A Radar Target Feature Extraction Method Based on Polarimetric Microwave Vision

Abstract

This invention proposes a radar target feature extraction method based on polarimetric microwave vision, comprising the following steps: Step 1, acquiring the full polarimetric measurement data of the radar target; Step 2, full polarimetric two-dimensional ISAR imaging of the radar target; Step 3, Pauli polarimetric decomposition of the full polarimetric ISAR image, followed by energy normalization; Step 4, constructing a pseudo-color image using the polarimetric decomposition results and representing it with quaternions; Step 5, constructing translational and rotational amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments; Step 6, calculating the quaternion translational and rotational amplitude-phase invariant moments of the pseudo-color image. This invention's radar target feature extraction method based on polarimetric microwave vision avoids the limitation of traditional coherent polarimetric decomposition methods, which can only extract local polarimetric scattering characteristics of radar targets. By utilizing color image visual processing technology, it ensures that the extracted features retain the integrity of the radar target, thus facilitating the understanding of radar target scattering characteristics and target recognition applications.

Description

[Technical Field]

[0001] This invention relates to a method for extracting features of man-made targets using fully polarimetric ISAR images, belonging to the field of polarimetric radar target recognition. Specifically, it involves using pseudo-color images after Pauli polarimetric decomposition to construct a translational and rotational amplitude-phase invariant moment as a target feature, and using this feature to achieve effective classification of different radar targets. [Background Technology]

[0002] Polarization is an important information dimension besides electromagnetic wave amplitude, phase, and frequency. It contains the phase relationship between various polarization channels. Jointly processing multi-channel polarization information of radar targets can yield richer and deeper target features. Since the geometry of man-made targets generally does not change, their polarization scattering characteristics are relatively robust to changes in time and space. If the polarimetric radar has sufficiently high range and azimuth resolution, it can reflect the polarization scattering characteristics of each strong scattering center of the man-made target in ISAR images. Traditional methods use coherent polarization decomposition methods (such as Pauli decomposition, Krogager decomposition, Cameron decomposition, etc.) to invert the geometric structure of the strong scattering centers of man-made targets. However, on the one hand, coherent polarization decomposition methods recognize target features based on the polarization scattering characteristics of the target's scattering centers, which is a "pixel-level" local target feature extraction method, not conducive to the effective recognition of the overall target structure; on the other hand, this method often obtains a scattering structure that does not match the actual target, which also increases the difficulty of target recognition.

[0003] To address the shortcomings of current coherent polarization decomposition methods, such as focusing on local aspects while neglecting the overall picture and insufficient robustness and accuracy in retrieving typical target structures, this invention, building upon traditional coherent polarization decomposition research, integrates color image visual processing and polarization information processing techniques to propose a radar target feature extraction method based on polarization microwave vision. Specifically, it constructs a translational and rotational amplitude-phase invariant moment based on quaternion Pseudo-Zernike moments as the polarization microwave visual feature of man-made targets. Because the invariant moment target feature extracted in this invention fully utilizes information from both the target's polarization and the image's visual dimensions, it exhibits better translational and rotational invariance. Experimental data shows that, compared to Pseudo-Zernike moments from single-channel images and the right-side Pseudo-Zernike moment feature from pseudo-color images, the invariant moment extracted in this invention demonstrates better imaging azimuth rotation invariance across the entire angular range of the "Pioneer" UAV target ISAR image, indicating stronger azimuth robustness for the target ISAR image and its potential as an effective feature for radar target identification. [Summary of the Invention]

[0004] The technical problem this invention aims to solve is: addressing the issues of current coherent polarization decomposition methods that "focus on the local while ignoring the overall" and lack robustness and accuracy in retrieving typical target structure types. This invention utilizes color image processing technology to fully leverage the fully polarized microwave scattering information of man-made targets. It extracts the overall target features of fully polarized ISAR images of radar targets by constructing a translation and rotation amplitude-phase invariant moment based on quaternion Pseudo-Zernike moments. Experimental data fully verify that this feature has better azimuth rotation invariance and can serve as an effective feature for radar target identification.

[0005] This invention discloses a radar target feature extraction method based on polarimetric microwave vision, the technical solution of which includes the following steps:

[0006] The first step is to acquire full polarization measurement data of the radar target.

[0007] Acquire broadband fully polarimetric scattering data of radar targets.

[0008] The second step is radar target fully polarimetric two-dimensional ISAR imaging.

