ISAR coherent translation compensation and imaging method and device based on deep network

Abstract

The invention provides an ISAR coherent translation compensation and imaging method and device based on a deep network, and relates to the technical field of radar signal processing. Comprising the following steps: performing pulse compression on ISAR echoes to obtain a plurality of echo signals after pulse compression; sequentially inputting the plurality of echo signals after pulse compression into a trained LSTM strategy network, so that the trained LSTM strategy network outputs a plurality of first distance grid offset estimated values; obtaining a target translation parameter according to the plurality of first distance grid offset estimation values and the resolution of the distance dimension; estimating the target translation parameter by using an Adam optimizer to obtain an optimal target translation parameter; and according to the optimal target translation parameter, performing compensation and coherent accumulation on the echo signals subjected to pulse compression to obtain a focused image. In this way, the compensation precision of the echo signals after the multiple pulses are compressed is high, and the accuracy of the imaging result of the focused image is high.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to an ISAR coherent translation compensation and imaging method and device based on a deep network. Background Art

[0002] Inverse Synthetic Aperture Radar (ISAR) has all-day, all-weather, long-range and high-resolution imaging capabilities, and has important application value in the fields of situational awareness, prevention and control warning, civil aviation control, etc. During ISAR imaging, the movement of non-cooperative targets can be equivalently decomposed into two processes: translation along the line of sight and rotation around the center of the turntable. Among them, translation along the line of sight is not only meaningless to ISAR imaging, but also introduces envelope translation and phase errors, resulting in poor quality of the final imaging result. The process of compensating for it is called translation compensation.

[0003] At present, translation compensation is usually achieved by using a new parameterized ISAR motion compensation method based on particle swarm optimizer to estimate polynomial coefficients, and selecting Average Range Profile Energy (ARPE) and Entire Image Contrast (EIC) as the fitness functions of envelope alignment and phase autofocus, respectively. However, the envelope offset estimation result with ARPE as the objective function often cannot meet the requirements of focused imaging for range resolution accuracy, and the envelope offset estimation error cannot be compensated in the subsequent phase error estimation, which limits the estimation accuracy of the entire algorithm, making the accuracy of ISAR coherent translation compensation low, resulting in low accuracy of imaging results. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide an ISAR coherent translation compensation and imaging method and device based on a deep network, so as to solve the problem that the accuracy of ISAR coherent translation compensation is low, resulting in low accuracy of imaging results.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of the present invention provides an ISAR coherent translation compensation and imaging method based on a deep network, the method comprising:

[0007] Perform pulse compression on the ISAR echo to obtain multiple pulse compressed echo signals;

[0008] Inputting the echo signals after multiple pulse compression into the trained LSTM strategy network in sequence, so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values, wherein the trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set having the same signal form as the echo signals after multiple pulse compression;

[0009] Obtaining target translation parameters according to a plurality of first range grid offset estimation values ​​and a resolution of the range dimension;

[0010] The Adam optimizer is used to estimate the target translation parameters and obtain the optimal target translation parameters;

[0011] According to the optimal target translation parameters, the echo signals after multiple pulse compression are compensated and coherently accumulated to obtain a focused image.

[0012] A second aspect of the present invention provides an ISAR coherent translation compensation and imaging device based on a deep network, the device comprising:

[0013] A pulse compression module is used to perform pulse compression on the ISAR echo to obtain multiple pulse compressed echo signals;

[0014] An input module is used to sequentially input the echo signals after multiple pulse compression into the trained LSTM strategy network, so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values, wherein the trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set with the same signal form as the echo signals after multiple pulse compression;

[0015] An acquisition module, used for acquiring target translation parameters according to a plurality of first range grid offset estimation values ​​and a resolution of the range dimension;

[0016] The estimation module is used to estimate the target translation parameters using the Adam optimizer to obtain the optimal target translation parameters;

[0017] The compensation and coherent accumulation module is used to compensate and coherently accumulate echo signals after multiple pulse compression according to the optimal target translation parameters to obtain a focused image.

