A radar cross section test method based on deep learning and imaging inversion

Abstract

The present application relates to a kind of radar scattering cross section test methods based on deep learning and ISAR imaging inversion, the method includes using ISAR imaging test mode to obtain target and calibration body broadband radar scattering echo signal, calibration processing and applying filter-inverse projection imaging algorithm to carry out ISAR imaging, application completes the deep learning model of accurate extraction ISAR image scattering signal area, application FFT transform-interpolation integral ISAR imaging inversion algorithm high-precision inversion reconstruction RCS value, change ISAR imaging aperture center azimuth angle traversal acquisition target all azimuth angle RCS data.The radar scattering cross section test method based on deep learning and ISAR imaging inversion provided by the present application solves the accurate extraction ISAR image scattering area and ISAR image high-precision inversion reconstruction RCS and other test practical problems, expands RCS data acquisition method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of target radar scattering characteristic testing, in particular to an inverse synthetic aperture radar (ISAR) image target radar cross section (RCS) extraction technology based on deep learning and imaging inversion. BACKGROUND

[0002] The main outdoor field target static radar scattering characteristic testing methods include point frequency testing and step frequency testing. The point frequency testing is the main way to obtain RCS data, which has the advantages of utilizing the random distribution characteristics of the background environment noise and the receiver thermal noise, and improving the signal-to-noise ratio to suppress the background environment noise through signal coherence accumulation. However, the data processing method is insufficient, and it is difficult to perform fine signal processing.

[0003] The step frequency testing obtains distance direction high resolution imaging by transmitting a wideband radar signal, and realizes azimuth direction high resolution imaging by utilizing the Doppler effect generated by the target relative to the radar rotation. Finally, the ISAR image is formed through two-dimensional imaging algorithm processing. The step frequency testing is mainly used for ISAR imaging, and realizes strong scattering point diagnosis. The advantages are that more abundant wideband echo information can be utilized to suppress background clutter through time-frequency domain signal processing technology. However, the signal accumulation is insufficient, which leads to the decrease of signal-to-noise ratio and the reduction of testing accuracy.

[0004] The ISAR imaging inversion can obtain the advantages of point frequency testing and step frequency testing. Firstly, the integral operation is adopted in the imaging process, which can obtain the advantage of point frequency coherence accumulation. Secondly, the rich wideband information of step frequency testing can be utilized to perform fine data processing and obtain the advantage of testing accuracy. The main technical difficulties of ISAR imaging inversion for obtaining target RCS are the accurate extraction of ISAR image scattering area and the realization of high-precision inversion reconstruction RCS algorithm. The existing CLEAN algorithm and traditional image segmentation algorithm can extract the ISAR image scattering center or area, but the segmentation effect is not ideal due to the influence of factors such as target diversity, scattering center complexity, and background noise clutter. The existing ISAR image inversion reconstruction RCS algorithm mainly adopts two-dimensional FFT transformation and coordinate interpolation method, and the interpolation in the azimuth and frequency directions will bring errors and cause the loss of inversion accuracy. SUMMARY

[0005] The purpose of the present application is to solve the problem of accurate extraction of radar image target scattering signal in RCS testing, and to provide a radar scattering cross section testing method based on deep learning and imaging inversion.

[0006] The technical scheme of the present application is as follows:

[0007] A radar scattering cross section testing method based on deep learning and imaging inversion, comprising the following steps:

[0008] S1, obtaining radar wideband echo signals of a target and a calibration body through a wideband ISAR imaging test mode; wherein the target rotates in a plane at a uniform speed, and a wideband measurement frequency point and an azimuth angle are recorded synchronously;

[0009] S2, performing calibration processing on the echo signals of the target and the calibration body, setting imaging parameters, and generating a target ISAR image by using a filtering-inverse projection algorithm; the imaging parameters include an aperture center azimuth angle, an imaging aperture angle, a transverse distance, a longitudinal distance, and an imaging point number;

[0010] S3, performing segmentation on the target ISAR image obtained in S2 by using a deep learning model that has been trained, and extracting a target scattering area;

[0011] S4, performing inversion reconstruction on the target scattering area, and solving a target radar scattering cross section;

[0012] S5, repeatedly performing steps S2-S4 to obtain radar scattering cross section data of the target under different aperture center azimuth angles, and forming an RCS-azimuth data result by traversing the aperture center azimuth angle.