[0009] ISAR imaging was performed on broadband scattering data from four polarization channels (HH, HV, VH, VV) to obtain fully polarimetric ISAR images.

[0010] The third step is to perform Pauli polarization decomposition on the fully polarimetric ISAR image and normalize it according to energy.

[0011] The fully polarimetric ISAR image was polarimetrically decomposed according to the Pauli basis into three components: |HH+VV|, |HH-VV|, and |2HV|, and then the energy was normalized.

[0012] The fourth step is to construct a pseudo-color image using the polarization decomposition results and represent it using quaternions.

[0013] The R channel of the pseudo-color image corresponds to the |HH-VV| component, the G channel corresponds to the |2HV| component, and the B channel corresponds to the |HH+VV| component; and this pseudo-color image is represented as a pure quaternion, i.e.: f RGB (x,y)=f R (x,y)i+f G (x,y)j+f B (x,y)k.

[0014] The fifth step is to construct translational and rotational amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments.

[0015] Quaternion Pseudo-Zernike moments possess translation invariance and amplitude rotation invariance, but lack phase rotation invariance. Therefore, it is necessary to construct quaternion moments with phase rotation invariance, which are called translation and rotation amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments, as features for target recognition.

[0016] Step 6: Calculate the quaternion translation and rotation amplitude-phase invariant moments of the pseudo-color image.

[0017] Calculate the phase invariant moments of the aforementioned quaternions for the pseudo-color image.

[0018] The advantages and beneficial effects of this invention are as follows: Applied to the field of polarimetric radar target recognition, this invention, through its proposed radar target feature extraction method based on polarimetric microwave vision, avoids the limitation of traditional coherent polarimetric decomposition methods, which can only extract local polarimetric scattering characteristics of radar targets. By utilizing color image visual processing technology, it ensures that the extracted features retain the integrity of the radar target, thus being more conducive to recognizing the scattering characteristics of radar targets and target recognition applications. Verification using anechoic chamber measurement data from the "Pioneer" UAV shows that, compared to the Pseudo-Zernike moments extracted using its single-channel ISAR images and the Pseudo-Zernike moments of the right-side quaternions in pseudo-color images, the invariant moments constructed in this invention exhibit better imaging azimuth rotation invariance across the entire angular range of the "Pioneer" pseudo-color images. This indicates that this feature is an effective feature for radar target recognition and has significant engineering application value in the field of polarimetric radar target recognition. [Attached Image Description]

[0019] Figure 1 This is a flowchart of a radar target feature extraction method based on polarized microwave vision.

[0020] Figure 2 This is a flowchart of the ISAR imaging process for radar targets.

[0021] Figure 3 This is a static measurement scene in a darkroom using the "Pioneer" drone.

[0022] Figure 4a This is the ISAR imaging result of the HH channel of the "Pioneer" UAV (the azimuth angle of the imaging center is 0°).

[0023] Figure 4b This is the ISAR imaging result of the HV channel of the "Pioneer" UAV (the azimuth angle of the imaging center is 0°).

[0024] Figure 4c This is the ISAR imaging result of the VH channel of the "Pioneer" UAV (the azimuth angle of the imaging center is 0°).

[0025] Figure 4dThis is the ISAR imaging result of the VV channel of the "Pioneer" UAV (the azimuth angle of the imaging center is 0°).

[0026] Figure 5a This is a false-color image of the "Pioneer" UAV (with an imaging center azimuth of 0°).

[0027] Figure 5b This is a false-color image of the "Pioneer" UAV (with an imaging center azimuth of 90°).

[0028] Figure 5c This is a false-color image of the "Pioneer" UAV (image center azimuth angle is -90°).

[0029] Figure 6a This is an R-channel image of the "Pioneer" UAV after energy normalization (image center azimuth angle is 0°).

[0030] Figure 6b This is a G-channel image of the "Pioneer" UAV after energy normalization (image center azimuth angle is 0°).

[0031] Figure 6c This is a B-channel image of the "Pioneer" UAV after energy normalization (image center azimuth angle is 0°).

[0032] Figure 7 The results show the comparison of the rotational invariance properties of different image moments (the azimuth angle of the imaging center ranges from -90° to 90°, and the highest order of each image moment is T = 6).

Detailed Implementation Methods

[0033] The following is in conjunction with the appendix Figure 1-7 The present invention will be further described below. The present invention is a radar target feature extraction method based on polarimetric microwave vision, and its implementation flowchart is shown below. Figure 1 As shown, the steps are as follows:

[0034] The first step is to acquire full polarization measurement data of the radar target.

