An isar imaging method for down-sampling data

Abstract

The application discloses an ISAR imaging method for down-sampling data, relates to the technical field of radar signal processing, and constructs an ISAR data set for training CSKNN; constructs the CSKNN for ISAR down-sampling data imaging; adopts a skip connection construction mode, and the skip connection refers to: replacing SK with a convolution layer to establish the connection between the network shallow feature representation and the network deep feature representation in the same level. Based on the ISAR data set, after a given loss function form, the parameters of the CSKNN are learned in combination with the back propagation and the gradient descent algorithm; when the training loss and the verification loss of the CSKNN are small enough and tend to be stable, the parameter updating of the CSKNN is stopped, and the trained CSKNN is obtained; and the CSKNN is used to realize the ISAR down-sampling data imaging. The application is superior to the OMP imaging method and the GKF imaging method in imaging quality and efficiency, and is superior to the FCNN imaging method in imaging quality. The target contour in the imaging result of the CSKNN is more complete, and the number of error reconstruction scattering points is reduced by 18% on average.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to an ISAR imaging method for downsampled data. Background Technology

[0002] Inverse Synthetic Aperture Radar (ISAR) can achieve high-resolution imaging of moving targets under all-weather, all-day, and long-range operating conditions. Traditional ISAR imaging methods are range-Doppler (RD) type imaging methods. RD type imaging methods have high imaging efficiency, but the imaging results are easily affected by sidelobe interference, and they are prone to failure when the imaging assumptions are not valid or the imaging model is mismatched.

[0003] Optimized ISAR imaging methods are a class of imaging methods constrained by the sparsity of the target scene to be reconstructed. They can overcome the shortcomings of RD-type methods and reduce the influence of sidelobe interference. Compressive / compressive Sensing (CS) based ISAR imaging methods are similar to optimized ISAR imaging methods, but CS imaging methods emphasize imaging using a very small amount of undersampled data.

[0004] Most ISAR imaging methods based on Convolutional Neural Networks (CNNs) focus on extracting deep abstract features from radar data for imaging, while neglecting shallow features. Full Convolutional Neural Networks (FCNNs), with their multi-level decomposition and multi-channel filtering architecture, can extract both deep and shallow features from input ISAR data at all scales, resulting in images with complete target contours and clear details. However, because shallow and deep features are directly connected using a series of nested skip connections (SK), semantic discrepancies inevitably arise between them, reducing the consistency of feature representations and limiting the performance improvement of FCNN-based ISAR imaging. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an ISAR imaging method for downsampled data. The present invention can reconstruct high-quality ISAR images. The imaging method proposed in the present invention produces more complete target contours and higher target image contrast in the imaging results.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] An ISAR imaging method for downsampled data proposed according to the present invention includes the following steps:

[0008] Step S1: Construct the ISAR dataset for training CSKNN, group the ISAR dataset to obtain the training set and validation set of CSKNN;

[0009] Step S2: Construct convolutional skip connections using convolutional layers, and use the constructed convolutional skip connections to construct CSKNN for ISAR downsampled data imaging; wherein, the construction method of convolutional skip connections refers to: replacing skip connections SK with convolutional layers, and establishing the connection between shallow feature representations and deep feature representations of the network in the same level.

[0010] Step S3: Based on the training set and validation set generated in step S1, and given the form of the loss function, the parameters of CSKNN are learned by combining backpropagation and gradient descent algorithms. When the training loss and validation loss of CSKNN are small enough and both tend to be stable, the CSKNN parameter update is stopped, and the trained CSKNN is obtained.

[0011] Step S4: Use the CSKNN obtained in step S3 to perform ISAR downsampling data imaging.

[0012] As a further optimization scheme for the ISAR imaging method for downsampling data described in this invention, in step S1, the electromagnetic scattering echo of the simulated target is generated using the electromagnetic simulation software FEKO to construct the ISAR dataset.

