters, 2015, 12(8): 1780-1784, and based on the selected scale sizes S = 5, S = 10, S = 15, S = 20, S = 30, S = 40, S = 50, 7 different scale superpixel segmentation results can be obtained. Please refer to Figure 3 , Figure 3 It is a schematic diagram of the segmentation result map of the selected scale S = 10 for the superpixel scale segmentation provided by the embodiment of the present invention.

[0184] S4. Using the superpixel segmentation results of different scales as indexes, calculate the average value of the partial polarization features randomly selected from multiple polarization features to obtain the polarization feature maps of different scales for each polarization feature in the partial polarization features.

[0185] Randomly select 15 features screened in S2. Using the 7 different scale superpixel segmentation results obtained in S3 as indexes, take the average value of each feature in the segmented regions at different scales. Each feature obtains 7 polarization feature maps of different scales. Arrange the 7 polarization feature maps from smallest to largest scale and from top to bottom, and arrange the original feature as S = 1 at the top. Thus, each of the 15 features obtains 8 polarization feature maps of different scales, which is called a feature pyramid.

[0186] S401. Randomly select partial polarization features from multiple polarization features.

[0187] Among the features with strong ability to distinguish targets and clutter screened from 85 features in S2, randomly select 15 features for obtaining polarization feature maps of different scales subsequently.

[0188] S402. Using the superpixel segmentation results of different scales as segmentation indexes, load them into the selected partial polarization features, and calculate the average value of all polarization features in the segmented regions divided according to the segmentation indexes as the value of the segmented region to obtain a feature map.

[0189] Load the different scale superpixel segmentation results obtained in S3 as indexes into the 15 extracted features. For each extracted feature f i (i = 1, 2...15), there are 7 superpixel segmentation indexes S j(j = 1, 2... 7), and the corresponding superpixel scales are [5, 10, 15, 20, 30, 40, 50] respectively.

[0190] For the extracted feature f i and the segmentation index S j , the extracted features are divided into regions according to the segmentation index, and the average value of all features within the region is taken. The average value of the k-th region is expressed as:

[0191]

[0192] where n represents the total number of pixels within the region, and f i (j) represents the feature value of the j-th pixel within the region. Use this average value μ k to replace all the feature values within the k-th region.

[0193] S403. Sort the feature maps of different scales to obtain a pyramid of polarization feature maps of different scales.

[0194] For each extracted feature f i (i = 1, 2... 15), 7 superpixel segmentation maps can be obtained. Arrange the 7 obtained superpixel segmentation maps in ascending order of segmentation scale from top to bottom. Then arrange the original feature map at the topmost. Finally, for each feature, eight feature maps of different scales are obtained. The eight polarization feature maps of different scales are used as a feature pyramid for subsequent experiments.

[0195] S5. According to the polarization feature maps of different sizes of each polarization feature, obtain a multi-level hierarchical saliency map, normalize the multi-level hierarchical saliency map to obtain a feature saliency map; add the feature saliency maps corresponding to each polarization feature among the 15 extracted polarization features to obtain a comprehensive feature saliency map.

[0196] For the feature pyramid of each feature (the topmost is the first layer, and the bottommost is the eighth layer), subtract the feature map of the first layer from the feature map of the fourth layer, subtract the feature map of the second layer from the feature maps of the fifth and sixth layers respectively, subtract the feature map of the third layer from the feature maps of the sixth and seventh layers respectively, and subtract the feature map of the fourth layer from the feature maps of the seventh and eighth layers respectively, that is, subtract the large scale from the small scale, which means subtracting the background feature of the large scale from the detailed feature of the small scale. Then, through a normalization model, normalize the 7 subtracted feature maps to the same order of magnitude, representing the hierarchical saliency map. Combine the 7 hierarchical saliency maps and perform normalization to obtain the saliency map under this feature. Since there are 15 features in each sampling, add the 15 feature saliency maps to obtain the comprehensive feature saliency map result of one sampling. Please refer to Figure 4 , Figure 4 is a flowchart of a schematic diagram of the process for obtaining the feature saliency map provided by an embodiment of the present invention.

