A method for intelligent segmentation and attitude estimation of space target ISAR image components

Abstract

The application provides a kind of space target ISAR image component intelligent segmentation and attitude estimation method, through inverse synthetic aperture radar to space target is continuously observed, and utilize distance-Doppler imaging algorithm to the echo is sequentially imaged, obtain the ISAR image sequence of space target;Utilize deep learning network Pix2pixGAN to the ISAR image sequence is segmented, and the segmentation accuracy is higher;After removing the invalid connected region in each component segmentation result, obtain the final segmentation image of each component;Linear structure extraction is carried out to the final each component segmentation image using the minimum circumscribed rectangle method, and the method is lower to the image component segmentation accuracy requirement, and the robustness is stronger;Finally, according to the imaging principle of radar observation and ISAR image, an unconstrained optimization problem for solving three-dimensional attitude is constructed, and a particle swarm optimization algorithm is used for solving, and the solving efficiency is higher, to realize the attitude inversion of space target key component.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a space target ISAR image component intelligent segmentation and attitude estimation method. BACKGROUND

[0002] Accurate estimation of the attitude of a space target in orbit is an important part of space target monitoring. Real-time monitoring of the attitude of a target can determine the running state of the target and perform fault analysis. For non-cooperative targets, attitude estimation can provide important support for behavior and intention analysis. Space target monitoring mainly relies on optical and radar devices for tracking and observation. Inverse synthetic aperture radar (ISAR) has the ability to perform high-resolution imaging of space targets at all times and in all weather. Further, using ISAR image sequences obtained at different observation angles, the attitude of a space target can be inverted.

[0003] Most existing space target attitude inversion methods use morphological methods to extract the linear structure or target contour of key components in the ISAR image sequence to estimate the attitude. Zhou Yiejian et al. use the typical linear structure of a target in an ISAR image in combination with target orbit information to estimate the attitude of key components. Kou Peng et al. use morphological erosion operations to extract the main axis direction of a target and estimate the attitude of the target based on the main axis direction. Wang Jiadong et al. use a deep learning network to segment the main body of a satellite and estimate the absolute pointing direction and main body size of the satellite based on the segmented main body.

[0004] The above methods all have high requirements for the quality of ISAR images. However, due to the effects of electromagnetic anisotropy and occlusion between different components of a target, the sparsity of ISAR images, and the limitations of the ability of a target to represent its complete shape and contour, the image segmentation effect is poor, the linear structure is difficult to extract, and the attitude of the target cannot be accurately estimated. SUMMARY

[0005] To solve the above problems in the prior art, the application provides a space target ISAR image component intelligent segmentation and attitude estimation method. The technical problem to be solved by the application is solved by the following technical scheme.

[0006] The space target ISAR image component intelligent segmentation and attitude estimation method provided by the application comprises the following steps.

[0007] Step 1: continuously observing a space target by an inverse synthetic aperture radar to obtain echoes of the space target;

[0008] Step 2: using a range-Doppler imaging algorithm to sequentially image the echoes to obtain an ISAR image sequence of the space target;

[0009] Step 3: component segmentation of the ISAR image sequence is performed by using a deep learning network Pix2pixGAN to obtain each component segmentation image;

[0010] Step 4: invalid connected regions in the each component segmentation image are removed to obtain each component final segmentation image;

[0011] Step 5: linear structure extraction is performed on the final each component segmentation image by using a minimum circumscribed rectangle method to obtain linear parameters of each component;

[0012] Step 6: according to observation and imaging parameters of an ISAR system, a projection matrix corresponding to each frame image and a projection expression of linear structure of each component on an imaging plane are calculated;

[0013] Step 7: the projection expression and the linear parameters of each component are used to construct a non-constrained optimization problem for solving a three-dimensional space posture of each component;

[0014] Step 8: the non-constrained optimization problem is solved by using a particle swarm optimization algorithm to obtain posture parameters of each component in the three-dimensional space, and posture inversion of a key component of a space target is realized.