[0013] As a further optimization scheme for the ISAR imaging method for downsampled data described in this invention, in step S1, the ISAR dataset is grouped to obtain the training set and validation set of CSKNN; specifically as follows:

[0014] Construct the ISAR dataset for training CSKNN: Radar parameters are set as follows: the simulation model is illuminated by a plane wave. Under the far-field assumption, the reflected echo of the simulation model is calculated using physical optics and collected by receivers placed around the target model; the transmitted plane wave frequency band is 8.75 GHz to 9.25 GHz, with 100 frequency sampling points; the elevation angle of the radar target is 46° to 86°, with a total of 40 elevation angles; the azimuth angle ranges from 0.05° to 360.05°, with an azimuth accumulation angle of 5°, accumulated 72 times; resulting in an ISAR dataset of 2880 elements.

[0015] The ISAR dataset was divided into training and validation sets for CSKNN. Echo data at elevation angles of 50°, 60°, 70°, 80°, and 90° were assigned to the validation set, while the rest were assigned to the training set. The validation set contained 360 samples, and the training set contained 2520 samples. Each sample consisted of an initial input image and a label image. The label and initial images were obtained using the simulation data described above. The generation method involved randomly downsampling each echo data point in both the range and azimuth directions in two dimensions, padding the downsampling results with zeros, and then performing a two-dimensional FFT to obtain a low-quality initial image. For the same data set, the imaging result obtained directly through the RD method was used as the target image. One initial image and one target image constituted a training sample.

[0016] As a further optimization scheme for the ISAR imaging method for downsampled data described in this invention, in step S2, a CSKNN for ISAR downsampled data imaging is constructed by using convolutional layers, maximum pooling layers, BN layers, activation function layers, deconvolutional layers, residual learning mechanisms, and SK cascades.

[0017] As a further optimization scheme for the ISAR imaging method for downsampled data described in this invention, the process of constructing CSKNN is as follows: The input of CSKNN is the initial image obtained from downsampled data, and the output is the final imaging result; when constructing CSKNN, multiple feature data are first extracted using a 3×3 convolution kernel, and the deviation between the feature data and the input data is reduced using a BN layer and a ReLU activation function; the feature data is reduced in dimensionality using a 2×2 max pooling operation; in the data reconstruction process, the feature data is reconstructed using a deconvolution layer with a deconvolution kernel size of 2×2; the error between the reconstructed feature data and the reduced-dimensional feature data is reduced using a BN layer and a ReLU activation function, and a residual learning mechanism is added to the last layer of CSKNN, summing the initial image and the network-reconstructed feature data at the end of CSKNN to obtain the final ISAR imaging result.

[0018] As a further optimization scheme for the ISAR imaging method for downsampled data described in this invention, in step S3, a batch gradient descent training strategy is adopted, and the minimum mean square error function is used as the loss function:

[0019]

[0020] Where T represents the total number of training samples, and i represents the i-th training sample. This is the reconstruction result output by CSKNN. Let {W} represent the initial image in the i-th training sample, {W} represent the CSKNN parameter set, f(·) represent CSKNN, and σ iLet L({W}) represent the labeled image of the i-th sample, and let L({W}) represent the mean square error of the reconstruction.

[0021] As a further optimization scheme for the ISAR imaging method for downsampled data described in this invention, step S4 is as follows: the ISAR data generated by the electromagnetic simulation software FEKO is downsampled in the range and azimuth directions to obtain downsampled data. The downsampled data is then subjected to a two-dimensional Fourier transform to obtain an initial image. The obtained initial image is used as the input of CSKNN, and the output of CSKNN is the final imaging result.

[0022] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0023] (1) This invention introduces a novel skip connection construction method, replacing the traditional SK with a convolutional layer and reducing the number of channels in the convolutional layer, thereby narrowing the semantic gap between shallow and deep features and improving the imaging quality of ISAR contours and details.