[0018] Compared with the prior art, the ISAR coherent translation compensation and imaging method and device based on a deep network provided by the present invention performs pulse compression on ISAR echoes to obtain multiple pulse compressed echo signals; the multiple pulse compressed echo signals are sequentially input into a trained LSTM strategy network so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values; the target translation parameters are obtained according to the multiple first distance grid offset estimation values ​​and the resolution of the distance dimension; the target translation parameters are estimated using an Adam optimizer to obtain the optimal target translation parameters; according to the optimal target translation parameters, the multiple pulse compressed echo signals are compensated and coherently accumulated to obtain a focused image. In this way, the Adam optimizer is used to achieve a more stable solution to non-convex problems, the target translation parameters can be accurately estimated, the influence of the target translation component on imaging can be eliminated, and the imaging result of the focused image obtained is clearer and more precise, so that the accuracy of the imaging result of the focused image is higher; using the optimal target translation parameters, the echo signals after multiple pulse compressions are directly compensated and coherently accumulated, which avoids the transmission of different target translation parameter estimation errors in the cascade compensation method, resulting in error accumulation in the final compensation result, and the compensation accuracy of the echo signals after multiple pulse compressions is higher, thereby improving the focusing degree and image calibration accuracy of the imaging result of the focused image, so that the accuracy of the imaging result of the focused image is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] By reading the detailed description below with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0020] Figure 1 The flowchart of the ISAR coherent translation compensation and imaging method based on the deep network is schematically shown;

[0021] Figure 2 The optical image of the experimental target is schematically shown;

[0022] Figure 3 A comparison diagram of the imaging results of the method of the present invention and the imaging results of the traditional method under different signal-to-noise ratios is schematically shown;

[0023] Figure 4 The loss function is schematically shown as a function of the signal-to-noise ratio;

[0024] Figure 5 The structure diagram of the ISAR coherent translation compensation and imaging device based on the deep network is schematically shown. DETAILED DESCRIPTION

[0025] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0026] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present invention should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0027] The method in the embodiment of the present invention is described in detail below.

[0028] Figure 1 The flowchart of the ISAR coherent translation compensation and imaging method based on the deep network in the embodiment of the present invention is schematically shown. Figure 1 As shown, the method may include:

[0029] S101 . Perform pulse compression on the ISAR echo to obtain a plurality of pulse compressed echo signals.

[0030] ISAR echo is a radar echo with large bandwidth and long time obtained by radar receiver.

[0031] Specifically, pulse compression is performed on the ISAR echo to obtain multiple pulse compressed echo signals, including:

[0032] Step A1: Divide the ISAR echo into multiple sub-apertures.

[0033] Step A2: performing range compression processing on multiple sub-apertures to obtain corresponding one-dimensional range images, and transforming the one-dimensional range images into the frequency domain to obtain multiple frequency domain signals.

[0034] The de-linear frequency modulation technique is used to perform range compression processing on multiple sub-apertures to obtain the corresponding one-dimensional range image.

[0035] Step A3: According to the pulse compression method, compensation factors of residual video phase (RVP) items and envelope ramp items corresponding to multiple frequency domain signals are constructed.

[0036] Step A4: using the compensation factors of the residual video term and the envelope ramp term, phase compensation is performed on the peak frequency of each target scattering point in the multiple frequency domain signals to obtain multiple pulse compressed echo signals.

[0037] The multiple pulse compressed echo signals may also be referred to as multiple compensated echo signals.

[0038] S102. Inputting the echo signals after multiple pulse compression into the trained LSTM strategy network in sequence, so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values.

[0039] The trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set having the same signal form as the echo signals after compression of a plurality of pulses.

[0040] Specifically, the echo signals after multiple pulse compression are sequentially input into the trained LSTM strategy network, so that the hidden nodes in the trained LSTM strategy network obtain multiple first distance grid offset estimation values ​​through a fully connected layer with an output dimension of 1.

[0041] The plurality of first distance grid offset estimation values ​​are distance grid offset estimation values ​​corresponding to the center position of the echo target.

[0042] Specifically, the LSTM strategy network is trained using the training set, including:

[0043] Step B1: Use the predicted distance parameters in the training set as the initial state s0 of the LSTM policy network.

[0044] The predicted distance parameter is a parameter including the predicted speed, the predicted acceleration and the predicted jerk.

[0045] Step B2: Based on the initial state s0, feedback the training action a0 of the LSTM policy network.

[0046] Step B3: Acquire the actual state s1 according to the actual distance parameters of the echo signals after multiple pulse compression.