[0015] The space target ISAR image component intelligent segmentation and posture estimation method provided by the application has the advantages that:

[0016] The space target ISAR image component intelligent segmentation and posture estimation method provided by the application has the advantages that:

[0017] The application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of a space target ISAR image component intelligent segmentation and posture estimation method provided by an embodiment of the application;

[0019] Figure 2is a Pix2pixGAN network structure schematic diagram provided by an embodiment of the application;

[0020] Figure 3 is a generator structure schematic diagram of Pix2pixGAN provided by an embodiment of the application;

[0021] Figure 4 is a segmentation result schematic diagram of each component provided by an embodiment of the application;

[0022] Figure 5 is a principle schematic diagram of the minimum circumscribed rectangle provided by an embodiment of the application;

[0023] Figure 6 is a process schematic diagram of determining the minimum circumscribed rectangle with the minimum area provided by an embodiment of the application;

[0024] Figure 7 is a space target ISAR observation and imaging model schematic diagram provided by an embodiment of the application;

[0025] Figure 8 is a posture schematic diagram of the observed target linear structure in three-dimensional space provided by an embodiment of the application. DETAILED DESCRIPTION

[0026] The application will be further described in detail below with specific embodiments, but the embodiments of the application are not limited thereto.

[0027] As shown in Figure 1 , the space target ISAR image component intelligent segmentation and posture estimation method provided by the application comprises:

[0028] Step 1: continuously observing a space target by inverse synthetic aperture radar to obtain echo of the space target;

[0029] Step 2: using a range-Doppler imaging algorithm to sequentially image the echo to obtain an ISAR image sequence of the space target;

[0030] The application uses inverse synthetic aperture radar to continuously observe and record target echo of a space target, and then uses a range-Doppler imaging algorithm to perform two-dimensional high-resolution sequential imaging on the echo to obtain an ISAR image sequence of the space target.

[0031] Step 3: using a deep learning network Pix2pixGAN to segment components of the ISAR image sequence to obtain segmented images of each component;

[0032] As an optional embodiment of the application, the step 3 comprises:

[0033] Step 3-1: taking U-Net++ as a generator and taking a Markov discriminator as a discriminator to construct a Pix2pixGAN network;

[0034] Step 3-2: obtaining an ISAR image as an initial image;

[0035] The initial image carries a real label, and the real label is a three-channel picture, and the three-channel pictures mark the regions of the solar panel, the main body and the background of the space target respectively.

[0036] Step 3-3: training the constructed Pix2pixGAN network based on the initial image pair to obtain a trained Pix2pixGAN network;

[0037] Step 3-4: using the trained Pix2pixGAN network to perform component segmentation on the ISAR image sequence to obtain component segmentation results;

[0038] The component segmentation images include the segmentation images of the regions of the solar panel, the main body and the background.

[0039] It is worth noting that Figure 2 and Figure 3 , the generative adversarial network is an important generative model in the field of deep learning, which approximates some unsolvable loss functions through the adversarial learning of the generator and the discriminator, and has a wide application in the generation of image, video, natural language and music. Pix2pixGAN is one of the derivative models of GAN, and the present application improves Pix2pixGAN by taking U-Net++ as its generation network. U-Net++ is improved on the basis of U-Net, which is a system architecture with nested and dense skip connections. Compared with U-Net, U-Net++ can capture features at different levels and integrate them through feature stacking, and has better segmentation effect. The discriminator network adopts Markov discriminator. The ISAR image is input to the generation network as an initial image for component segmentation, and the real label is a three-channel picture, and the three channels mark the regions of the solar panel, the main body and the background. The real label and the segmentation result are input into the discriminator, and the discriminator is trained to distinguish them as much as possible. Through the continuous adversarial learning of the generation network and the discriminator network, a better component segmentation effect is obtained.

[0040] The trained network model is used for component segmentation of the ISAR image sequence for pose estimation, and the solar panel and the main body of the target are extracted.