[0013] Further, the wideband ISAR imaging test in S1 is performed on the target rotating in a plane at a uniform speed, and the azimuth angle sampling interval The following requirements should be met:

[0014] (1)

[0015] wherein, is the highest frequency of the wideband ISAR imaging test, D is the maximum size of the target, is the speed of light.

[0016] Further, S2 is specifically:

[0017] S21, performing calibration processing according to the echo signals of the target and the calibration body, and calculating a complex radar scattering cross section of the target by using the following formula .

[0018] (2)

[0019] wherein, denotes the complex radar scattering cross section of the target after calibration processing, and the subscript denotes the target, and the subscript denotes the calibration body, and the superscript 1 denotes the complex radar scattering cross section after calibration processing, , and , These represent the distances between the target and the testing equipment, and between the calibration object and the testing equipment, respectively. For the theoretical complex radar cross section of the calibration body, For broadband measurement frequency points, The azimuth angle of rotation around the target;

[0020] S22. The calibrated signal is processed using a filtering-inverse projection algorithm to generate the target ISAR image. ;

[0021] (3)

[0022] (4)

[0023] in, For wave number, ; To represent a complex number; , The lowest and highest frequencies for broadband ISAR imaging testing. The speed of light; , for , The corresponding wave number; For projection lines, when the far-field condition is satisfied, ; The horizontal and vertical distances of the ISAR image; The starting point of the imaging aperture angle, The endpoint of the imaging aperture angle; This is an intermediate variable representing the azimuth angle. superior The projection value.

[0024] Furthermore, the training process of the deep learning model described in S3 is as follows:

[0025] S31. Using the same test position, for each type of target that has been tested, set the imaging parameters and generate ISAR image samples; label the scattering region and background clutter on the image samples and construct a training dataset.

[0026] S32. Adjust the resolution and augment the data of the training dataset;

[0027] S33. Input the enhanced dataset into the deep learning model for iterative training;

[0028] S34. Evaluate the model using the test dataset, analyze the intersection-over-union ratio and Dice coefficient, and verify the model's segmentation effect.

[0029] Furthermore, S3 specifically refers to:

[0030] The ISAR image samples are feature extracted by an encoder, and each layer is processed by a residual module and max pooling.

[0031] Feature upsampling is performed through a decoder, with each layer passing through a residual module and an upsampling layer;

[0032] A skip connection is used to connect shallow features of the encoder and deep features of the decoder in a cascaded encoder;

[0033] Feature redundancy is reduced and feature representation is enhanced by using channel attention modules and spatial attention modules;

[0034] A deep learning model incorporating residual modules and attention mechanisms is used for image segmentation to generate target scattering regions.

[0035] The residual module consists of a batch normalization layer, two convolutional layers, and two ReLU functions. After the input features undergo one normalization, two convolutions, and two activation functions, they are finally summed with the input features and output, as shown in the following formula:

[0036] (5)

[0037] in, This represents the feature map output by the residual module. Indicates the input feature map, Indicates batch normalization. This represents the convolution operation. express Activation function.

[0038] The attention mechanism is a Convolutional Block Attention Module (CBAM), which includes a channel attention module and a spatial attention module. Its output feature map is the result of multiplying the input feature map by the output feature maps of the channel attention module and the spatial attention module, as shown in the following equation.

[0039] (6)

[0040] in, This represents the feature map output by the CBAM attention module. Indicates the input feature map, This represents the feature map output by the channel attention module. This represents the feature map output by the spatial attention module. This indicates element-wise multiplication.