[0024] (2) The IASR imaging method based on CSKNN was compared with the ISAR method based on FCNN, the ISAR method based on Greedy Kalman Filtering (GKF), and the ISAR imaging method based on Orthogonal Matching Pursuit (OMP). Quantitative indicators were used to evaluate the imaging results. Simulation data showed that the imaging method proposed in this invention produces more complete target contours and higher target image contrast. The CSKNN-based imaging method using downsampled data disclosed in this invention outperforms the OMP and GKF imaging methods in both imaging quality and efficiency, and outperforms the FCNN imaging method in terms of imaging quality. The CSKNN imaging results show more complete target contours and an average reduction of 18% in the number of incorrectly reconstructed scattering points. Attached Figure Description

[0025] Figure 1 This is a diagram of the CSKNN network structure.

[0026] Figure 2 These are the imaging results of CSKNN on the same unknown data with different training times; where (a) is 30,000 times, (b) is 40,000 times, (c) is 50,000 times, and (d) is 100,000 times.

[0027] Figure 3 The images are UAV data with a downsampling rate of 25% and the imaging results obtained by CSKNN, FCNN and compressed sensing imaging methods. Among them, (a) is the CSKNN imaging result, (b) is the FCNN imaging result, (c) is the GKF imaging result and (d) is the OMP imaging result.

[0028] Figure 4 The images are UAV data with a downsampling rate of 25% and the imaging results obtained by CSKNN, FCNN and compressed sensing imaging methods. Among them, (a) is the CSKNN imaging result, (b) is the FCNN imaging result, (c) is the GKF imaging result and (d) is the OMP imaging result.

[0029] Figure 5 This is a schematic diagram of the CSKNN imaging principle. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] To further extract sufficient deep and shallow abstract features from ISAR data and improve the imaging quality of ISAR contours and details, this invention introduces a novel skip connection construction method. It replaces the traditional SK with convolutional layers and reduces the number of channels in the convolutional layers, thus narrowing the semantic gap between shallow and deep features. The designed imaging network is called a Convolutional Skip Neural Network (CSKNN).

[0032] A quadcopter-based simulated UAV target was constructed using the electromagnetic simulation software FEKO, generating a large amount of simulated UAV target radar data with different imaging geometries. A large-scale simulated UAV target radar training dataset was established to update the CSKNN parameters, resulting in a parameter-optimized CSKNN. Based on the parameter-optimized CSKNN, an ISAR imaging method based on CSKNN was proposed. The input to the CSKNN is a low-quality initial image. The initial image is obtained by two-dimensional random downsampling of each echo data in the range and azimuth directions. The downsampling results are zero-padded and a two-dimensional FFT is performed to obtain a low-quality image. The output of the CSKNN is a high-quality imaging result.

[0033] The IASR imaging method based on CSKNN is compared with the ISAR method based on FCNN, the ISAR method based on Greedy Kalman Filtering (GKF), and the ISAR imaging method based on Orthogonal Matching Pursuit (OMP). Quantitative indicators are used to evaluate the imaging results. Simulation data imaging results show that the imaging method proposed in this invention produces more complete target contours and higher target image contrast.

[0034] This invention is based on the CSKNN structure, employing convolution and pooling operations to obtain feature representations of the data. It then reconstructs the feature data through deconvolution operations, replacing the traditional SK with convolutional layers and reducing the number of channels in the convolutional layers to improve abstract feature extraction capabilities. Feature connections are built between layers, ultimately reconstructing a high-quality ISAR image. The invention includes the following steps:

[0035] S1. Construct the ISAR dataset for training CSKNN. Use the electromagnetic simulation software FEKO to generate electromagnetic scattering echoes of simulated targets to construct the ISAR dataset. Divide the dataset into training and validation sets for CSKNN.

[0036] S2. A new skip connection construction method is adopted, which utilizes convolutional layers, max pooling layers, BN layers, activation function layers, deconvolutional layers, residual learning mechanisms, and SK cascades to construct CSKNN for ISAR downsampled data imaging. The new skip connection construction method refers to replacing skip connections SK with convolutional layers and establishing the connection between shallow feature representations and deep feature representations in the same level of the network.