[0041] Step 4: removing invalid connected regions in the component segmentation images to obtain final component segmentation images;

[0042] As an optional embodiment of the present application, the step 4 comprises:

[0043] Step 4-1: determining a pixel matrix of each component segmentation image;

[0044] Step 4-2: binarizing the pixel matrix of each component segmentation image to obtain a binarization matrix;

[0045] Step 4-3: filling each binarization matrix with a row of zero elements above and below and a column of zero elements on the left and right to obtain a filled binarization matrix;

[0046] It is worth noting that, as shown in Figure 4 , in most cases, due to the anisotropy of the target electromagnetic scattering characteristics, mutual shielding between components and the influence of noise, the target morphology of the IASR image is incomplete, the contour features are not obvious, and false extraction results may occur, which affects the component segmentation result. For example, Figure 4 , as shown in Fig. (b) of the drawings, the solar panel segmentation result cannot accurately represent the shape structure of the solar panel, and in order to ensure the accuracy of linear structure extraction, the connected region outside the solar panel needs to be removed. For an image with a size of MxN, the upper left corner coordinate is defined as (1, 1), the lower right corner coordinate is defined as (M, N), the horizontal direction is the Doppler axis, and the vertical direction is the distance axis. The pixel matrix of the component segmentation result is binarized, and a row of zero elements is filled above and below the matrix and a column of zero elements is filled on the left and right. The size of the filled matrix is (M+2)x(N+2).

[0047] Step 4-4: taking the element in the second row and the second column of the filled binarization matrix as a starting point, traversing the binarization matrix from left to right, and determining the connected region in which the pixel point is located according to the binarization values of the pixel point and its surrounding pixel points;

[0048] As an optional embodiment of the present application, the step 4-4 comprises:

[0049] Step 4-4-1: taking the element in the second row and the second column of the filled binarization matrix as a starting point, traversing the pixel matrix from left to right in the order of 2~M+1 rows and 2~N+1 columns. For a pixel point with index (x, y), if P(x, y) = 0, it is determined that the pixel point does not belong to any connected region;

[0050] Step 4-4-2: if P(x, y) = 1 and P(x-1, y-1) = 0, P(x-1, y) = 0, P(x-1, y+1) = 0, P(x, y-1) = 0, it is determined that the pixel point belongs to a new connected region;

[0051] wherein the pixel points belonging to the same connected region are assigned the same value, and the pixel points of different connected regions are assigned different values.

[0052] Step 4-4-3: If the pixel point P(x, y) = 1 and does not belong to the judgment condition in step 4-4-1 and step 4-4-2, it is judged that the pixel point belongs to the connected region in which the adjacent non-zero element point it has traversed is located.

[0053] Step 4-5: The number of pixel points contained in each connected region is counted, and for the segmented image of the solar panel, the two connected regions with the most pixel points are selected as the final segmentation result of the solar panel; for the segmented image of the main body, the connected region with the most pixel points is selected as the final segmentation result of the main body.

[0054] It is worth noting that the subscript (2, 2) is the starting point, and the 2~M+1 rows of the pixel matrix are traversed from left to right in the order of 2~N+1 columns. Assuming that the pixel matrix is P, the subscript of a pixel point is (x, y), if P(x, y) = 0, it is judged that the point does not belong to any connected region; if P(x, y) = 1 and P(x-1, y-1) = 0, P(x-1, y) = 0, P(x-1, y+1) = 0, P(x, y-1) = 0, it is judged that the point belongs to a new connected region, and the pixel values of the point and other points determined to belong to the connected region by traversal are set to the same value, and the pixel values of different connected regions are different; other cases of P(x, y) = 1 are judged that the point belongs to the connected region in which the adjacent non-zero element point it has traversed is located. The number of pixel points contained in each connected region is counted, and for the solar panel, the two connected regions with the most pixel points are retained as the solar panel segmentation image; for the main body, the connected region with the most pixel points is retained as the main body segmentation image.