[0041] The channel attention module applies global average pooling and max pooling to the input feature map, and then performs summation using a multilayer perceptron (MLP). The activation function generates the output feature map of the channel attention module, as shown in the following equation:

[0042] (7)

[0043] in, This represents the activation function. This represents a multilayer perceptron. Indicates feature splicing, Indicates global average pooling. This indicates global max pooling.

[0044] The spatial attention module takes the result of element-wise multiplication of the input and output feature maps of the channel attention module as its input feature map. Global average pooling and max pooling are applied to the input feature map, and the number of channels is reduced to 1 through convolution. The sigmoid function is then used to generate the output feature map of the spatial attention module, as shown in the following equation:

[0045] (8)

[0046] The deep learning model with residual module and attention mechanism uses a combined loss function of Tversky loss and BCE loss to calculate the difference between predicted and true values, optimizing model training, as shown in the following equation: (9)

[0047] in, This indicates the number of imaging points in the ISAR image. This represents the total number of imaging points in an ISAR image. Indicates the training sample number. Represents the number of training samples. Indicates the predicted value. Represents the true value. Indicates weight, and These represent the penalty values ​​for false negatives and false positives, respectively, and are set to 0.5.

[0048] Furthermore, S4 specifically refers to:

[0049] S41. The target scattering region is inverted and reconstructed using Fourier transform and interpolation integration algorithms to obtain the target complex radar cross section. ;

[0050] (10)

[0051] (11)

[0052] in, This represents the complex radar cross section of the target after imaging inversion and reconstruction. The superscript 2 indicates the complex radar cross section after inversion and reconstruction. Represents the target ISAR image. To represent a complex number; , For the frequency to be inverted, At the speed of light, The azimuth angle of rotation around the target; The horizontal and vertical distances of the ISAR image; , The minimum and maximum horizontal distances of the ISAR image. , The minimum and maximum longitudinal distances of the ISAR image; As an intermediate variable, representing the horizontal distance of the ISAR image. The result of performing a one-dimensional Fourier transform on the vertical data column;

[0053] S42. Define the ideal point target. ;

[0054]

[0055] in, The complex radar cross section (RCS) of an ideal point target is defined as 1, using the subscript "ones". Correspondingly, the S2 filtering-inverse projection algorithm is used to acquire ISAR images. ,and Correspondingly, the S41 inversion method is used to obtain the complex radar cross section of an ideal point target. ,and Correspondingly;

[0056] S43. The following formula is used to eliminate the influence of the inversion point spread function for calibration, and the final target radar cross section is obtained. ;

[0057]

[0058]

[0059] in, This represents the complex radar cross section of the target after inversion reconstruction and calibration. The superscript 3 indicates the target radar cross section after inversion, reconstruction, and calibration.

[0060] Furthermore, S5 specifically refers to:

[0061] S51. Within the azimuth range of the ISAR imaging test, divide the aperture center azimuth at equal intervals and traverse it.

[0062] S52. Generate an ISAR image for the center azimuth of each aperture;

[0063] S53. Repeat the segmentation and inversion reconstruction steps on the generated image;

[0064] S54. Summarize the radar cross section data under different aperture center azimuth angles to form a comprehensive data result.

[0065] The beneficial effects of this invention are:

[0066] The method of this invention achieves high-precision extraction of the scattering region of ISAR images through a deep learning model that includes residual modules and attention mechanisms. It combines the FFT transform-interpolation integral ISAR image high-precision RCS inversion and reconstruction algorithm to achieve refined signal processing for RCS testing, thereby obtaining accurate RCS data of the target.

[0067] The deep learning model proposed in this invention, which incorporates a residual module and an attention mechanism, introduces a residual module on the basis of the UNet model, which is beneficial for accurate extraction of scattering regions and improves segmentation accuracy. The fusion of the attention mechanism module adapts to complex and changing scenarios of scattering centers and improves the robustness of the algorithm. The use of Tversky loss and BCE loss functions improves detection stability and effectively suppresses background clutter.