[0037] S3. Based on the training and validation sets generated in step S1, and given the form of the loss function, learn the parameters of CSKNN using backpropagation and gradient descent algorithms. When the training and validation losses of the network are sufficiently small and both tend to stabilize, stop updating the network parameters. At this point, a CSKNN that meets the task requirements is obtained.

[0038] S4. ISAR downsampled data imaging using CSKNN. Downsampled data is obtained through two-dimensional downsampling in the range and azimuth directions, and the initial image is obtained through two-dimensional Fourier transform. The obtained initial image is used as the input of CSKNN, and the network output is the final imaging result.

[0039] Electromagnetic scattered echoes of the simulated target were generated using the electromagnetic simulation software FEKO. Radar parameters were set as follows: the simulation model was illuminated by a plane wave; under the far-field assumption, the reflected echoes of the simulation model were calculated using physical optics and collected by receivers positioned around the target model. The transmitted plane wave frequency band was 8.75 GHz to 9.25 GHz, with 100 frequency sampling points. The radar observed the target's elevation angles from 46° to 86°, totaling 40 elevation angles. The azimuth angle ranged from 0.05° to 360.05°, with an azimuth accumulation angle of 5°, accumulating 72 times. Therefore, the total number of radar echo data points was 2880.

[0040] The 2880 radar echo data points were divided into training and validation sets. Echo data at elevation angles of 50°, 60°, 70°, 80°, and 90° were assigned to the validation set, while the rest were assigned to the training set. The validation set contains 360 samples, and the training set contains 2520 samples. Each sample consists of one input image (initial image) and one label image. The label and initial images were obtained using the simulation data described above. The generation method involved randomly downsampling each echo data point in both the range and azimuth directions in two dimensions. The downsampling results were then zero-padded, and a two-dimensional FFT was performed to obtain a low-quality initial image. For the same data point with full data, the imaging result obtained directly through the RD method was used as the target image. One initial image and one target image constituted one training sample.

[0041] Next, a CSKNN needs to be constructed. The input of the CSKNN is the initial image obtained from downsampled data, and the output is the final imaging result. In constructing the CSKNN in this invention, multiple feature data are first extracted using a 3×3 convolutional kernel, and the bias between the feature data and the input data is reduced using BN layers and ReLU activation functions. Dimensionality reduction of the feature data is achieved using 2×2 max pooling. In the data reconstruction process, feature data is reconstructed using deconvolution layers with a 2×2 kernel size. The error between the reconstructed feature data and the dimensionality-reduced feature data is reduced using BN layers and ReLU activation functions. A residual learning mechanism is added to the last layer of the network, summing the initial image and the reconstructed feature data at the end of the network to obtain the final ISAR imaging result. The CSKNN constructed in this invention contains 26 hidden layers, specifically:

[0042] The first layer is a convolutional layer conv1, which contains 64 3×3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0043] The second layer is a convolutional layer conv2, which contains 64 3×3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0044] The third layer is a convolutional layer conv3, which contains 64 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0045] The fourth layer is a convolutional layer conv4. conv4 contains 16 3×3 filters that perform convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0046] The fifth layer is a convolutional layer conv5. conv5 contains 16 3×3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0047] The sixth layer is a convolutional layer conv6. conv6 contains 16 3×3 filters that perform convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0048] The seventh layer is the maxpooling1 pooling layer, which uses max pooling during the pooling process. The pooling layer has a 2β2 kernel and a stride of 2.

[0049] The eighth layer is a convolutional layer conv10, which contains 128 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0050] The ninth layer is a convolutional layer conv11, which contains 128 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0051] The tenth layer is a convolutional layer conv12, which contains 64 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0052] The eleventh layer is a convolutional layer conv13. conv13 contains 64 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated using the ReLU function.

[0053] The twelfth layer is the maxpooling2 pooling layer, which uses max pooling during the pooling process. The pooling layer has a 2β2 kernel and a stride of 2.