[0047] Step B4: Calculate the gap loss between the initial state and the actual state.

[0048] The gap loss can be expressed as c0(s0,s1).

[0049] Step B5: Update the initial state according to the training action, gap loss and learning rate, and return to the step of feeding back the training action of the LSTM policy network according to the initial state until the total number of training times reaches the preset total number of times, and then stop the training to obtain the trained LSTM policy network.

[0050] According to the training action, gap loss and learning rate, the initial state is updated and the process returns to step B2 until the total number of training cycles reaches the preset total number of training cycles, thereby obtaining a trained LSTM policy network.

[0051] The preset total number of times may be 200 times, and there may be multiple preset total numbers, which are not limited here.

[0052] In addition to the LSTM strategy network, the present invention can also use other timing-dependent recurrent neural network structures, such as a gated recurrent unit (Gated Recurrent Unit, GRU) network, a recurrent neural network (Recurrent Neural Network, RNN), a bidirectional long short-term memory (Bidirectional Long Short Term Memory, Bi LSTM) network, a stacked long short-term memory (Stacked Long Short Term Memory, Stacked LSTM), etc. to perform feature extraction to estimate the target translation parameters.

[0053] The deep network, namely the LSTM policy network, is used to roughly estimate the prediction distance parameters, avoiding the problem that the parameterized initial state setting cannot adapt to the environment, while improving the overall computing efficiency.

[0054] S103: Acquire target translation parameters according to a plurality of first range grid offset estimation values ​​and a resolution of the range dimension.

[0055] Among them, the target translation parameters include target velocity, target acceleration and target jerk.

[0056] Specifically, obtaining the target translation parameter according to the multiple first range grid offset estimation values ​​and the resolution of the range dimension includes:

[0057] Step C1: multiplying a plurality of first distance grid offset estimation values ​​and a resolution of the distance dimension to obtain a plurality of second distance grid offset estimation values.

[0058] The plurality of second range grid offset estimates is a coarse estimate.

[0059] Step C2: Fitting multiple second distance grid offset estimation values ​​to obtain corresponding fitting curves.

[0060] Step C3: Obtain the target translation parameters by fitting the curve.

[0061] By fitting the curve, we can obtain the target velocity v, target acceleration a and target jerk a. r .

[0062] S104. Estimate the target translation parameters using the Adam optimizer to obtain optimal target translation parameters.

[0063] In addition to the Adam optimizer, the present invention may also use other gradient descent optimizers, such as a root mean square propagation (RMSprop) optimizer, an adaptive gradient (Adagrad) optimizer, and the like.

[0064] Specifically, the Adam optimizer is used to estimate the target translation parameters to obtain the optimal target translation parameters, including:

[0065] Step D1: According to the target translation parameters, translation compensation and imaging are performed on the echo signals after multiple pulse compression to obtain the imaging results of the current round.

[0066] Specifically, according to the target translation parameters, translation compensation and imaging are performed on the echo signals after multiple pulse compression to obtain the imaging results of the current round, including:

[0067] Step D11: Set the target translation parameters (v, a, a r ) as the learning parameters of the imaging network to obtain the imaging parameter estimation value

[0068] Step D12: constructing a first translation compensation term corresponding to a plurality of pulse compressed echo signals according to the imaging parameter estimation values.

[0069] Step D13: Compensating the echo signals after the plurality of pulse compressions according to the first translation compensation term to obtain a plurality of compensated echo signals.

[0070] Step D14: coherently accumulate multiple compensated echo signals to obtain the imaging result of the current round.

[0071] Step D2: Determine the corresponding loss function of the current round according to the imaging result of the current round.

[0072] The loss function of the current round is selected as the image entropy of the imaging result of the current round.

[0073] In addition to the image entropy of the current round of imaging results, the loss function can also be other image indicators, such as mean, sharpness, etc.

[0074] Specifically, the expression of the loss function of the current round is:

[0075]

[0076] Among them, Entropy(·) is the loss function of the current round, that is, the image entropy of the imaging result of the current round, is the average energy value of the imaging result of the current round, I(y,x) is the imaging result of the current round, X is the image azimuth length, Y is the image distance length, x is the image azimuth index, and y is the image distance index.