[0068] The high-precision RCS inversion and reconstruction algorithm for ISAR images based on FFT transform-interpolation integral proposed in this invention can effectively suppress azimuth interpolation errors compared with traditional two-dimensional FFT transform methods. Practical verification shows that it can adapt to a wider range of imaging aperture angles and has higher accuracy. Attached Figure Description

[0069] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0070] Figure 1 A flowchart of a radar cross section testing method based on deep learning and imaging inversion according to an embodiment of the present invention is shown;

[0071] Figure 2 An ISAR imaging result diagram of an embodiment of the present invention is shown;

[0072] Figure 3 A schematic diagram of a deep learning model network structure including a residual module and an attention mechanism is shown in an embodiment of the present invention.

[0073] Figure 4 A schematic diagram of a residual module according to an embodiment of the present invention is shown;

[0074] Figure 5 A schematic diagram of the attention mechanism of an embodiment of the present invention is shown;

[0075] Figure 6 A schematic diagram of a channel attention module according to an embodiment of the present invention is shown;

[0076] Figure 7 A schematic diagram of a spatial attention module according to an embodiment of the present invention is shown;

[0077] Figure 8 An ISAR image segmentation result diagram of an embodiment of the present invention is shown;

[0078] Figure 9 A flowchart of an FFT transform-interpolation integral ISAR image inversion and reconstruction algorithm according to an embodiment of the present invention is shown;

[0079] Figure 10 This figure shows a comparison between S-band RCS data and spot frequency test data according to an embodiment of the present invention;

[0080] Figure 11 A comparison chart of X-band RCS data and point frequency test data according to an embodiment of the present invention is shown. Detailed Implementation

[0081] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0082] like Figure 1 As shown, this invention provides a radar cross section testing method based on deep learning and imaging inversion. The specific implementation steps of this method are as follows:

[0083] S1. Install the target to be measured on the target support turntable and the calibration body on the calibration support. Set the measurement mode to broadband ISAR imaging test, acquire the broadband radar echo signals of the target and the calibration body, and record the broadband measurement frequency and the target rotation azimuth angle accordingly. The calibration body can be measured simultaneously at a different location from the target or measured at different times at the same location.

[0084] The target rotates in a plane at a constant speed, with the azimuth angle sampling interval... The following requirements should be met:

[0085] (1)

[0086] in, This is the highest frequency for broadband ISAR imaging tests. D For the target maximum size, It is the speed of light.

[0087] S2. The echo signals of the target and the calibration body are calibrated, imaging parameters are set, and a filter-inverse projection algorithm is used to generate the target ISAR image; the imaging parameters include the aperture center azimuth angle, imaging aperture angle, lateral distance, longitudinal distance, and number of imaging points.

[0088] S21. Based on the echo signals from the target and the calibration body, perform calibration processing and calculate the target's complex radar cross section using the following formula. ;

[0089] (2)

[0090] in, Indicates the target's complex radar cross section after calibration, subscript Indicates the target, subscript This indicates the calibration body; the superscript 1 indicates the target's complex radar cross section after calibration. , These are the broadband scattered echo signals of the target and the calibration body, respectively. , These represent the distances between the target and the testing equipment, and between the calibration object and the testing equipment, respectively. For the theoretical complex radar cross section of the calibration body, For broadband measurement frequency points, The azimuth angle of rotation around the target;

[0091] S22. The calibrated signal is processed using a filtering-inverse projection algorithm to generate the target ISAR image. ;

[0092] (3)

[0093] (4)

[0094] in, For wave number, ; To represent a complex number; , The lowest and highest frequencies for broadband ISAR imaging testing. The speed of light; , for , The corresponding wave number; For projection lines, when the far-field condition is satisfied, ; The horizontal and vertical distances of the ISAR image; The starting point of the imaging aperture angle, The endpoint of the imaging aperture angle; This is an intermediate variable representing the azimuth angle. superior The projection value.