[0035] Methods for acquiring broadband scattering data from radar targets include electromagnetic calculation, anechoic chamber measurement, and experimental testing. Choosing any one method is sufficient; essentially, they are all equivalent to frequency and angle sweeps on a fixed target. Frequency sweeping accumulates range-up data, while angle sweeping accumulates azimuth-up data. The specific process is as follows: Fix a radar target under test, and at a certain angle, transmit a single-frequency signal with a frequency step of Δf. The frequency scanning range is [f...]. min ,f maxSubsequently, the signal incident angle is changed in increments of Δθ, and the above process is repeated until the entire angular range of the target (i.e., [0°, 360°]) is traversed, thereby obtaining the broadband scattering data of the radar target. If H and V polarized signals are emitted to excite the target during the data acquisition process, and the echoes are received using H and V polarization, then the fully polarized broadband measurement data of the radar target can be obtained.

[0036] The second step is radar target fully polarimetric two-dimensional ISAR imaging.

[0037] The fully polarimetric measurement data includes four channels: HH, HV, VH, and VV. ISAR images are generated from the echo scattering data of each of the four channels. The imaging process is as follows: First, the echo data undergoes pulse compression processing. Depending on the processing domain, this can be divided into two methods: dechirp and matched filtering. Next, translational compensation is performed, including envelope alignment and phase correction, to obtain the turntable model. Finally, azimuth compression is applied to the rotationally compensated turntable model to obtain a two-dimensional ISAR image of the radar target. The ISAR imaging process for the radar target is as follows: Figure 2 As shown.

[0038] The third step is to perform Pauli polarization decomposition on the fully polarimetric ISAR image of the radar target and normalize the energy.

[0039] For a reciprocal target, the polarization scattering matrix S, in the backscattering case, S HV =S VH At this point, the three-dimensional Pauli feature vector can be represented as:

[0040]

[0041] Among them, S HH S HV S VH S VV These are the four elements of the scattering matrix S. It can be seen that the Pauli decomposition decomposes the target into... and This is a combination of the three ingredients.

[0042] In the above case, the energy of the scattering matrix S can be expressed as:

[0043] Span(S) = |K| 2 =|S HH | 2 +2|S HV | 2 +|S VV | 2 (2)

[0044] Similarly, after Pauli polarization decomposition, the energies of the three components can be expressed as follows:

[0045]

[0046]

[0047]

[0048] After Pauli decomposition, the fully polarimetric ISAR image of the radar target can be represented as three M×N matrices, where M is the number of pixels contained in the longitudinal range of the target, and N is the number of pixels contained in the lateral range of the target. Each element in the three ISAR images can be represented as P. z (x,y), x∈[1,M], y∈[1,N], z∈[1,3]. For example, P1(x,y) corresponds to the first component in the Pauli basis. And so on. Let P... z All elements in (x, y) are normalized in terms of energy, and the normalized value range is set to [0, 255] to facilitate representation as a pseudo-color image. This process is expressed as:

[0049]

[0050] Among them, P z Span(x,y) represents the energy value of each pixel in the energy-normalized ISAR image. max ) is for traversing P z (x,y) represents the maximum energy obtained by all elements.

[0051] The fourth step involves constructing a pseudo-color image using the Pauli polarization decomposition results and representing it using quaternions.

[0052] Generally, the R channel of a pseudo-color image corresponds to the |S| of the Pauli basis. HH -S VV |Component, corresponding to G channel|2S HV |Component, B channel corresponds to|S HH +S VV |Components. A quaternion consists of one real part and three imaginary parts, represented as:

[0053] q=a+bi+cj+dk (7)

[0054] Where a, b, c, and d are real numbers, and i, j, and k are three imaginary units.

[0055] A pseudo-color image can be represented as a pure quaternion (i.e., without the real part 'a'):

[0056] f RGB(x,y)=f R (x,y)i+f G (x,y)j+f B (x,y)k (8)

[0057] Among them, f R (x,y) represents the R channel of the pseudo-color image, f R (x,y)=P2′(x,y); f G (x,y) represents the G channel of the pseudo-color image. For the B channel of a pseudo-color image, f B (x,y)=P1′(x,y).

[0058] The fifth step is to construct translational and rotational amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments.