[0077] Step D3: Based on the target translation parameters and the loss function of the current round, the Adam optimizer is used to solve the target translation parameters of the next round.

[0078] Specifically, according to the target translation parameters and the loss function of the current round, the Adam optimizer is used to solve the target translation parameters of the next round, including:

[0079] Step D31: Set the target translation parameters (v, a, a r ) as the optimization parameter of the current round And the first-order moment m of the Adam optimizer l and the second moment v l Initialize.

[0080] The first-order moment m of the Adam optimizer l and the second moment v l Initialized to 0.

[0081] Step D32: Calculate the gradient of the loss function of the current round according to the loss function of the current round and the optimization parameters of the current round.

[0082] The gradient expression of the loss function of the current round is:

[0083]

[0084] Among them, g l is the gradient of the loss function of the current round, Entropy(I(y,x)) is the loss function of the current round, is the optimization parameter of the current round, I(y,x) is the imaging result of the current round, x is the image azimuth index, and y is the image distance index.

[0085] Step D33: Update the initialized first-order moment according to the gradient, the initialized first-order moment and the first attenuation rate to obtain an updated first-order moment.

[0086] The expression of the updated first-order moment is:

[0087]

[0088] Among them, m l ′ is the updated first-order moment, is the first decay rate, m l is the first-order moment after initialization, g l is the gradient.

[0089] First decay rate To control the decay rate of the first-order moment after initialization, the first decay rate β1 l is close to 1, for example,

[0090] Step D34: Update the initialized second-order moment according to the gradient, the initialized second-order moment and the second attenuation rate to obtain an updated second-order moment.

[0091] The expression of the updated second-order moment is:

[0092]

[0093] Among them, v l ′ is the updated second-order moment, is the second decay rate, v l is the second-order moment after initialization, g l is the gradient.

[0094] Second decay rate To control the decay rate of the second-order moment after initialization, the second decay rate is close to 1, for example,

[0095] Step D35: Perform deviation correction on the updated first-order moment and the updated second-order moment to obtain a corrected first-order moment and a corrected second-order moment.

[0096] The expression of the corrected first-order moment is:

[0097]

[0098] in, is the corrected first-order moment, m l ′ is the updated first-order moment, is the first decay rate.

[0099] The expression of the corrected second-order moment is:

[0100]

[0101] in, is the corrected second-order moment, is the updated second-order moment, β2 l is the second decay rate.

[0102] Step D36: According to the optimization parameters of the current round, the corrected first-order moment, the corrected second-order moment and the learning rate, the optimization parameters of the next round are obtained, and the optimization parameters of the next round are the target translation parameters of the next round.

[0103] The expression of the optimization parameter for the next round is:

[0104]

[0105] in, is the optimization parameter for the next round, is the optimization parameter of the current round, is the corrected first-order moment, is the corrected second-order moment, η is the learning rate, and ε is a non-zero real number, which is a very small number, usually set to 10 -8 .

[0106] Step D4: Determine whether the difference between the loss function of the current round and the loss function of the previous round is less than a preset value. If so, the loss function of the current round converges, and the target translation parameters of the next round are determined as the optimal target translation parameters; if not, the target translation parameters of the next round are used as the target translation parameters, and the step of performing translation compensation and imaging on the echo signals after multiple pulse compression according to the target translation parameters is returned to obtain the imaging results of the current round, and the iteration is stopped until the difference between the loss function of the next round and the loss function of the current round is less than the preset value.

[0107] The default value can be 10 -9 .

[0108] Specifically, determine whether the difference between the current round's loss function and the previous round's loss function is less than 10 -9 If yes, the loss function of the current round converges, indicating that the optimal target translation parameters have been obtained, and the target translation parameters of the next round are determined as the optimal target translation parameters. If no, the target translation parameters of the next round are used as the target translation parameters, and the process returns to step D1 until the difference between the loss function of the next round and the loss function of the current round is less than 10 -9 The iteration is stopped when , and the optimal target translation parameters are obtained.

[0109] The Adam optimizer is used to achieve a more stable solution to non-convex problems, and then the target translation parameters are accurately estimated to eliminate the influence of the target translation component on imaging.

[0110] S105 . According to the optimal target translation parameters, compensation and coherent accumulation are performed on the echo signals after multiple pulse compression to obtain a focused image.