[0095] Figure 2 This is an ISAR image of a low-scattering carrier obtained using the filter-inverse projection algorithm. The ISAR image includes background clutter and a scattered signal region. This scattered signal region of the target exhibits a local scattering center in the high-frequency band. Local scattering centers generally have a wide visible angle range and are fixed in a small area. They are considered to be the most important scattering center feature in ISAR images.

[0096] S3. The target ISAR image obtained in S2 is segmented using a trained deep learning model to extract the target scattering region;

[0097] The training process of the deep learning model described in S3 is as follows:

[0098] S31. Using the same test position, for each type of target that has been tested, set the imaging parameters and generate ISAR image samples; label the scattering region and background clutter on the image samples and construct a training dataset.

[0099] S32. Adjust the resolution and augment the data of the training dataset;

[0100] S33. Input the enhanced dataset into the deep learning model for iterative training;

[0101] S34. Evaluate the model using the test dataset, analyze the intersection-over-union ratio and Dice coefficient, and verify the model's segmentation effect.

[0102] The deep learning model described in S3 is specifically:

[0103] like Figure 3 As shown, the deep learning model is a deep learning model based on residual modules and attention mechanisms. On the basis of the UNet deep learning model, residual modules and attention mechanisms are introduced, including two parts: the left side is the encoder, and each layer goes through a residual module and a max pooling layer; the right side is the decoder, and each layer goes through a residual module and an upsampling layer. In order to reduce the redundant information of shallow features, an attention mechanism is introduced after the left residual module, and skip connections are used to cascade the shallow features on the left and the deep features on the right.

[0104] The residual module, such as Figure 4 As shown, it consists of a batch normalization layer, two convolutional layers, and two ReLU functions. The input features are normalized once, convolved twice, and activated twice. Finally, the sum of the input features and the output is calculated, as shown in the following formula:

[0105] (5)

[0106] in, This represents the feature map output by the residual module. Indicates the input feature map, Indicates batch normalization. This represents the convolution operation. express Activation function.

[0107] The attention mechanism is a Convolutional Block Attention Module (CBAM), such as... Figure 5 As shown, it includes a channel attention module and a spatial attention module. Its output feature map is the result of multiplying the input feature map by the output feature maps of the channel attention module and the spatial attention module, as shown in the following formula:

[0108] (6)

[0109] in, This represents the feature map output by the CBAM attention module. Indicates the input feature map, This represents the feature map output by the channel attention module. This represents the feature map output by the spatial attention module. This indicates element-wise multiplication.

[0110] The channel attention module, such as Figure 6 As shown, global average pooling and max pooling are applied to the input feature maps, and then the sums are obtained through a multilayer perceptron (MLP). The output feature map of the channel attention module is generated using the sigmoid activation function, as shown in the following equation:

[0111] (7)

[0112] in, This represents the activation function. This represents a multilayer perceptron. Indicates feature splicing, Indicates global average pooling. This indicates global max pooling.

[0113] The spatial attention module, such as Figure 7 As shown, the result of element-wise multiplication of the input and output feature maps of the channel attention module is used as the input feature map. Global average pooling and max pooling are applied to the input feature map, and the number of channels is reduced to 1 through convolution. The function generates the output feature map of the spatial attention module, as shown in the following equation:

[0114] (8)

[0115] The deep learning model, which includes residual modules and attention mechanisms, uses a combined loss function of Tversky loss and BCE loss to calculate the difference between predicted and true values, optimizing model training, as shown in the following equation:

[0116] (9)

[0117] in, This indicates the number of imaging points in the ISAR image. This represents the total number of imaging points in an ISAR image. Indicates the training sample number. Represents the number of training samples. Indicates the predicted value. Represents the true value. Indicates weight, and These represent the penalty values ​​for false negatives and false positives, respectively, and are set to 0.5.