[0197] S501. Subtract the feature map of the first layer from that of the fourth layer according to the pyramid of polarization feature maps at different scales to obtain the first-level saliency map; subtract the feature map of the second layer from that of the fifth layer to obtain the second-level saliency map; subtract the feature map of the second layer from that of the sixth layer to obtain the third-level saliency map; subtract the feature map of the third layer from that of the sixth layer to obtain the fourth-level saliency map; subtract the feature map of the third layer from that of the seventh layer to obtain the fifth-level saliency map; subtract the feature map of the fourth layer from that of the seventh layer to obtain the sixth-level saliency map; subtract the feature map of the fourth layer from that of the eighth layer to obtain the seventh-level saliency map.

[0198] Calculating the differences between different levels of the feature pyramid. Subtracting the feature map of a small scale from that of a large scale can highlight the high-intensity superpixel blocks in the small scale and suppress the background clutter.

[0199] S502. Add the saliency values of the multi-level saliency maps at corresponding positions and then perform normalization to obtain the feature saliency map.

[0200] Normalize the unnormalized saliency map to obtain the level saliency map. The normalization method is to first divide the entire unnormalized level saliency map by the value of the maximum saliency point to normalize it to the range of [0, 1]. Traverse the image with a search step of 16, extract the local maximum value within the range of 16×16, and use the local maximum value to adjust the entire image.

[0201] The specific operation is as follows:

[0202] Find the maximum saliency point in the unnormalized saliency map, and divide the entire unnormalized saliency map by the value of the maximum saliency point to initially normalize it to the range of [0, 1].

[0203] max = max(saliency values of all pixel points in the unnormalized saliency map);

[0204] Initial normalized saliency map = unnormalized saliency map / max;

[0205] After initial normalization, the value of the maximum saliency point in the initially normalized saliency map is M, and M = 1.

[0206] Select step = 16, traverse the image with a search step of 16, and extract the local maximum value m within the range of step×step, that is, within the i-th step×step range G i where m i is the local maximum value. By using the local maximum value m i to adjust the normalized image, within G iThe normalized hierarchical saliency image img within Gi is:

[0207] img Gi = preliminary normalized saliency map × (M - m i ) 2 ;

[0208] The advantage of normalization is that it can reduce the influence of the cluttered saliency maps obtained from poor features and focus on the saliency maps obtained from better features.

[0209] Therefore, the operation steps to obtain the seven - layer hierarchical saliency map are as follows:

[0210]

[0211] S503. Add the feature saliency maps corresponding to each of the 15 extracted polarization features to obtain a comprehensive feature saliency map.

[0212] Add the saliency values of the seven - layer hierarchical saliency map at the corresponding positions and then perform the normalization algorithm. The normalization algorithm here is the same as the one in S502. Thus, a feature saliency map is obtained from the seven - layer hierarchical saliency map. For each of the 15 features extracted in S4, a feature saliency map can be obtained. Add the 15 feature saliency maps to obtain the comprehensive feature saliency map for one sampling.

[0213] S6. Enhance the contrast of the comprehensive feature saliency map to obtain a binary saliency result map.

[0214] Select the super - pixel segmentation scale that best matches the target pixel size, perform super - pixel contrast enhancement on the SPAN intensity features of each super - pixel block to obtain a binary map of the strong - contrast region, and multiply this binary map of the strong - contrast region by each comprehensive feature saliency map to obtain the binary saliency result map for each sampling.

[0215] S601. According to the size of the target in the original image, select the corresponding - scale super - pixel segmentation result to obtain the target super - pixel block, and calculate the spatial contrast between the target super - pixel block and the preset window super - pixel block.