[0111] Specifically, according to the optimal target translation parameters, the echo signals after multiple pulse compression are compensated and coherently accumulated to obtain a focused image, including:

[0112] According to the optimal target translation parameter, a second translation compensation term corresponding to the echo signals after multiple pulse compression is constructed; according to the second translation compensation term, the echo signals after multiple pulse compression are compensated to obtain multiple final compensated echo signals; the multiple final compensated echo signals are coherently accumulated to perform two-dimensional imaging to obtain a focused image.

[0113] The target translation parameter estimation process aims at the loss function of the final imaging result, namely the image entropy. The translation compensation term is directly constructed to jointly compensate the envelope error and phase error caused by the target translation. This avoids the transmission of different target translation parameter estimation errors in the cascade compensation method, which leads to the error accumulation of the final compensation result, and improves the focusing degree and image calibration accuracy of the imaging result of the focused image. The Adam optimizer is used for non-convex optimization problems. The convergence speed is better than other gradient descent algorithms, and the target translation parameter update is not affected by the scaling transformation of the gradient. The learning rate can be automatically adjusted, which is suitable for unstable objective functions. At the same time, using image entropy as the optimization objective function can simultaneously obtain the energy gain brought by the two-dimensional coherent accumulation of distance and azimuth, ensuring that the algorithm has high noise robustness.

[0114] Figure 2 The optical image of the experimental target is schematically shown. Figure 3 The following diagram schematically shows the comparison between the imaging results of the method of the present invention and the imaging results of the traditional method under different signal-to-noise ratios. Figure 3 Yes Figure 2 For the results of the imaging, see Figure 3 As shown, Figure 3 The horizontal coordinates of all the pictures in the figure are the azimuth unit numbers, and the vertical coordinates are the distance unit numbers. The two pictures in the first row are respectively the imaging results of the traditional method and the imaging results of the method of the present invention under the condition of 15dB signal-to-noise ratio, the two pictures in the second row are respectively the imaging results of the traditional method and the imaging results of the method of the present invention under the condition of 10dB signal-to-noise ratio, the two pictures in the third row are respectively the imaging results of the traditional method and the imaging results of the method of the present invention under the condition of 5dB signal-to-noise ratio, and the two pictures in the last row are respectively the imaging results of the traditional method and the imaging results of the method of the present invention under the condition of 0dB signal-to-noise ratio. Figure 3 For electromagnetic data, no matter under any signal-to-noise ratio conditions, the imaging results of the traditional method have obvious defocusing, while the method of the present invention can ultimately estimate the corresponding optimal parameters according to different test data, and then obtain better focused imaging results, which verifies the effectiveness and noise resistance of the method of the present invention.

[0115] Figure 4 The loss function is schematically shown as a function of the signal-to-noise ratio. Figure 4As shown, the horizontal axis is the signal-to-noise ratio, the vertical axis is the image entropy value, that is, the value of the loss function, the blue line is the image entropy value of the imaging result of the traditional method, and the orange line is the image entropy value of the imaging result of the method of the present invention. It can be seen intuitively that when the signal-to-noise ratio is the same, the image entropy value of the imaging result of the method of the present invention is significantly lower than the image entropy value of the imaging result of the traditional method, which means that the image focusing effect of the method of the present invention is better and the algorithm compensation ability is more accurate.

[0116] Based on the above Figure 1 It can be seen from the implementation method that the ISAR coherent translation compensation and imaging method based on the deep network in the embodiment of the present invention includes pulse compression of ISAR echoes to obtain multiple pulse compressed echo signals; inputting the multiple pulse compressed echo signals into the trained LSTM strategy network in sequence, so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values; obtaining target translation parameters according to the multiple first distance grid offset estimation values ​​and the resolution of the distance dimension; estimating the target translation parameters using the Adam optimizer to obtain the optimal target translation parameters; according to the optimal target translation parameters, compensating and coherently accumulating the multiple pulse compressed echo signals to obtain a focused image. In this way, the Adam optimizer is used to achieve a more stable solution to non-convex problems, the target translation parameters can be accurately estimated, the influence of the target translation component on imaging can be eliminated, and the imaging result of the focused image obtained is clearer and more precise, so that the accuracy of the imaging result of the focused image is higher; using the optimal target translation parameters, the echo signals after multiple pulse compressions are directly compensated and coherently accumulated, which avoids the transmission of different target translation parameter estimation errors in the cascade compensation method, resulting in error accumulation in the final compensation result, and the compensation accuracy of the echo signals after multiple pulse compressions is higher, thereby improving the focusing degree and image calibration accuracy of the imaging result of the focused image, so that the accuracy of the imaging result of the image is higher.