[0118] Figure 8 This is a segmentation result of an ISAR image using a deep learning model that includes residual modules and attention mechanisms, in which the target scattering center region is accurately extracted.

[0119] S4. Perform inversion and reconstruction on the target scattering region to calculate the target radar cross section;

[0120] S41. The target scattering region is inverted and reconstructed using Fourier transform and interpolation integration algorithms to obtain the target complex radar cross section. ;

[0121] (10)

[0122] (11)

[0123] in, This represents the complex radar cross section of the target after imaging inversion and reconstruction. The superscript 2 indicates the complex radar cross section after inversion and reconstruction. Represents the target ISAR image. To represent a complex number; ， For the frequency to be inverted, At the speed of light, The azimuth angle of rotation around the target; x, y The horizontal and vertical distances of the ISAR image; , The minimum and maximum horizontal distances of the ISAR image. , The minimum and maximum longitudinal distances of the ISAR image; As an intermediate variable, representing the horizontal distance of the ISAR image. The result of performing a one-dimensional Fourier transform on the vertical data column;

[0124] S42. Define the complex radar cross section of an ideal point target. ;

[0125] (12)

[0126] in, The complex radar cross section (RCS) of an ideal point target is defined as 1, using the subscript "ones". Correspondingly, the S2 filtering-inverse projection algorithm is used to acquire ISAR images. ,and Correspondingly, the S41 inversion method is used to obtain the complex radar cross section of an ideal point target. ,and Correspondingly;

[0127] S43. The following formula is used to eliminate the influence of the inversion point spread function for calibration, and the final target radar cross section is obtained. ;

[0128] (13)

[0129] (14)

[0130] in, This represents the complex radar cross section of the target after inversion reconstruction and calibration. The superscript 3 indicates the target radar cross section after inversion, reconstruction, and calibration.

[0131] Figure 9 This is a flowchart of the FFT transform-interpolation integral ISAR image inversion algorithm proposed in this invention, which includes six steps:

[0132] The first step is to sort the ISAR image by azimuth. Dewindowing eliminates the influence of window functions during the imaging process;

[0133] The second step is to examine the radial (column-wise) ISAR image column by column. y (Direction) Perform a one-dimensional FFT transformation to obtain , Points are recorded as ;

[0134] The third step is based on the inversion... Calculate the corresponding position ,right Rounding up , This refers to the imaging step size;

[0135] Step 4: Check column by column coordinate , Vector interpolation is performed to obtain and in Horizontal ( (Direction) Integrate according to formula (14), Indicates the number of imaging points;

[0136] (15)

[0137] The fifth step is to perform frequency-based windowing to eliminate the influence of the window function during the imaging process and obtain... ;

[0138] The sixth step is to perform image inversion on the ideal point target, eliminate the influence of the system point spread function, and calibrate the inversion results.

[0139] Figure 10 , 11 This is a comparison chart of S-band and X-band RCS data obtained by the method of this invention and RCS data obtained by point frequency testing. The imaging inversion results and point frequency results show a high degree of agreement, verifying the effectiveness of the method proposed in this invention.

[0140] S5. Traverse the aperture center azimuth angles and repeat steps S2-S4 to obtain radar cross section data of the target under different aperture center azimuth angles, forming RCS-azimuth data results.

[0141] S51. Within the azimuth range of the ISAR imaging test, divide the aperture center azimuth at equal intervals and traverse it. If the test azimuth range is -45° to 45°, the aperture center azimuth can be divided into 0°, ±10°, ±20°, ±30°, ±40°, and the aperture angle can be set to 10°.

[0142] S52. Generate an ISAR image for the center azimuth of each aperture;

[0143] S53. Repeat the segmentation and inversion reconstruction steps on the generated image;

[0144] S54. Summarize the radar cross section data under different aperture center azimuth angles to form a comprehensive data result.