[0055] Step 5: using the minimum circumscribed rectangle method, linear structure extraction is performed on the final component segmentation image to obtain the linear parameters of each component;

[0056] As an optional embodiment of the present application, the step 5 comprises:

[0057] Step 5-1: determining the minimum circumscribed rectangle containing each component final segmentation image;

[0058] Step 5-2: taking the center of each component final segmentation image as the rotation point, rotating the image pixel matrix of the final segmentation image by an angle of record the minimum circumscribed rectangle area and its vertex coordinates once every rotation until the minimum circumscribed rectangle with the smallest area is found;

[0059] Step 5-3: determining the length and direction of the component according to the rotation angle and vertex coordinates of the minimum circumscribed rectangle with the smallest area;

[0060] As an optional embodiment of the present application, the step 5-3 comprises:

[0061] Step 5-3-1: judging whether the side of the minimum circumscribed rectangle is parallel to the Doppler axis when the area of the minimum circumscribed rectangle is the minimum;

[0062] Step 5-3-2: if parallel, determining whether the parallel side is a long side or a short side, if a long side, determining the length of the component as L i =y maxi -y mini , then is the angle between the long side of the minimum circumscribed rectangle at the initial position and the distance axis, and let be converted into the angle with the Doppler axis;

[0063] Step 5-3-3: if a short side, determining the length of the component as L i =x maxi -x mini , then is the angle between the long side of the minimum circumscribed rectangle at the initial position and the Doppler axis, and let

[0064] wherein x mini is the horizontal coordinate of the top-left corner coordinate of the minimum circumscribed rectangle in the i-th rotation, y mi is the vertical coordinate, x maxi is the horizontal coordinate of the bottom-right corner coordinate of the minimum circumscribed rectangle in the i-th rotation, and y maxi is the vertical coordinate.

[0065] Step 5-4: determining the length and the direction of the component as the linear parameters of the component.

[0066] Taking a solar panel as an example, there are connected region 1 and connected region 2 in the solar panel segmentation result, the pixel values of the points in connected region 1 are all ξ1, and the pixel values of the points in connected region 2 are all ξ2. The coordinates of the points with the pixel value of ξ1 are obtained by traversing the pixel matrix, that is, the coordinates of the points in connected region 1 are ((x1, y1), (x2, y2)···(x n , y n )), and the coordinates of the points in connected region 2 are obtained by traversing, that is, n is the number of pixel points contained in connected region 1, and j is the number of pixel points contained in connected region 2.

[0067] If two adjacent sides of a rectangle are parallel to the distance axis and the Doppler axis of the image respectively, and contain all pixel points of a connected region, when the total number of pixel points contained in the rectangle is the minimum, the rectangle is called the minimum circumscribed rectangle of the connected region. The minimum value in x min is assigned to x , and the maximum value is assigned to xmax , The minimum value is given to y min , and the maximum value is given to y max , (x min , y min ), (x min , y max ), (x max , y min ), (x max , y max ) four points as the top of the rectangle as the minimum circumscribed rectangle of the solar panel segmentation result, define its area S = (x max -x min ) × (y max -y min ).

[0068] The solar panel and the main body are approximated as rectangles, and it can be known from Figure 5 that when the approximate rectangles of each component are rotated to be parallel to the distance axis and the Doppler axis, the minimum circumscribed rectangle area is minimum. With the image center as the rotation point, the image pixel matrix is rotated counterclockwise by an angle The rotation angle and the area S i of the minimum circumscribed rectangle at each time are recorded, and the coordinates (x min , y min ), (x min , y max ), (x max , y min ), (x max , y max ) of the top points of the minimum circumscribed rectangle are recorded, i ≤ k, k is the number of rotations.

[0069] For the solar panel, when the area S i of the minimum circumscribed rectangle is minimum, the long side of the solar panel points in the same direction as the long side of the minimum circumscribed rectangle. As shown in Figure 6 , if the side parallel to the Doppler axis is the short side of the rectangle at this time, then is the angle between the long side of the minimum circumscribed rectangle and the distance axis at the initial position, let be converted into the angle with the Doppler axis, and the length L i of the long side of the solar panel = y maxi -y mini , in units of pixels; if the side parallel to the Doppler axis is the long side of the rectangle, then is the angle between the long side of the minimum circumscribed rectangle and the Doppler axis at the initial position, let the length L i of the long side of the solar panel = X maxi -X mini, in pixels. For the body, the principal axis is directed in the same way as the length.