[0117] Based on the same inventive concept, as an implementation of the above-mentioned ISAR coherent translational compensation and imaging method based on a deep network, an embodiment of the present invention also provides an ISAR coherent translational compensation and imaging device based on a deep network. Figure 5 FIG. 1 is a structural diagram of an ISAR coherent translation compensation and imaging device based on a deep network in an embodiment of the present invention, see Figure 5 As shown, the device may include:

[0118] A pulse compression module 501 is used to perform pulse compression on the ISAR echo to obtain a plurality of pulse compressed echo signals;

[0119] An input module 502 is used to sequentially input the echo signals after multiple pulse compression into a trained LSTM strategy network, so that the trained LSTM strategy network outputs multiple first distance grid offset estimation values, wherein the trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set with the same signal form as the echo signals after multiple pulse compression;

[0120] An acquisition module 503 is used to acquire target translation parameters according to a plurality of first range grid offset estimation values ​​and a resolution of the range dimension;

[0121] An estimation module 504 is used to estimate the target translation parameters using an Adam optimizer to obtain the optimal target translation parameters;

[0122] The compensation and coherent accumulation module 505 is used to compensate and coherently accumulate echo signals after multiple pulse compression according to the optimal target translation parameters to obtain a focused image.

[0123] The pulse compression module 501 is specifically used to divide the ISAR echo into multiple sub-apertures; perform range compression processing on the multiple sub-apertures to obtain corresponding one-dimensional range images, and transform the one-dimensional range images into the frequency domain to obtain multiple frequency domain signals; construct compensation factors of residual video items and envelope tilt items corresponding to the multiple frequency domain signals according to the pulse compression method; use the compensation factors of the residual video items and envelope tilt items to perform phase compensation on the peak frequency of each target scattering point in the multiple frequency domain signals to obtain multiple echo signals after pulse compression.

[0124] The acquisition module 503 is specifically used to multiply the multiple first distance grid offset estimation values ​​and the resolution of the distance dimension to obtain multiple second distance grid offset estimation values; fit the multiple second distance grid offset estimation values ​​to obtain corresponding fitting curves; and obtain the target translation parameters through the fitting curve.

[0125] The estimation module 504 is specifically used to perform translation compensation and imaging on the echo signals after multiple pulse compression according to the target translation parameters to obtain the imaging results of the current round; determine the corresponding loss function of the current round according to the imaging results of the current round; solve the target translation parameters of the next round by using the Adam optimizer according to the target translation parameters and the loss function of the current round; determine whether the difference between the loss function of the current round and the loss function of the previous round is less than a preset value, if so, the loss function of the current round converges, and the target translation parameters of the next round are determined as the optimal target translation parameters; if not, the target translation parameters of the next round are used as the target translation parameters, and return to the step of performing translation compensation and imaging on the echo signals after multiple pulse compression according to the target translation parameters to obtain the imaging results of the current round, until the difference between the loss function of the next round and the loss function of the current round is less than the preset value, and the iteration is stopped.

[0126] The estimation module 504 performs translation compensation and imaging on the echo signals after multiple pulse compression according to the target translation parameters to obtain the imaging results of the current round, including: taking the target translation parameters as the learning parameters of the imaging network to obtain the imaging parameter estimation value; constructing the first translation compensation term corresponding to the echo signals after multiple pulse compression according to the imaging parameter estimation value; compensating the echo signals after multiple pulse compression according to the first translation compensation term to obtain multiple compensated echo signals; and coherently accumulating the multiple compensated echo signals to obtain the imaging results of the current round.

[0127] In the estimation module 504, the expression of the loss function of the current round is:

[0128]

[0129] Where Entropy(·) is the loss function of the current round, I(y,x) is the average energy value of the imaging result of the current round, I(y,x) is the imaging result of the current round, X is the image azimuth length, Y is the image distance length, x is the image azimuth index, and y is the image distance index.