[0145] The method of this invention provides a feasible approach for testing and practical problems such as accurate extraction of scattering regions in ISAR images and high-precision inversion and reconstruction of RCS in ISAR images. This method expands the application scenarios of ISAR imaging testing and can also be applied to the acquisition of RCS of target areas of interest and the acquisition of RCS of the whole machine by component stitching.

[0146] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A radar cross section testing method based on deep learning and imaging inversion, characterized in that, Includes the following steps: S1. Obtain the wideband radar echo signals of the target and the calibration body through the wideband ISAR imaging test method; wherein the target rotates in a plane at a constant speed, and the wideband measurement frequency and azimuth angle are recorded simultaneously. S2. The echo signals of the target and the calibration body are calibrated, imaging parameters are set, and a filter-inverse projection algorithm is used to generate the target ISAR image; the imaging parameters include the aperture center azimuth angle, imaging aperture angle, lateral distance, longitudinal distance, and number of imaging points. S3. The target ISAR image obtained in S2 is segmented using a trained deep learning model to extract the target scattering region; S4. Perform inversion and reconstruction on the target scattering region to calculate the target radar cross section; S5. Traverse the aperture center azimuth angles and repeat steps S2-S4 to obtain radar cross section data of the target under different aperture center azimuth angles, forming RCS-azimuth data results.

2. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that... In S1, the target rotates in a plane at a constant speed, and the azimuth angle sampling interval is... The following requirements should be met: （1） in, This is the highest frequency for broadband ISAR imaging tests. D For the target maximum size, It is the speed of light.

3. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that S2 include: S21. Based on the echo signals from the target and the calibration body, perform calibration processing and calculate the target's complex radar cross section using the following formula. ; （2） in, The term represents the complex radar cross section of the target after calibration. The subscript T represents the target, the subscript C represents the calibration body, and the superscript 1 represents the complex radar cross section of the target after calibration. , These are the broadband scattered echo signals of the target and the calibration body, respectively. , These represent the distances between the target and the testing equipment, and between the calibration object and the testing equipment, respectively. For the theoretical complex radar cross section of the calibration body, For broadband measurement frequency points, The azimuth angle of rotation around the target; S22. The calibrated signal is processed using a filtering-inverse projection algorithm to generate the target ISAR image. ; （3） （4） in, For wave number, ; To represent a complex number; 、 The lowest and highest frequencies for broadband ISAR imaging testing. The speed of light; , for 、 The corresponding wave number; For projection lines, when the far-field condition is satisfied, ; The horizontal and vertical distances of the ISAR image; The starting point of the imaging aperture angle, The endpoint of the imaging aperture angle; This is an intermediate variable representing the azimuth angle. superior The projection value.

4. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that... In S3, the training process of the deep learning model is as follows: S31. Using the same test position, for each type of target that has been tested, set the imaging parameters and generate ISAR image samples; label the scattering region and background clutter on the image samples and construct a training dataset. S32. Adjust the resolution and augment the data of the training dataset; S33. Input the enhanced dataset into the deep learning model for iterative training; S34. Evaluate the model using the test dataset, analyze the intersection-over-union ratio and Dice coefficient, and verify the model's segmentation effect.

5. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that S3 include: The ISAR image samples are feature extracted by an encoder, and each layer is processed by a residual module and max pooling. Feature upsampling is performed through a decoder, with each layer passing through a residual module and an upsampling layer; A skip connection is used to connect shallow features of the encoder and deep features of the decoder in a cascaded encoder; Feature redundancy is reduced and feature representation is enhanced by using channel attention modules and spatial attention modules; A deep learning model incorporating residual modules and attention mechanisms is used for image segmentation to generate target scattering regions.