[0070] Step 6: According to the observation and imaging parameters of the ISAR system, a projection matrix corresponding to each frame of image and a projection expression of the linear structure of each component on the imaging plane are calculated.

[0071] As an optional embodiment of the present application, step 6 comprises:

[0072] Step 6-1: According to the observation and imaging parameters of the ISAR system, a projection matrix corresponding to each frame of image is calculated.

[0073] Step 6-2: The linear structure of each component is represented by a vector and multiplied by the projection matrix to obtain a projection expression of the linear structure of each component on the range axis and a projection expression on the Doppler axis.

[0074] It is worth mentioning that: the ISAR observation and imaging model of a space target is as shown in Figure 7 In the target body coordinate system, the present application adopts a classical 3D turntable model to describe the spatial geometric relationship between the target and the radar, the radar line of sight is l, the target relative to the radar line of sight elevation angle is φ, the azimuth angle is θ, and two linear structures on the target are s and m respectively. Assuming that at a certain time, the elevation angle and the azimuth angle of the target relative to the radar line of sight are φ(t) and θ(t), t is the observation time, then the instantaneous radar line of sight can be expressed as

[0075] l = [-cosφ(t)cosθ(t), -cosφ(t)sinθ(t), -sinφ(t)] T

[0076] Assuming that the instantaneous distance between the radar and the target center is represented by r0(t), then the instantaneous position of the radar is

[0077]

[0078] Assuming that the scattering point p on the target is n = [x n , y n , z n ] T The projection of the instantaneous distance between the radar and the scattering point on the radar line of sight is

[0079] r(t) = (p n -q t ) T ×l

[0080] Then

[0081] r n (t) = r0(t) - x ncosφ(t)cosθ(t)-y n cosφ(t)sinθ(t)-z n sinφ(t)

[0082] The Doppler frequency of the scattering point is

[0083]

[0084] where λ represents the wavelength of electromagnetic wave. Combining the above formula, we can get

[0085]

[0086] where,

[0087]

[0088]

[0089]

[0090] Let

[0091] As shown in Figure 8 , define the attitude angles α and β of the linear structure s, then

[0092] s=(cosαsinβ,cosαcosβ,sinα) T ,

[0093] Assuming that the attitude angle parameters α and β are constant during the observation time of the space target by the ISAR system, the projection of the linear structure s on the imaging plane is

[0094]

[0095] where,

[0096] .

[0097] represents the projection vector of the scattering point in the Doppler dimension,

[0098]

[0099] represents the projection vector of the scattering point in the distance dimension.

[0100] The radar image distance and Doppler pixel resolution are defined as

[0101]

[0102]

[0103] where c is the speed of light, f s is the radar sampling frequency, M R is the number of range sampling points of the radar echo, γ is the signal frequency modulation, M A is the number of echoes corresponding to the ISAR image, and PRF is the pulse repetition frequency.

[0104] According to the projection of the linear structure s on the imaging plane and the definition of the radar image range and Doppler pixel resolution, the projection of the linear structure s on the 2D ISAR image range and Doppler can be expressed as

[0105]

[0106]

[0107] where the unit is pixel.

[0108] For the k-th ISAR image, there are

[0109]

[0110] where -90°≤α≤90°, 0°≤β≤180°, P k is the projection matrix. The attitude angles ψ 1 , ψ 2 ,..., ψ k and the lengths L1, L2,..., L k of the linear structure in the ISAR image are taken as observation values; the attitude with the smallest difference from the observation values is searched as the attitude of the linear structure in space.

[0111] Step 7: using the projection expression and the linear parameters of each component, the solving problem of the attitude of each component in the three-dimensional space is constructed as an unconstrained optimization problem;

[0112] As an optional implementation of the present application, the step 7 comprises:

[0113] Step 7-1: substituting the projection expression into the linear parameters to obtain a transformed projection expression;

[0114] Step 7-2: using the transformed projection expression to construct a cost function to obtain the unconstrained optimization problem.