[0130] The estimation module 504 uses the Adam optimizer to solve the target translation parameters of the next round according to the target translation parameters and the loss function of the current round, including: taking the target translation parameters as the optimization parameters of the current round, and initializing the first-order moments and second-order moments of the Adam optimizer; calculating the gradient of the loss function of the current round according to the loss function of the current round and the optimization parameters of the current round; updating the initialized first-order moment according to the gradient, the initialized first-order moment and the first decay rate to obtain an updated first-order moment; updating the initialized second-order moment according to the gradient, the initialized second-order moment and the second decay rate to obtain an updated second-order moment; performing deviation correction on the updated first-order moment and the updated second-order moment to obtain a corrected first-order moment and a corrected second-order moment; obtaining the optimization parameters of the next round according to the optimization parameters of the current round, the corrected first-order moment, the corrected second-order moment and the learning rate, and the optimization parameters of the next round are the target translation parameters of the next round.

[0131] In acquisition module 503, target translation parameters include target velocity, target acceleration and target jerk.

[0132] It should be pointed out here that: the above description of the embodiment of the ISAR coherent translation compensation and imaging device based on a deep network is similar to the description of the embodiment of the ISAR coherent translation compensation and imaging method based on a deep network, and has similar beneficial effects as the embodiment of the ISAR coherent translation compensation and imaging method based on a deep network. For technical details not disclosed in the embodiment of the ISAR coherent translation compensation and imaging device based on a deep network of the embodiment of the present invention, please refer to the description of the embodiment of the ISAR coherent translation compensation and imaging method based on a deep network of the present invention for understanding.

[0133] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A deep network-based ISAR coherent translation compensation and imaging method, characterized in that: include: Perform pulse compression on the ISAR echo to obtain multiple pulse compressed echo signals; Inputting the echo signals after the compression of the multiple pulses into the trained LSTM strategy network in sequence, so that the trained LSTM strategy network outputs a plurality of first distance grid offset estimation values, wherein the trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set having the same signal form as the echo signals after the compression of the multiple pulses; Acquire target translation parameters according to the plurality of first range grid offset estimation values ​​and the resolution of the range dimension; The target translation parameters are estimated using the Adam optimizer to obtain the optimal target translation parameters; According to the optimal target translation parameters, the echo signals after the compression of the multiple pulses are compensated and coherently accumulated to obtain a focused image.

2. The ISAR coherent translation compensation and imaging method based on deep network according to claim 1, characterized in that: The pulse compression of the ISAR echo to obtain a plurality of pulse compressed echo signals includes: Dividing the ISAR echo into a plurality of sub-apertures; Performing range compression processing on the multiple sub-apertures to obtain corresponding one-dimensional range images, and transforming the one-dimensional range images into a frequency domain to obtain multiple frequency domain signals; constructing compensation factors of residual video items and envelope ramp items corresponding to the plurality of frequency domain signals according to a pulse compression method; Phase compensation is performed on the peak frequency of each target scattering point in the multiple frequency domain signals using the compensation factors of the residual video item and the envelope ramp item to obtain the multiple pulse compressed echo signals.

3. The ISAR coherent translation compensation and imaging method based on deep network according to claim 1, characterized in that: The method of training the LSTM strategy network using the training set includes: Using the predicted distance parameters in the training set as the initial state of the LSTM strategy network, the predicted distance parameters are parameters including predicted speed, predicted acceleration and predicted jerk; According to the initial state, feeding back the training action of the LSTM strategy network; Acquiring an actual state according to actual distance parameters of the echo signals after the plurality of pulse compressions; Calculating the difference loss between the initial state and the actual state; According to the training action, the gap loss and the learning rate, the initial state is updated, and the step of feeding back the training action of the LSTM strategy network according to the initial state is returned until the total number of training times reaches a preset total number of times, so as to obtain the trained LSTM strategy network.

4. The ISAR coherent translation compensation and imaging method based on deep network according to claim 1, characterized in that: The step of acquiring the target translation parameter according to the plurality of first distance grid offset estimation values ​​and the resolution of the distance dimension comprises: Multiplying the plurality of first distance grid offset estimation values ​​and the resolution of the distance dimension to obtain a plurality of second distance grid offset estimation values; Fitting the plurality of second distance grid offset estimation values ​​to obtain a corresponding fitting curve; The target translation parameters are obtained through the fitting curve.