[0054] The thirteenth layer is a convolutional layer conv16. conv16 contains 256 3×3 filters that perform convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0055] The fourteenth layer is a convolutional layer conv17. conv17 contains 256 3×3 filters that perform convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0056] The fifteenth deconvolution layer, uconv1, contains 128 2×2 filters performing convolution operations with a stride of 2 and an edge padding of 1. After deconvolution, the output of the deconvolution layer is normalized by a BN layer, and then the output of the BN layer is nonlinearly activated using the ReLU function.

[0057] The sixteenth layer is the cropping layer crop1, which uses the output of the fifteenth layer as a reference to crop the output of the eleventh layer;

[0058] The seventeenth layer is the concat1 connection layer, which connects the output of the sixteenth layer with the output of the fifteenth layer to form the feature data of the seventeenth layer.

[0059] The eighteenth layer is a convolutional layer conv14, which contains 128 3β3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the output of the BN layer is non-linearly activated by the ReLU function.

[0060] The nineteenth layer is a convolutional layer conv15. conv15 contains 128 3×3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized through a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0061] The twentieth layer is the deconvolution layer uconv2. uconv2 has 64 2β2 filters performing convolution operations with a stride of 2 and an edge padding of 1. After the deconvolution is completed, the output of the deconvolution layer is normalized by a BN layer, and then the output of the BN layer is nonlinearly activated by the ReLU function.

[0062] The twenty-first layer is the crop2 layer, which uses the output of the twentieth layer as a reference to crop the output of the sixth layer.

[0063] The twenty-second layer is the concat2 connection layer, which connects the output of the twenty-first layer with the output of the twenty-second layer to form the feature data of the twenty-second layer.

[0064] The 23rd layer is a convolutional layer conv7. conv7 contains 64 3×3 filters performing convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized through a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0065] The 24th layer is a convolutional layer conv8. conv8 contains 64 3×3 filters that perform convolution operations with a stride of 1 and an edge padding value of 1. After the convolution is completed, the output of the convolutional layer is normalized by a BN layer, and then the ReLU function is used to non-linearly activate the output of the BN layer.

[0066] The 25th convolutional layer is conv9. conv9 contains a 1×1 filter that performs a convolution operation with a stride of 1. After the convolution, it passes through a BN normalization layer, and then uses the ReLU function to perform non-linear activation on the output of the BN layer.

[0067] The 26th layer is a summation layer that sums the initial input image with the output of the 25th layer point by point to obtain the output.

[0068] The 27th layer is the loss function layer, which calculates the error between the output of the 26th layer and the final imaging result in the training samples, and then updates the network parameters through backpropagation and gradient descent algorithms.

[0069] Based on the construction of CSKNN, this invention introduces a novel skip connection construction method, replacing the traditional SK with convolutional layers, and reducing the number of channels in the convolutional layers to improve their abstract feature extraction capabilities. It establishes the connection between the shallow feature representations and deep feature representations of the network in the same level, and constructs feature connections between layers.

[0070] After constructing CSKNN, given the corresponding training strategy and loss function, the network parameters can be updated. This invention employs a batch gradient descent training strategy and uses the minimum mean square error function as the loss function:

[0071]

[0072] Where T represents the total number of training samples, and i represents the i-th training sample. This is the reconstruction result output by CSKNN. Let {W} represent the initial image in the i-th training sample, {W} represent the CSKNN parameter set, f(·) represent CSKNN, and σi iLet L({W}) represent the labeled image of the i-th sample. L({W}) represents the mean square error of the reconstruction.

[0073] The training strategy of this invention is as follows: the baseline learning rate is 1e-5, the learning rate remains unchanged, 2520 training samples are used, 360 validation samples are used, and 179 validation iterations are performed. When both the training loss and the testing loss of the network are sufficiently small and stable, the CSKNN training is complete, and a network that meets the task requirements is obtained.

[0074] After obtaining the trained CSKNN, it can be used to achieve ISAR downsampled data imaging. ISAR data generated using the electromagnetic simulation software FEKO is downsampled in both the range and azimuth directions to obtain downsampled data at different angles with a sampling rate of 25%. The downsampled data is then subjected to a two-dimensional Fourier transform to obtain the initial image. This initial image serves as the input to the CSKNN, and the output of the CSKNN is the final imaging result. The CSKNN imaging principle diagram is shown below. Figure 5 As shown.