[0216] For the target, select the super - pixel segmentation scale that best suits its size. For the target size in this data, select the scale of S = 10. In the P×P window area around the target super - pixel, where P = 30, traverse the super - pixel blocks and calculate the spatial contrast between the target super - pixel block and the window super - pixel block.

[0217] The spatial contrast is expressed as:

[0218]

[0219] Among them, n represents the number of all superpixel blocks, s i represents the target superpixel block to be enhanced in contrast, B represents the set of superpixel blocks within the P×P region around the target superpixel, and b j represents the superpixel block in this set;

[0220] The expression of ω(s i , b j ) is as follows:

[0221]

[0222] Among them, represents the similarity of the polarization coherence matrix of the central pixel points of the superpixel block s i and b j , ω(s i , b i ) represents the position similarity of the central pixel points of the superpixel block s i and b j , D spatial represents the spatial position distance of the central pixel points of the superpixel block s i and b j , d wishart represents the Wishart distance of the central pixel points of the superpixel block s i and b j , and the formula is as follows;

[0223] The calculation formula for the spatial position distance of the central pixel points of two superpixel blocks is:

[0224]

[0225] Among them, (x s , y s ), (x b , y b ) respectively represent the coordinate positions of the central pixel points of the superpixel blocks s i , b j .

[0226] The calculation formula for the Wishart distance of the central pixel points of the superpixel blocks s i , b j is:

[0227]

[0228] Among them, Trace(·) represents the trace of the matrix, and q represents the dimension of the matrix.

[0229] S602. Use the superpixel segmentation result of the corresponding scale as the segmentation index, calculate the average value of the SPAN intensity features in the segmentation region divided according to the segmentation index, and use it as the result of the superpixel segmentation region.

[0230] S603. Use the spatial contrast between the target superpixel block and the preset window superpixel block to enhance the superpixel contrast of the result of the superpixel segmentation region.

[0231] S604. Set a contrast threshold. If the superpixel contrast enhancement result is greater than the contrast threshold, the superpixel region value is 1. If the superpixel contrast enhancement result is less than the contrast threshold, the superpixel region value is 0, and a binary map of the strong contrast region is obtained.

[0232] SPAN = |S HH | 2 +|S HV | 2 +|S VH | 2 +|S VV | 2 , indicating the ability of the target to intercept and reflect the incident wave.

[0233] Please refer to Figure 5 , Figure 5 is a schematic diagram of the SPAN intensity feature of the input PolSAR intensity image provided by the embodiment of the present invention. Use the superpixel segmentation result with S = 10 as the index, calculate the average value of the SPAN intensity features in each region, and use this average value as the result of the superpixel segmentation region.

[0234] SPAN(s i ) = s i Average value of the SPAN feature within the superpixel block;

[0235] Perform superpixel contrast enhancement on the SPAN intensity feature of each superpixel block. The objective function is the product of the spatial contrast and the SPAN intensity feature of the superpixel block, and the expression is as follows:

[0236] K(s i ) = SPAN(s i )D LS (s i );

[0237] Take the threshold T p , here set T p = 1000, screen out the strong contrast region, set the superpixel region value where K(s i ) is greater than the threshold to 1, and set the superpixel region value where K(s i ) is less than the threshold to 0, and obtain a binary map of the strong contrast region.

[0238] S605. Multiply the binary image of the strong contrast region with the comprehensive feature saliency map, and obtain a binary saliency result map after comparison with a threshold.

[0239] For each extraction of 15 features, a feature saliency map can be obtained for each feature. Add the 15 feature saliency maps to obtain a comprehensive feature saliency map. Multiplying the corresponding positions of the comprehensive feature saliency map and the binary image of the strong contrast region can reduce the influence of the weak contrast region on subsequent detection.

[0240] S606. Compare the saliency result with a threshold T l Here, set T l = 1.3. Set the superpixel blocks with saliency greater than the threshold to 1, and set the superpixel blocks with saliency less than the threshold to 0 to obtain a binary saliency result map.