5. The ISAR coherent translation compensation and imaging method based on deep network according to claim 1, characterized in that: The method of estimating the target translation parameters by using the Adam optimizer to obtain the optimal target translation parameters includes: According to the target translation parameters, performing translation compensation and imaging on the echo signals after the plurality of pulse compressions to obtain the imaging results of the current round; Determine the corresponding loss function of the current round according to the imaging result of the current round; According to the target translation parameter and the loss function of the current round, using the Adam optimizer to solve the target translation parameter of the next round; Determine whether the difference between the loss function of the current round and the loss function of the previous round is less than a preset value. If so, the loss function of the current round converges, and the target translation parameters of the next round are determined as the optimal target translation parameters. If not, the target translation parameters of the next round are used as the target translation parameters, and return to the step of performing translation compensation and imaging on the echo signals after compression of the multiple pulses according to the target translation parameters to obtain the imaging results of the current round, and stop iteration until the difference between the loss function of the next round and the loss function of the current round is less than the preset value.

6. The ISAR coherent translation compensation and imaging method based on deep network according to claim 5, characterized in that: The step of performing translation compensation and imaging on the echo signals after the plurality of pulse compressions according to the target translation parameters to obtain the imaging results of the current round includes: Using the target translation parameter as a learning parameter of an imaging network to obtain an imaging parameter estimation value; constructing a first translation compensation term corresponding to the echo signals after the plurality of pulse compressions according to the imaging parameter estimation value; Compensating the echo signals after the plurality of pulse compressions according to the first translation compensation item to obtain a plurality of compensated echo signals; The multiple compensated echo signals are coherently accumulated to obtain the imaging result of the current round.

7. The ISAR coherent translation compensation and imaging method based on deep network according to claim 5, characterized in that: The expression of the loss function of the current round is: Wherein, Entropy(·) is the loss function of the current round, is the average energy value of the imaging result of the current round, I(y,x) is the imaging result of the current round, X is the image azimuth length, Y is the image distance length, x is the image azimuth index, and y is the image distance index.

8. The ISAR coherent translation compensation and imaging method based on deep network according to claim 5, characterized in that: The step of solving the target translation parameter of the next round by using the Adam optimizer according to the target translation parameter and the loss function of the current round includes: The target translation parameter is used as the optimization parameter of the current round, and the first-order moment and the second-order moment of the Adam optimizer are initialized; Calculate the gradient of the loss function of the current round according to the loss function of the current round and the optimization parameters of the current round; According to the gradient, the initialized first-order moment and the first attenuation rate, updating the initialized first-order moment to obtain an updated first-order moment; According to the gradient, the initialized second-order moment and the second attenuation rate, the initialized second-order moment is updated to obtain an updated second-order moment; Performing deviation correction on the updated first-order moment and the updated second-order moment to obtain a corrected first-order moment and a corrected second-order moment; According to the optimization parameters of the current round, the corrected first-order moment, the corrected second-order moment and the learning rate, the optimization parameters of the next round are obtained, and the optimization parameters of the next round are the target translation parameters of the next round.

9. The ISAR coherent translation compensation and imaging method based on deep network according to claim 1, characterized in that: The target translation parameters include target velocity, target acceleration and target jerk.

10. An ISAR coherent translation compensation and imaging device based on a deep network, characterized in that: include: A pulse compression module is used to perform pulse compression on the ISAR echo to obtain multiple pulse compressed echo signals; An input module, used for sequentially inputting the echo signals after the plurality of pulse compressions into a trained LSTM strategy network, so that the trained LSTM strategy network outputs a plurality of first distance grid offset estimation values, wherein the trained LSTM strategy network is a network obtained by training the LSTM strategy network using a training set, and the training set is a data set having the same signal form as the echo signals after the plurality of pulse compressions; An acquisition module, configured to acquire target translation parameters according to the plurality of first distance grid offset estimation values ​​and the resolution of the distance dimension; An estimation module, used for estimating the target translation parameters using an Adam optimizer to obtain optimal target translation parameters; The compensation and coherent accumulation module is used to compensate and coherently accumulate the echo signals after the plurality of pulse compressions according to the optimal target translation parameters to obtain a focused image.

Images