[0115] The length of the linear structure of the target in the ISAR image extracted by the minimum circumscribed rectangle method should be consistent with the length predicted by the projection relationship in the distance axis and the Doppler axis. The angle between the linear structure of the kth ISAR image extracted in step 4 and the Doppler axis is The length of the linear structure of the target in the ISAR image extracted by the minimum circumscribed rectangle method should be consistent with the length predicted by the projection relationship in the distance axis and the Doppler axis. The angle between the linear structure of the kth ISAR image extracted in step 4 and the Doppler axis is k The length of the linear structure of the target in the ISAR image extracted by the minimum circumscribed rectangle method should be consistent with the length predicted by the projection relationship in the distance axis and the Doppler axis. The angle between the linear structure of the kth ISAR image extracted in step 4 and the Doppler axis is

[0116]

[0117]

[0118] The cost function is constructed as follows

[0119]

[0120]

[0121] wherein, are the optimal values of the pitch angle and the azimuth angle obtained by searching.

[0122] Step 8: solving the unconstrained optimization problem by using the particle swarm optimization algorithm to obtain the attitude parameters of each component in the three-dimensional space, and realizing the attitude inversion of the key components of the space target.

[0123] As an optional embodiment of the application, the step 8 comprises:

[0124] Step 8-1: taking the unconstrained optimization problem as the objective function of the particle swarm;

[0125] Step 8-2: solving the objective function by using the particle swarm optimization algorithm to obtain the attitude parameters of each component in the three-dimensional space at the minimum of the objective function, and realizing the attitude inversion of the key components of the space target.

[0126] The solving process of the particle swarm optimization algorithm for the objective function is as follows:

[0127] (1) initializing the particle swarm algorithm: setting the particle number and the maximum iteration number of the PSO algorithm. Initializing the state of each particle, including the position and the speed. The particle position corresponds to the attitude parameters (α, β), which represent the elevation angle and the azimuth angle of the observed linear structure of the target. T p represents the minimum value of the objective function in the current iteration, and T g is the minimum value of the objective function in the whole search history, and T g is initialized as 1000000.

[0128] (2) objective function calculation:

[0129] According to the analysis in step 7, the objective function is

[0130]

[0131] where N represents the number of ISAR images used for pose estimation, ψ k represents the pose angle of the linear structure in the k-th ISAR image, L k represents the length of the linear structure in the k-th ISAR image, and (α,β) are the pose angle parameters of the linear structure in the target body coordinate system.

[0132] (3) Current optimal pose record

[0133] At each iteration, find the minimum target function value T and assign it to T p If T p <T g , the value of T p is assigned to T g , and the corresponding pose parameters of T p are recorded. The recorded pose is the current optimal pose.

[0134] (4) PSO termination judgment:

[0135] When the number of iterations of the algorithm is equal to the set threshold, the algorithm is terminated, and the minimum target function value T g and the corresponding pose parameters (α,β) are output. These parameters represent the elevation angle and azimuth angle of the observed target linear structure. Otherwise, update the state of each particle and go to step (2) target function calculation.

[0136] The present application uses a deep learning network Pix2pixGAN to intelligently extract multiple components in the ISAR image of a spatial target, and has higher segmentation accuracy. Compared with the existing method of extracting component pointing after image segmentation, the present application uses morphological processing and minimum circumscribed rectangle optimization method to extract component pointing, which has lower requirements for image component segmentation accuracy and stronger robustness. The present application uses a particle swarm optimization (PSO) algorithm for solving, and has higher solving efficiency.

[0137] The effects of the present application can be further illustrated by the following simulation experiment.

[0138] 1. Simulation conditions

[0139] The target echo is sequentially imaged by using a range-Doppler imaging algorithm, and the horizontal and vertical axes of the image are the Doppler axis and the range axis, respectively. The deep learning network Pix2pixGAN is trained by using an ISAR image dataset, and the trained model is saved. The ISAR image used for pose estimation is segmented by using the trained model, and then combined with a morphological processing method, the minimum circumscribed rectangle method is used to extract the linear structure of each component, and finally the pose estimation of the linear structure of each component is performed.

[0140] 2、Example

[0141] The Tiangong-1 satellite ISAR image sequence is selected as the dataset, and the components of the ISAR image are labeled as the true labels of the image segmentation results for training. The dataset information is shown in the following table.