[0241] S7. Repeat steps S4 - S6 to obtain multiple binary saliency result maps. Add the multiple binary saliency result maps by voting, and compare with a voting threshold to obtain the final detection result map.

[0242] Repeat the above steps S4 - S6, that is, repeat sampling 15 features. For each feature, obtain a polarization feature pyramid through different superpixel segmentation scales, obtain a feature saliency map according to the feature pyramid, add the 15 feature saliency maps to obtain a comprehensive feature saliency map, perform a superpixel-level contrast enhancement operation on the comprehensive feature saliency map, that is, multiply the corresponding positions of the comprehensive feature saliency map and the binary image of the strong contrast region, and take the threshold of the multiplied result as the result of one extraction for one vote. Accumulatively add the results of 500 repeated samplings, and compare with the voting threshold T v If it is greater than or equal to the voting threshold T v then set the value of the pixel point in the final detection result map to 1. If it is less than the voting threshold T v , then set the value of the pixel point in the final detection result map to 0 to obtain the final binary detection result. Please refer to Figure 6 , Figure 6 which is a schematic diagram of the final detection result map of 500 samplings and votes provided by an embodiment of the present invention. The red frame indicates a false alarm, that is, wrongly considering that there is a target, and the green frame indicates that the detection result is correct.

[0243] It should be noted that the selection of the voting threshold is set according to the situation. In this embodiment, the threshold is not limited. The voting threshold T v can be set to 300.

[0244] In summary, the PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement provided by the present invention extracts features from the PolSAR polarization scattering matrix and the polarization scattering matrix transformation matrix, and filters out high signal-to-clutter ratio features to distinguish clutter from targets. Combined with superpixel segmentation, saliency calculation is performed after feature fusion, which reduces a large amount of noise false alarms compared with the pixel-level detection results. Subsequently, contrast enhancement operations are performed on the superpixel blocks to enhance the saliency difference between the targets and the background, which can greatly improve the performance of target detection.

[0245] In an optional embodiment of the present invention, the effect of the PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement provided by the above embodiment is verified through simulation experiments, specifically:

[0246] A set of publicly available datasets is selected, and the experimental data is imaged in full polarization mode. The image size is 768×768, the targets in the image are all vehicle targets, the clutter regions are mainly forest trees and sand, and some vehicle targets are located in the forest and are blocked by trees.

[0247] Table 1 below presents the detection results of the method of the present invention on the above dataset and compares them with existing PolSAR image target detection algorithms: the Polarimetric Whitening Filter Detector (PWF) (from the paper "Optimal polarimetric processing for enhanced target detection [J]. IEEE Transactions on Aerospace and Electronic Systems, 1993, 29(1): 34-244"), the Reflection Symmetry (RS) target detector (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 the Polarimetric Correlation Pattern (PCP) target detection algorithm (from the paper "PolSAR ship detection based on polarimetric correlation pattern [J]. IEEE Geoscience and Remote Sensing Letters, 2020, 18(3): 471-475"). Three commonly used target-level detection evaluation metrics are used in Table 1: Precision (Pre), Recall (Rec), and F1-score to quantitatively analyze the performance of the detection algorithm, and the definitions are as follows:

[0248]

[0249] In the above formula, TP represents the number of true targets detected, FP represents the number of false alarms, and NP represents the total number of true targets in the image. The false alarm rate P at the pixel level f represents the ratio of the number of misdetected pixel points to the total number of background pixel points.