[0075] Finally, the results of the CSKNN imaging method are compared in detail with the experimental results of the FCNN, GKF, and OMP imaging methods in terms of imaging quality, imaging characteristics, and computational complexity, and qualitative and quantitative comparative analysis results are given.

[0076] In practice, this invention is divided into a training phase and an imaging phase.

[0077] In the training phase of CSKNN, the first step is to construct the training dataset, obtaining the training and validation sets. Electromagnetic scattering echoes of the simulated target are generated using the electromagnetic simulation software FEKO. Radar parameters are set as follows: the simulation model is illuminated by a plane wave; under the far-field assumption, the reflected echoes of the simulation model are calculated using physical optics and collected by receivers positioned around the target model. The transmitted plane wave frequency band is 8.75 GHz to 9.25 GHz, with 100 frequency sampling points. The radar observes the target's elevation angle from 46° to 86°, totaling 40 elevation angles. The azimuth angle ranges from 0.05° to 360.05°, with an azimuth accumulation angle of 5°, accumulating 72 times. Therefore, the total number of radar echo data is 2880.

[0078] The 2880 radar echo data points were divided into training and validation sets. Echo data at elevation angles of 50°, 60°, 70°, 80°, and 90° were assigned to the validation set, while the rest were assigned to the training set. The validation set contains 360 samples, and the training set contains 2520 samples. Each sample consists of one input image (initial image) and one label image. The label and initial images were obtained using the simulation data described above. The generation method involved randomly downsampling each echo data point in both the range and azimuth directions in two dimensions. The downsampling results were then zero-padded, and a two-dimensional FFT was performed to obtain a low-quality initial image. For the same data point with full data, the imaging result obtained directly through the RD method was used as the target image. One initial image and one target image constituted one training sample.

[0079] The next step is to construct a CSKNN and learn its parameters using a dataset. The CSKNN structure is constructed as follows: Figure 1 As shown, the first, second, and third levels each undergo convolution, batch normalization, and corrected linear unit (SLU) operations. Specifically, layers C6 to D2 and C13 to D1 perform skip connections, cropping, and concatenation operations; layers C3 to M1 and C11 to M2 perform max pooling operations from the first to the second and from the second to the third levels, respectively; layers C17 to D1 and C15 to D2 perform deconvolution, batch normalization, and corrected SLU operations from the third to the second and from the second to the first levels, respectively; and finally, layer C9, after convolution, is summed with the initial image after skip connections to obtain the label image. The strategy adopted is to introduce a novel skip connection construction method on CSKNN, replacing the traditional SK with convolutional layers and reducing the number of channels in the convolutional layers to improve their abstract feature extraction capabilities, establishing connections between shallow and deep layers of the network, and constructing connections between shallow and deep feature representations within the same level. Features are extracted from the input initial image and dimensionality reduced, and the final imaging result is reconstructed using the low-dimensional feature data. Feature extraction and dimensionality reduction are accomplished using convolutional layers and max pooling layers, respectively. Batch normalization (BN) layers, ReLU layers, and deconvolutional layers are added to CSKNN to reduce the error between the reconstructed result and the initial input image, resulting in a network architecture with stronger abstract feature extraction capabilities. The minimum mean squared error loss function is used to evaluate the reconstruction results, and the error is backpropagated, then the network parameters are updated using stochastic gradient descent. Training is complete when both the training loss and the testing loss stabilize and become sufficiently small.