[0059] Let f RGB (r,θ) is a pseudo-color image in polar coordinates. The right-hand quaternion Pseudo-Zernike moment of order n and repetition m is defined as:

[0060]

[0061] n≥0, m can take positive or negative integer values, and |m|≤n.

[0062] Where μ is the identity quaternion. (r,θ) represents the magnitude and phase angle in polar coordinates, and |r|≤1; The superscript R represents the right side; R n,m (r) is the Pseudo-Zernike real-valued radial polynomial, expressed as:

[0063]

[0064] Where s is a parameter value that satisfies the range of n-|m|.

[0065] Since the quaternion Pseudo-Zernike moments are defined in polar coordinates (r, θ), the image coordinates need to be linearly transformed to a suitable region within the unit circle during calculation. Through this mapping transformation, the discrete form of equation (9) is approximated as follows:

[0066]

[0067] Where x and y are the pseudo-color images f RGB The x and y coordinates of (x, y). The mapping transformation from image coordinates to the unit circle is given by the following equation:

[0068]

[0069] in,

[0070] This shows that calculating the Pseudo-Zernike moments of quaternions requires ensuring that the number of elements on the horizontal and vertical axes of the image is equal, i.e., the image is in the form of N×N (M=N). Therefore, it is generally necessary to perform preprocessing such cropping on the image and recalibrate its horizontal or vertical distances.

[0071] As is known from the fundamental properties of quaternions, their multiplication does not satisfy the commutative law. Therefore, the definition of the left-hand side quaternion Pseudo-Zernike moment of order n and repetition degree m is also given:

[0072]

[0073] in, The superscript L represents the left side.

[0074] Similarly, the discrete form of the above equation can be expressed as:

[0075]

[0076] Through theoretical derivation and The relationship between them can be represented as:

[0077]

[0078] in,() * Represents the conjugate operation of quaternions.

[0079] Inverse transformations of equations (9) and (13) yield the pseudo-color image f. RGB (r,θ) can be approximately reconstructed by the T-order right-hand or left-hand quaternion Pseudo-Zernike moments, expressed as:

[0080]

[0081] The larger the highest order T, the smaller the error between the reconstructed image and the original image. This shows that the quaternion Pseudo-Zernike moments reflect the pseudo-color image f RGB The essential characteristics of (r,θ) indicate that this feature is expected to achieve effective classification of different targets.

[0082] A high-quality image moment should possess translation, rotation, and scale invariance. Since the lateral and longitudinal resolutions of ISAR images of man-made targets are usually consistent and unchanging, scale variations in ISAR images generally do not need to be considered. Furthermore, quaternion Pseudo-Zernike moments have inherent translation invariance; therefore, the focus is primarily on whether these moments possess rotation invariance.

[0083] Let f′ RGB (r,θ)=f RGB (r, θ - α), where α represents the rotation angle, can be obtained as follows:

[0084]

[0085]

[0086] The above two equations show and It possesses magnitude rotation invariance, but its phase changes during rotation. Therefore, it is necessary to construct image moments with translation and rotation amplitude-phase invariance, expressed as:

[0087]

[0088] in, That is, translation and rotation amplitude- and phase-invariant moments based on quaternion Pseudo-Zernike moments; The subscript k is a non-negative parameter variable, which is different from the imaginary unit vector k in formula (7).

[0089] like or The highest order is T, and according to the values ​​of n and m, the number of its moments is (T+1). 2 As for The number of its moments is

[0090] Step 6: Calculate the quaternion translation and rotation amplitude-phase invariant moments of the pseudo-color image.

[0091] Iterate through the imaging azimuth angles of the radar target and calculate the quaternion translation and rotation amplitude-phase invariant moments of the pseudo-color image of the radar target with different azimuth angles as the imaging center. The invariant moment multi-view dataset of the radar target is obtained and used as the target recognition feature.