[0250] Draw the detection rectangle according to the final binary detection result in the following way. Obtain the rectangular detection frame based on the boundary of the eight-connected region with a value of 1 in the final detection result map. Similarly, obtain the rectangular frame corresponding to the real target based on the boundary of the eight-connected region of the real target in the GroundTruth. Calculate the IoU (Intersection over Union) between the detection frame and the real frame to determine whether the detection frame is a correct detection result. If the IoU between the detection frame and a certain target frame is greater than or equal to 0.5, it is considered that the detection frame matches the target frame, and the target detection is correct, indicating that the detection frame is a True Positive (TP), and the color of the rectangular frame is green. If the IoU between a certain detection frame and a certain target frame is less than 0.5 or does not match any target frame, it is considered that the detected target is a false alarm, and the detection frame is a False Positive (FP), and the color of the rectangular frame is red. If a real target that actually exists is not detected, it is considered that the target is missed, and the color of the rectangular frame is blue. The IoU calculation formula is as follows:

[0251]

[0252] Among them, A and B are the sets of pixel points of the detection frame and the real frame respectively, A∩B represents the set of pixel points of the intersection of the two rectangular frames, and A∪B represents the set of pixel points of the union of the two rectangular frames.

[0253] Table 1 Comparison of the detection performance between the method of the present invention and the existing methods

[0254]

[0255] As can be seen from the above table, and please refer to Figures 7 - 9 , Figure 7 is a schematic diagram of the detection result map obtained according to the comparison algorithm PNF provided by an embodiment of the present invention, Figure 8 is a schematic diagram of the detection result map obtained according to the comparison algorithm PWF provided by an embodiment of the present invention, Figure 9 is a schematic diagram of the detection result map obtained according to the comparison algorithm RS provided by an embodiment of the present invention. While maintaining a high recall rate, the precision rate of the method of the present invention is higher than that of the existing PWF, RS, and PNF detection algorithms.

[0256] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant are intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the article or device comprising said element. Similar words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0257] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are 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. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0258] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement, characterized in that, Including: S1. Obtain the original image, obtain the polarization scattering matrix according to the data of the original image, and perform transformation on the polarization scattering matrix to obtain the transformed polarization scattering matrix; S2. Obtain a plurality of polarization features according to the polarization scattering matrix and the transformed polarization scattering matrix; S3. Use a preset superpixel segmentation algorithm to perform superpixel segmentation on the original image to obtain superpixel segmentation results of different scales; S4. Use the superpixel segmentation results of different scales as indexes, calculate the average values of some randomly selected polarization features from the plurality of polarization features, and obtain polarization feature maps of different scales for each polarization feature in the some polarization features; S5. Obtain a multi-level significance map according to the polarization feature maps of different sizes of each polarization feature, perform normalization processing on the multi-level significance map to obtain a feature significance map; add the feature significance maps corresponding to all the polarization features in the some polarization features to obtain a comprehensive feature significance map; S6. Enhance the contrast of the comprehensive feature significance map to obtain a binary significance result map; S7. Repeat S4 to S6 to obtain a plurality of binary significance result maps, add the plurality of binary significance result maps, and compare with a voting threshold to obtain a final detection result map.

2. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 1, wherein The S2 includes: S201. Obtain a plurality of polarization features to be processed according to the polarization scattering matrix and the transformed polarization scattering matrix; S202. Calculate the signal-to-clutter ratio of each polarization feature to be processed, where the expression of the signal-to-clutter ratio of the i-th polarization feature to be processed is: Among them, SCR i represents the signal-to-clutter ratio of the i-th polarization feature to be processed, f i represents the i-th polarization feature to be processed, i represents the index of the polarization feature to be processed, μ t (f i ) represents the average value of the i-th polarization feature to be processed in the target area, μ c (f i ) represents the average value of the i-th polarization feature to be processed in the clutter area; S203. Delete the polarization features to be processed with the signal-to-clutter ratio less than the signal-to-clutter ratio threshold, and the remaining ones obtain a plurality of polarization features.

3. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 1, wherein The S4 includes: S401. Randomly select some polarization features from the plurality of polarization features; S402. Use the superpixel segmentation results of different scales as segmentation indexes, load them into the selected some polarization features, and calculate the average value of all the polarization features in the segmentation regions divided according to the segmentation indexes as the value of the segmentation region to obtain feature maps of different scales; S403. Sort the feature maps of different scales and add them to the original image to obtain a pyramid of polarization feature maps of different scales.

4. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 1, wherein The S5 includes: S501. Obtain a multi-level significance map according to the pyramid of polarization feature maps of different scales; S502. Add the significance values of the multi-level significance map at the corresponding positions, and then perform normalization processing to obtain a feature significance map; S503. Add the feature significance maps corresponding to all the polarization features in the some polarization features to obtain a comprehensive feature significance map.

5. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 4, characterized in that, The S501 includes: Subtract the first-layer feature map from the fourth-layer feature map in the pyramid of polarization feature maps of different scales to obtain the first-layer hierarchical saliency map, subtract the second-layer feature map from the fifth-layer feature map in the pyramid of polarization feature maps of different scales to obtain the second-layer hierarchical saliency map, subtract the second-layer feature map from the sixth-layer feature map in the pyramid of polarization feature maps of different scales to obtain the third-layer hierarchical saliency map, subtract the third-layer feature map from the sixth-layer feature map in the pyramid of polarization feature maps of different scales to obtain the fourth-layer hierarchical saliency map, subtract the third-layer feature map from the seventh-layer feature map in the pyramid of polarization feature maps of different scales to obtain the fifth-layer hierarchical saliency map, subtract the fourth-layer feature map from the seventh-layer feature map in the pyramid of polarization feature maps of different scales to obtain the sixth-layer hierarchical saliency map, and subtract the fourth-layer feature map from the eighth-layer feature map in the pyramid of polarization feature maps of different scales to obtain the seventh-layer hierarchical saliency map.

6. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 1, wherein The S6 includes: S601. Select the superpixel segmentation result of the corresponding scale according to the size of the target in the original image to obtain the target superpixel block, and calculate the spatial contrast between the target superpixel block and the preset window superpixel block; S602. Use the superpixel segmentation result of the corresponding scale as the segmentation index, calculate the average value of the SPAN intensity features in the segmentation region divided according to the segmentation index, and use it as the result of the superpixel segmentation region; S603. Use the spatial contrast between the target superpixel block and the preset window superpixel block to enhance the superpixel contrast of the result of the superpixel segmentation region; among them, the objective function is the product of the spatial contrast and the SPAN intensity feature of the superpixel block, expressed as: K(s i ) = SPAN(s i )D LS (s i ); where SPAN = |S HH | 2 + |S HV | 2 + |S VH | 2 + |S VV | 2 , representing the ability to target intercept and reflect incident waves; S604. Set a contrast threshold. If the result of the superpixel contrast enhancement is greater than the contrast threshold, the superpixel region value is 1. If the result of the superpixel contrast enhancement is less than the contrast threshold, the superpixel region value is 0 to obtain a binary map of the strong contrast region; S605. Multiply the binary map of the strong contrast region by the comprehensive feature saliency map to obtain a binary map of the saliency result.

7. The PolSAR target detection method based on superpixel-level saliency improvement and contrast enhancement according to claim 6, characterized in that, The spatial contrast between the target superpixel block and the preset window superpixel block is expressed as: Among them, n represents the number of all superpixel blocks, and s i represents the target superpixel block to be enhanced in contrast, B represents the set of superpixel blocks within the P×P region around the target superpixel, and b j represents the superpixel block in this set. represents the similarity of the polarization coherence matrix of the central pixel points of the superpixel blocks s i and b j . ω(s i , b i ) represents the position similarity of the central pixel points of the superpixel blocks s i and b j . D spatial represents the spatial position distance of the central pixel points of the superpixel blocks s i and b j . d wishart represents the Wishart distance of the central pixel points of the superpixel blocks s i and b j . (x s , y s ), (x b , y b ) respectively represent the coordinate positions of the central pixel points of the superpixel blocks s i and b j . Trace(·) represents the trace of a matrix, and q represents the dimension of the matrix.

Images