[0080] In the CSKNN imaging stage, the input to CSKNN is downsampled drone data at 25% different angles, designated as Data 1 and Data 2. The parameter information for the drone data is listed in Table 1. The initial image obtained after FFT of the downsampled drone data is used as the input to CSKNN, and the output of CSKNN is the final imaging result. The final imaging result of CSKNN is shown below. Figure 3 and Figure 4 As shown, Figure 3 The images are from UAV data with a downsampling rate of 25%, processed using CSKNN, FCNN, and compressed sensing imaging methods. Figure 3 (a) in the image is the CSKNN imaging result. Figure 3 (b) in the image is the FCNN imaging result. Figure 3 (c) in the image is the GKF imaging result. Figure 3 (d) in the figure represents the OMP imaging result. Figure 4 The images are obtained from UAV data with a downsampling rate of 25% using CSKNN, FCNN, and compressed sensing imaging methods; among them, Figure 4 (a) in the image is the CSKNN imaging result. Figure 4 (b) in the image is the FCNN imaging result. Figure 4 (c) in the image is the GKF imaging result. Figure 4 (d) in the figure represents the OMP imaging result.

[0081] Table 1 shows the parameter information for UAV data.

[0082]

[0083] Implementation Examples

[0084] New ISAR echo data, different from the training set, was selected, downsampled by 25%, and imaged using a pre-trained CSKNN. The results are as follows. Figure 2 As shown. Figure 2 Tables (a)-(d) show the imaging results of the network obtained after different training iterations. Figure 2 As can be seen, with the increase of training iterations, the number of false scattering points in the background region in the CSKNN imaging reconstruction results is reduced, and the target image is clearer. Considering both hardware facilities and computational efficiency, and in order to conserve resources, this invention will... Figure 2 (c) is used as the optimal network selection, and 50,000 training iterations are used as the optimal training metric for CSKNN.

[0085] To objectively evaluate the performance of the proposed imaging method, the CSKNN imaging method is compared with the FCNN, GKF, and OMP imaging methods, clearly demonstrating the superiority of the CSKNN-based ISAR imaging method. It is important to emphasize that the imaging FCNN uses the same training dataset as the imaging CSKNN to update its network parameters.

[0086] Quantitative evaluation of 25% UAV data imaging results using various ground truth values: False Alarm (FA) reflects the number of incorrectly reconstructed scattering points in the imaging results; Missed Detection (MD) reflects the number of scattering points in the target area that was not reconstructed in the imaging results; Relative Root Mean Square Error (RRMSE) reflects the degree of deviation between the predicted or estimated value and the true value in the imaging results; and there is also the traditional image quality evaluation metric, Target-To-Clutter Ratio (TCR).

[0087] As shown in Table 2, the CSKNN imaging method has the lowest FA (Average Fraction), indicating that compared to the FCNN and compressed sensing methods, the CSKNN reconstruction results have the fewest incorrectly reconstructed scattering points, and the number of false scattering points in the background region should be the smallest. Further comparison of the RRMSE index reveals that the CSKNN-reconstructed image has the lowest RRMSE, indicating that the scattering point amplitude reconstruction error of the CSKNN imaging method is significantly smaller than that of the FCNN and compressed sensing methods.

[0088] Table 2 presents a quantitative evaluation of the imaging results of 25% downsampling data under different imaging methods.

[0089]

[0090] The TCR of CSKNN imaging results increases with the number of training iterations, resulting in improved image contrast and decreased image entropy. A well-trained CSKNN imaging result exhibits a higher TCR than FCNN and compressed sensing methods, while also displaying lower image entropy and higher contrast.