[0092] Taking the "Pioneer" UAV target as an example, its fully polarimetric echo scattering data and fully polarimetric ISAR imaging were first obtained through anechoic chamber measurements. The "Pioneer" comprises major components such as the fuselage, wings, tail, engine, and propeller. It is 2.3 meters long, has a wingspan of 2.9 meters, a height of 0.66 meters, and an empty weight of 11 kg. It is constructed from complex materials, including fiberglass, carbon fiber, wood, and metal, with a structure primarily consisting of a lightweight frame covered by a skin. The static anechoic chamber measurement scenario for the "Pioneer" UAV is shown below. Figure 3As shown. During anechoic chamber measurements, the UAV's nose was positioned along the x-axis. The test frequency was 8GHz–12GHz, with a center frequency of 10GHz and a frequency interval of 20MHz; the pitch angle was 0°, and the azimuth angle was -180°–180° with an angle interval of 0.2°; linear full polarization (HH, HV, VH, VV). The ISAR imaging results of the four channels (HH, HV, VH, VV) of the "Pioneer" UAV with an azimuth angle of 0° as the imaging center angle are shown below. Figures 4a-4d As shown.

[0093] Next, pseudo-color images are constructed using its fully polarimetric ISAR imagery. The pseudo-color images with the UAV azimuth imaging center at 0°, 90°, and -90° are shown below. Figures 5a-5c As shown. It can be seen that, compared to Figure 5a , Figure 5b and Figure 5c The UAV imaging results show some missing target component structures, and the polarization decomposition results of each component have also changed. This indicates that as the azimuth imaging center changes, the quaternion translational and rotation amplitude-phase invariant moments of the UAV target also change. These will also change. However, note that this change is due to the different azimuth of the target in ISAR imaging, and the invariant moments... It still possesses strict translation and rotation amplitude and phase invariance.

[0094] Finally, to demonstrate the effectiveness of this method, the amplitude-phase invariant moments of quaternion translation and rotation of the "Pioneer" pseudo-color image are compared. The rotation-invariant properties of the Pseudo-Zernike moments of the single-channel image and the Pseudo-Zernike moments of the right-hand quaternions of the pseudo-color image are compared. The single-channel image includes the three basis channels after Pauli polarization decomposition and the HH polarization channel. The three-channel ISAR image of the radar target after Pauli polarization decomposition corresponds to the R, G, and B channels of the pseudo-color image. After energy normalization, the R, G, and B channel images of the "Pioneer" UAV with an azimuth angle of 0° as the imaging center angle are shown below. Figures 6a-6c As shown.

[0095] Similarly, let f s (r,θ) is a single-channel image in polar coordinates. The Pseudo-Zernike moments of order n and repetition m are defined as follows:

[0096]

[0097] Unlike the unit quaternion μ, j in the above formula is the imaginary unit.

[0098] Similarly, the discrete form of equation (20) can be expressed as:

[0099]

[0100] Different from and Since the above equation satisfies the commutative law of multiplication, there are no right-hand or left-hand Pseudo-Zernike moments, and it can be uniformly described by Pseudo-Zernike moments.

[0101] Similarly, by iterating through the azimuth imaging center angles of the "Pioneer" UAV, the pseudo-color image quaternion invariant moments at different azimuth angles are calculated. Compared with single-channel image Pseudo-Zernike moments Z n,m (f) and the Pseudo-Zernike moments of the quaternion on the right side of the pseudocolor image. Because their feature vectors have different dimensions, the dimensions of each moment need to be normalized by zero-padding. Then, the T-Sne algorithm is used to reduce the high-dimensional feature vectors to two dimensions and visualize them. The rotation-invariant performance of each moment is compared by the intra-class distance between the samples of each moment. Assuming the highest order of the above moments is T=6, the comparison results when the imaging center azimuth angle traverses from -90° to 90° are as follows. Figure 7 As shown.

[0102] Figure 7 This indicates that quaternion invariant moments Because it utilizes the full polarization information of the "Pioneer" UAV, the intra-class distance between samples is smaller, indicating stronger azimuth robustness when the azimuth angle of the imaging center traverses from -90° to 90°; while for the single-channel Pseudo-Zernike moment Z... n,m (f) Whether it's the three bases after Pauli polarization decomposition or the HH polarization channels, the intra-class distance between samples is greater than... A larger moment indicates weaker orientation robustness; furthermore, using the right-hand quaternion Pseudo-Zernike moment, which also utilizes fully polarized information, the intra-class distance of the samples is between the HH channel Pseudo-Zernike moment and... The comparison between the moments indicates that utilizing fully polarized information can increase the orientation robustness of Pseudo-Zernike moments, and the newly constructed moments exhibit superior performance; in particular, the distribution of sample points for the HH channel, the Pauli basis |HH+VV| component, and the fully polarized right-hand quaternion Pseudo-Zernike moments shows strong consistency. The above comparison demonstrates that the quaternion-invariant moments constructed in this invention... It exhibits better imaging azimuth rotation invariance across its entire angular range, indicating that this feature can serve as an effective characteristic for radar target identification and has significant engineering application value in the field of polarimetric radar target identification.