6. The radar cross section testing method based on deep learning and imaging inversion as described in claim 5, characterized in that: The residual module includes a batch normalization layer, two convolutional layers, and two ReLU functions. After the input features undergo one normalization, two convolutions, and two activation functions, they are finally summed with the input features and output, as shown in the following formula: （5） in, This represents the feature map output by the residual module. Indicates the input feature map, Indicates batch normalization. This represents the convolution operation. express Activation function; The attention mechanism is a Convolutional Block Attention Module (CBAM), which includes a channel attention module and a spatial attention module. Its output feature map is the result of multiplying the input feature map by the output feature maps of the channel attention module and the spatial attention module, as shown in the following formula: （6） in, This represents the feature map output by the CBAM attention module. Indicates the input feature map, This represents the feature map output by the channel attention module. This represents the feature map output by the spatial attention module. This indicates element-wise multiplication; The channel attention module applies global average pooling and max pooling to the input feature map, and then performs summation using a multilayer perceptron (MLP). The activation function generates the output feature map of the channel attention module, as shown in the following equation: （7） in, This represents the activation function. This represents a multilayer perceptron. Indicates feature splicing, Indicates global average pooling. Indicates global max pooling; The spatial attention module takes the result of element-wise multiplication of the input and output feature maps of the channel attention module as its input feature map. Global average pooling and max pooling are applied to the input feature map, and a convolution operation is used to reduce the number of channels to 1. The sigmoid function is then used to generate the output feature map of the spatial attention module, as shown in the following equation: （8）。 7. The radar cross section testing method based on deep learning and imaging inversion as described in claim 5, characterized in that: The deep learning model with residual module and attention mechanism uses a combined loss function of Tversky loss and BCE loss to calculate the difference between predicted and true values, optimizing model training, as shown in the following equation: （9） in, This indicates the number of imaging points in the ISAR image. N This represents the total number of imaging points in an ISAR image. Indicates the training sample number. Indicates the number of training samples. Indicates the predicted value. Represents the actual value. Indicates weight, and These represent the penalty values ​​for false negatives and false positives, respectively, and are set to 0.

5.

8. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that S4 include: S41. The target scattering region is inverted and reconstructed using Fourier transform and interpolation integration algorithms to obtain the target complex radar cross section. ; （10） （11） in, This represents the complex radar cross section of the target after imaging inversion and reconstruction. The superscript 2 indicates the complex radar cross section after inversion and reconstruction. Represents the target ISAR image. To represent a complex number; , For the frequency to be inverted, At the speed of light, The azimuth angle of rotation around the target; The horizontal and vertical distances of the ISAR image; , The minimum and maximum horizontal distances of the ISAR image. , The minimum and maximum longitudinal distances of the ISAR image; As an intermediate variable, representing the horizontal distance of the ISAR image. The result of performing a one-dimensional Fourier transform on the vertical data column; S42. Define the complex radar cross section of an ideal point target. ; （12） in, The complex radar cross section (RCS) of an ideal point target is defined as 1, using the subscript "ones". Correspondingly, the S2 filtering-inverse projection algorithm is used to acquire ISAR images. ,and Correspondingly, the S41 inversion method is used to obtain the complex radar cross section of an ideal point target. ,and Correspondingly; S43. The following formula is used to eliminate the influence of the point spread function for inversion calibration, and the final target radar cross section is obtained. ; （13） （14） in, This represents the complex radar cross section of the target after inversion reconstruction and calibration. The superscript 3 indicates the target radar cross section after inversion, reconstruction, and calibration.

9. The radar cross section testing method based on deep learning and imaging inversion as described in claim 1, characterized in that S5 include: S51. Within the azimuth range of the ISAR imaging test, divide the aperture center azimuth at equal intervals and traverse it. S52. Generate an ISAR image for the center azimuth of each aperture; S53. Repeat the segmentation and inversion reconstruction steps on the generated image; S54. Summarize the radar cross section data under different aperture center azimuth angles to form a comprehensive data result.

Images