[0142] Table 1 Experimental data information

[0143] Satellite type Number of training images (frames) Number of test images (frames) Tiangong-1 1200 24

[0144] According to the method of the application, the pitch angle a and the azimuth angle β of the linear structure s of the solar panel and the linear structure m of the main body in the target body coordinate system are estimated, and the results are as follows:

[0145] Table 2 Linear structure pose estimation table

[0146]

[0147] As can be seen from the above table, the ISAR image component linear structure pose estimation algorithm of the application is very accurate, and the pitch angle and azimuth angle errors of the estimated main body and solar panel linear structure pose are all within 2 degrees.

[0148] Compared with the existing morphological processing image component segmentation method, the application uses a deep learning network to intelligently extract multiple components in the ISAR image of a space target, and improves the Pix2pixGAN, uses U-Net++ as its generation network, and has higher segmentation accuracy; compared with the existing image segmentation method for extracting the pointing of the component, the application uses the minimum circumscribed rectangle optimization method to extract the pointing of the component, which has lower requirements for the image component segmentation accuracy and higher robustness; the application uses the particle swarm optimization algorithm for solving, and has higher solving efficiency.

[0149] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0150] Although the application has been described in connection with specific embodiments thereof, it will be understood that it is capable of further modifications and this application is intended to cover any variations using the principles of the application. For example, "comprising" shall not exclude other elements or steps. The inclusion of one particular limitation does not exclude others where these are not also mentioned. This specification is not intended to recite a complete list of all possible combinations.

[0151] The above description is further detailed in connection with specific preferred embodiments of the application, and it is not to be construed that the specific implementation of the application is limited to these descriptions. For those skilled in the art of the application, a number of simple deductions or substitutions can be made without departing from the concept of the application, and all of these should be considered as falling within the scope of protection of the application.

Claims

1. A method for intelligent segmentation and attitude estimation of space target ISAR image components, characterized in that, include: Step 1: Continuously observe the space target using inverse synthetic aperture radar to obtain the echo of the space target; Step 2: Use the range-Doppler imaging algorithm to perform sequential imaging of the echoes to obtain an ISAR image sequence of the space target; Step 3: Use the deep learning network Pix2pixGAN to segment the ISAR image sequence into components, and obtain segmented images of each component; Step 4: Remove invalid connected regions from the segmented images of each component to obtain the final segmented images of each component; Step 5: Using the minimum bounding rectangle method, perform linear structure extraction on the final segmented images of each component to obtain the linear parameters of each component; Step 6: Based on the observation and imaging parameters of the ISAR system, calculate the projection matrix corresponding to each frame of image and the projection expression of the linear structure of each component on the imaging plane; Step 7: Using the projection expression and the linear parameters of each component, the problem of solving the attitude of each component in three-dimensional space is constructed into an unconstrained optimization problem; Step 8: Solve the unconstrained optimization problem using the particle swarm optimization algorithm to obtain the attitude parameters of each component in three-dimensional space, thereby realizing the attitude inversion of key components of the space target.

2. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 1, characterized in that, Step 3 includes: Step 3-1: Use U-Net++ as the generator and Markov discriminator as the discriminator to construct the Pix2pixGAN network; Step 3-2: Acquire an ISAR image as the initial image; The initial image carries a real label, which is a three-channel image. The three channels respectively mark the areas where the solar panel, the main body, and the background of the space target are located. Step 3-3: Train the constructed Pix2pixGAN network based on the initial image to obtain the trained Pix2pixGAN network; Steps 3-4: Use the trained Pix2pixGAN network to segment the ISAR image sequence into components and obtain the segmentation results for each component; The segmentation results for each component include segmented images of the solar panel, the main body, and the background.

3. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 2, characterized in that, Step 4 includes: Step 4-1: Determine the pixel matrix of the segmented image for each component; Step 4-2: Binarize the pixel matrix of the segmented image of each component to obtain a binarized matrix; Step 4-3: Fill each binarized matrix with a row of zero elements at the top and bottom, and a column of zero elements at the left and right, to obtain the filled binarized matrix; Step 4-4: Starting from the element in the second row and second column of the filled binarized matrix, traverse the binarized matrix from left to right, and determine the connected region of the pixel based on the binarized value of the pixel and its surrounding pixels. Steps 4-5: Count the number of pixels contained in each connected region. For the segmented image of the solar panel, select the two connected regions with the most pixels as the final segmentation result of the solar panel; for the segmented image of the subject, select the connected region with the most pixels as the final segmentation result of the subject.

4. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 3, characterized in that, Step 4-4 includes: Step 4-4-1: Starting from the element in the second row and second column of the filled binary matrix, traverse the pixel matrix from left to right in order from row 2 to M+1 and column 2 to N+1. For a pixel with index (x,y), if P(x,y)=0, then determine that the pixel does not belong to any connected region. Step 4-4-2: If P(x,y)=1 and P(x-1,y-1)=0, P(x-1,y)=0, P(x-1,y+1)=0, P(x,y-1)=0, then the pixel is determined to belong to a new connected region; Pixels belonging to the same connected region are assigned the same value, while pixels in different connected regions are assigned different values. Step 4-4-3: If pixel P(x,y) = 1 and does not belong to the judgment conditions in Step 4-4-1 and Step 4-4-2, then the pixel is determined to belong to the connected region where its traversed adjacent non-zero element points are located.

5. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 1, characterized in that, Step 5 includes: Step 5-1: Determine the minimum bounding rectangle that contains the final segmented image of each part; Step 5-2: Using the center of the final segmented image of each component as the rotation point, rotate the image pixel matrix of the final segmented image counterclockwise by an angle. Record the area of ​​the smallest bounding rectangle and the coordinates of its vertices for each rotation, until the smallest bounding rectangle with the smallest area is found; Step 5-3: Determine the length and orientation of the component based on the rotation angle and vertex coordinates of the smallest bounding rectangle with the smallest area; Step 5-4: Determine the length and orientation of the component as linear parameters of the component.

6. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 5, characterized in that, Step 5-3 includes: Step 5-3-1: When the area of ​​the smallest bounding rectangle is the smallest, determine whether the side of the smallest bounding rectangle is parallel to the Doppler axis; Step 5-3-2: If they are parallel, determine whether the parallel side is the longer or shorter side. If it is the longer side, determine the length of the component as L. i =y maxi -y mini ,but Let be the angle between the longest side of the smallest bounding rectangle at the initial position and the distance axis. Converted to the angle with the Doppler axis; Step 5-3-3: If it is the shorter side, then determine the length of the component as L. i =x maxi -x mini ,but Let the angle between the longest side of the smallest circumscribed rectangle at the initial position and the Doppler axis be... Where, x m i ni Let y be the x-coordinate of the top-left vertex of the smallest bounding rectangle after the i-th rotation. m i ni x is the ordinate. maxi Let y be the x-coordinate of the lower right vertex of the smallest bounding rectangle after the i-th rotation. maxi The vertical axis is denoted as y.

7. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 1, characterized in that, Step 6 includes: Step 6-1: Calculate the projection matrix corresponding to each frame of image based on the observation and imaging parameters of the ISAR system; Step 6-2: Represent the linear structure of each component as a vector and multiply it by the projection matrix to obtain the projection expression of the linear structure of each component on the distance axis and the projection expression on the Doppler axis.

8. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 7, characterized in that, Step 7 includes: Step 7-1: Substitute the projection expression into the linear parameters to transform it, and obtain the transformed projection expression; Step 7-2: Construct the cost function using the transformed projection expression to obtain the unconstrained optimization problem.

9. The intelligent segmentation and attitude estimation method for space target ISAR image components according to claim 8, characterized in that, Step 8 includes: Step 8-1: Treat the unconstrained optimization problem as the objective function of the particle swarm optimization. Step 8-2: Solve the objective function using the particle swarm optimization algorithm to obtain the attitude parameters of each component in three-dimensional space that minimize the objective function, thereby realizing the attitude inversion of key components of the space target.

Images