Claims

1. A radar target feature extraction method based on polarimetric microwave vision, characterized in that, The steps are as follows: Step 1: Acquire the full polarization measurement data of the radar target: Acquire broadband fully polarimetric scattering data of radar targets; Step 2, Radar target fully polarimetric two-dimensional ISAR imaging: ISAR imaging was performed on broadband scattering data from the four polarization channels HH, HV, VH, and VV respectively to obtain a fully polarimetric ISAR image; Step 3: Perform Pauli polarization decomposition on the fully polarimetric ISAR image and normalize it according to energy: The fully polarimetric ISAR image was polarimetrically decomposed according to the Pauli basis into three components: |HH+VV|, |HH-VV|, and |2HV|, and the energy was normalized. Step four: Construct a pseudo-color image using the polarization decomposition results and represent it using quaternions: The R channel of the pseudo-color image corresponds to the |HH-VV| component, the G channel corresponds to the |2HV| component, and the B channel corresponds to the |HH+VV| component; and this pseudo-color image is represented as a pure quaternion, i.e.: f RGB (x,y)=f R (x,y)i+f G (x,y)j+f B (x,y)k; where i, j, and k are three imaginary units; f R (x,y) represents the R channel of the pseudo-color image, f G (x,y) represents the G channel of the pseudo-color image, f B (x,y) represents the B channel of the pseudo-color image; Step 5: Construct translation and rotation amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments: Constructing quaternion moments with phase rotation invariance, known as translation and rotation amplitude-phase invariant moments based on quaternion Pseudo-Zernike moments, as features for target recognition; Step 6: Calculate the quaternion translation and rotation amplitude-phase invariant moments of the pseudo-color image: Calculate the quaternion amplitude-phase invariant moments of a pseudo-color image.

2. The radar target feature extraction method based on polarization microwave vision according to claim 1, characterized in that: In step one, the methods for acquiring broadband scattering data of the radar target include electromagnetic calculation, anechoic chamber measurement, and field experiment; one of these methods is selected. The specific process is as follows: A radar target to be tested is fixed, and a single-frequency signal with a frequency step of Δf is emitted at a certain angle, with a frequency scanning range of [f]. min ,f max Then, the signal incident angle is changed in increments of Δθ, and the above process is repeated until the target's entire angular range, i.e., [0°, 360°], is traversed, thus obtaining the wideband scattering data of the radar target; if H and V polarized signals are emitted to excite the target during the data acquisition process, and H and V polarized signals are used to receive the echo, then the fully polarized wideband measurement data of the radar target is obtained.

3. The radar target feature extraction method based on polarization microwave vision according to claim 1, characterized in that: In step two, the fully polarimetric measurement data includes four channels: HH, HV, VH, and VV. ISAR images are generated from the echo scattering data of the four channels. The imaging process is as follows: First, the echo data is processed by pulse compression. Depending on the processing domain of the compression, there are two processing methods: deskewing and matched filtering. Next, translational compensation is performed, including envelope alignment and phase correction, to obtain the turntable model. Finally, the azimuth compression is performed on the turntable model after rotation compensation to obtain the two-dimensional ISAR image of the radar target.

4. A radar target feature extraction method based on polarimetric microwave vision according to claim 1, 2, or 3, characterized in that: In step three, for the polarization scattering matrix S of the reciprocal target, in the case of backscattering, S HV =S VH At this point, the three-dimensional Pauli feature vector is represented as: Among them, S HH S HV S VH S VV These are the four elements of the scattering matrix S; Pauli decomposition decomposes the target into... and The combination of these three ingredients; The energy of the scattering matrix S is expressed as: Span(S)=|K| 2 =|S HH | 2 +2|S HV | 2 +|S VV | 2 (2) Similarly, after Pauli polarization decomposition, the energies of the three components are expressed as follows: After Pauli decomposition, the fully polarimetric ISAR image of the radar target is represented as three M×N matrices, where M is the number of pixels contained in the longitudinal range of the target, and N is the number of pixels contained in the lateral range of the target; each element in the three ISAR images is represented as P. z (x,y), x∈[1,M],y∈[1,N],z∈[1,3]; P1(x,y) corresponds to the first component in the Pauli basis. And so on; P z All elements in (x, y) are normalized in terms of energy, and the normalized value range is set to [0, 255] to facilitate representation as a pseudo-color image, as follows: Among them, P z Span(x,y) represents the energy value of each pixel in the energy-normalized ISAR image. max ) is for traversing P z (x,y) represents the maximum energy obtained by all elements.