[0091] Regarding computation time, as shown in Table 2, which presents a quantitative evaluation of the imaging results under different imaging methods with 25% downsampled data, once CSKNN is trained, the imaging time of the CSKNN method can reach the order of 0.1 seconds, which is significantly more efficient than other compressed sensing methods.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An ISAR imaging method for downsampled data, characterized in that, Includes the following steps: Step S1: Construct the ISAR dataset for training CSKNN, group the ISAR dataset to obtain the training set and validation set of CSKNN; Step S2: Construct convolutional skip connections using convolutional layers, and use the constructed convolutional skip connections to construct CSKNN for ISAR downsampled data imaging; wherein, the construction method of convolutional skip connections refers to: replacing skip connections SK with convolutional layers, establishing the connection between shallow feature representations and deep feature representations in the same level of the network; reducing the number of channels in the convolutional layers to narrow the semantic gap between shallow features and deep features. Step S3: Based on the training set and validation set generated in step S1, and given the form of the loss function, the parameters of CSKNN are learned by combining backpropagation and gradient descent algorithms. When the training loss and validation loss of CSKNN are small enough and both tend to be stable, the CSKNN parameter update is stopped, and the trained CSKNN is obtained. Step S4: Use the CSKNN obtained in step S3 to perform ISAR downsampling data imaging; In step S2, CSKNN for ISAR downsampled data imaging is constructed using convolutional layers, max pooling layers, BN layers, activation function layers, deconvolutional layers, residual learning mechanisms, and SK cascades. The process of constructing a CSKNN is as follows: the input to a CSKNN is the initial image obtained from downsampled data, and the output is the final imaging result; when constructing a CSKNN, firstly by utilizing... The convolutional kernel extracts multiple feature data, and the BN layer and ReLU activation function are used to reduce the deviation between the feature data and the input data; Maximization pooling is used to reduce the dimensionality of the feature data; in the data reconstruction process, deconvolution layers are used to reconstruct the feature data, with a deconvolution kernel size of [missing value]. The error between the reconstructed feature data and the dimensionality-reduced feature data is reduced by using BN layers and ReLU activation functions. A residual learning mechanism is added to the last layer of CSKNN to sum the initial image and the feature data reconstructed by the network at the end of CSKNN to obtain the final ISAR imaging result.

2. The ISAR imaging method for downsampled data according to claim 1, characterized in that, In step S1, the electromagnetic scattering echo of the simulated target is generated using the electromagnetic simulation software FEKO to construct an ISAR dataset.

3. The ISAR imaging method for downsampled data according to claim 2, characterized in that, In step S1, the ISAR dataset is grouped to obtain the training and validation sets for CSKNN; specifically as follows: Construct the ISAR dataset for training CSKNN: The radar parameters are set as follows: the simulation model is illuminated by a plane wave. Under the far-field assumption, the reflected echo of the simulation model is calculated by physical optics and collected by receivers set around the target model; the transmitted plane wave frequency band is 8.75GHz~9.25GHz, and the number of frequency sampling points is 100. The elevation angle of the radar target is 46°~86°. o There are a total of 40 pitch angles; Azimuth from 0.05 o Up to 360.05 o The azimuth accumulation angle was 5°, and a total of 72 accumulations were performed; the number of ISAR datasets obtained was 2880. The ISAR dataset was grouped to obtain the training and validation sets for CSKNN; the elevation angle was set to 50°. o 60 o 70 o 80 o 90 o The echo data was divided into a validation set and the rest into a training set. The validation set contained 360 samples, and the training set contained 2520 samples. Each sample consisted of an initial input image and a label image. The label image and the initial image were obtained using the simulation data mentioned above, which refers to 2880 ISAR data points in the ISAR dataset. The generation method was to first randomly downsample each echo data point in two dimensions in the range and azimuth directions, pad the downsampling results with zeros, and then perform a two-dimensional FFT to obtain a low-quality initial image. For the same data, the imaging result obtained directly through the RD method was used as the target image in the case of the full dataset. An initial image and a target image constitute a training sample.

4. The ISAR imaging method for downsampled data according to claim 1, characterized in that, In step S3, a batch gradient descent training strategy is adopted, and the minimum mean squared error function is used as the loss function: (1)； in, This represents the total number of training samples. Indicates the first training samples, This is the reconstruction result output by CSKNN. Indicates the first The initial image from each training sample, This represents the set of CSKNN parameters. Represents CSKNN, Indicates the first Labeled images of each sample, This represents the mean square error of the reconstruction.

5. The ISAR imaging method for downsampled data according to claim 1, characterized in that, Step S4 is as follows: The ISAR data generated by the electromagnetic simulation software FEKO is downsampled in the range and azimuth directions to obtain downsampled data. The downsampled data is then subjected to a two-dimensional Fourier transform to obtain an initial image. The obtained initial image is used as the input of CSKNN, and the output of CSKNN is the final imaging result.

Images