5. The radar target feature extraction method based on polarization microwave vision according to claim 4, characterized in that: In step four, the R channel of the pseudo-color image corresponds to the |S| of the Pauli basis. HH -S VV |Component, corresponding to G channel|2S HV |Component, B channel corresponds to|S HH +S VV |Components; A quaternion consists of one real part and three imaginary parts, represented as: q=a+bi+cj+dk (7) Where a, b, c, and d are real numbers, and i, j, and k are three imaginary units; A pseudo-color image is represented as a pure quaternion: f RGB (x,y)=f R (x,y)i+f G (x,y)j+f B (x,y)k (8) Among them, f R (x,y) represents the R channel of the pseudo-color image, f R (x,y)=P2′(x,y); f G (x,y) represents the G channel of the pseudo-color image, f G (x,y)=P3′(x,y); f B (x,y) represents the B channel of the pseudo-color image. f B (x,y)=P1′(x,y)。 6. The radar target feature extraction method based on polarization microwave vision according to claim 5, characterized in that: In step five, let f RGB (r,θ) is a pseudo-color image in polar coordinates. The right-hand quaternion Pseudo-Zernike moment of order n and repetition m is defined as: Where μ is the identity quaternion. Let r be the magnitude and phase angle in polar coordinates, and |r|≤1; The superscript R represents the right side; R n,m (r) is the Pseudo-Zernike real-valued radial polynomial, expressed as: Where s is a parameter value that satisfies the range of n-|m|.

7. The radar target feature extraction method based on polarization microwave vision according to claim 6, characterized in that: Since the quaternion Pseudo-Zernike moments are defined in polar coordinates (r, θ), the image coordinates need to be linearly transformed to a suitable region within the unit circle during calculation; through mapping transformation, the discrete form of equation (9) is approximated as follows: Where x and y are the pseudo-color images f RGB The x and y coordinates of (x, y); the mapping transformation from image coordinates to the unit circle is given by the following formula: in, This shows that calculating the Pseudo-Zernike moments of quaternions requires ensuring that the number of elements on the horizontal and vertical axes of the image is equal, i.e., the image is in the form of N×N (M=N). Therefore, the image needs to be preprocessed by cropping and its horizontal or vertical distances recalibrated.

8. The radar target feature extraction method based on polarimetric microwave vision according to claim 7, characterized in that: The definition of a left-hand quaternion Pseudo-Zernike moment of order n and repetition degree m is: in, The superscript L represents the left side; Similarly, the discrete form of the above equation can be expressed as: and The relationship between them is represented as follows: in,() * Represents the conjugate operation of quaternions; Inverse transformations of equations (9) and (13) yield the pseudo-color image f. RGB (r,θ) is approximately reconstructed by the T-order right-hand or left-hand quaternion Pseudo-Zernike moments, and is expressed as: The larger the highest order T, the smaller the error between the reconstructed image and the original image; thus, the quaternion Pseudo-Zernike moments reflect the pseudo-color image f RGB The essential characteristics of (r,θ) indicate that this feature is expected to achieve effective classification of different targets.

9. A radar target feature extraction method based on polarization microwave vision according to claim 8, characterized in that: Let f′ RGB (r,θ)=f RGB (r, θ - α), where α represents the rotation angle, yields: The above two equations show and It possesses magnitude rotation invariance, but its phase changes during rotation; therefore, it is necessary to construct image moments with translation and rotation amplitude-phase invariance, expressed as: in, That is, translation and rotation amplitude- and phase-invariant moments based on quaternion Pseudo-Zernike moments; The subscript k is a non-negative parameter variable, which is different from the imaginary unit vector k in formula (7); like or The highest order is T, and according to the values ​​of n and m, the number of its moments is (T+1). 2 As for The number of its moments is 10. A radar target feature extraction method based on polarization microwave vision according to claim 1, characterized in that: Iterate through the imaging azimuth angles of the radar target and calculate the quaternion translation and rotation amplitude-phase invariant moments of the pseudo-color image of the radar target with different azimuth angles as the imaging center. The invariant moment multi-view dataset of the radar target is obtained and used as the target recognition feature.

Images