Airborne array radar foresight imaging method and device based on improved U-Net, and medium

Abstract

The invention discloses an airborne array radar foresight imaging method and device based on improved U-Net, and a medium, and the method comprises the steps: carrying out the preprocessing of a high-resolution SAR image according to an established airborne foresight array radar signal geometric model, and generating a foresight echo sample, generating a simulation foresight data set in which the echo samples are in one-to-one correspondence with the image tags, and driving deep learning training; constructing an improved U-Net network model, wherein the improved U-Net network model comprises an encoder and a decoder; inputting an echo sample into an encoder, obtaining a foresight image by a decoder, comparing the foresight image with a high-resolution image in a data set, and optimizing and updating a network weight through iterative return of a loss function; and inputting a to-be-processed echo into the trained network to obtain a high-resolution forward view image. According to the method, the forward-looking echo features can be extracted to a greater extent, the high-resolution forward-looking image is restored, the consumed time is short, and the method is suitable for real-time airborne processing.

Description

Technical Field

[0001] The present invention belongs to the field of radar imaging technology, and relates to airborne radar forward-looking imaging signal processing technology, and specifically to an airborne array radar forward-looking imaging method, device, and medium based on an improved U-Net. Background Art

[0002] High-resolution, two-dimensional imaging of the forward-looking area of an airborne radar is a crucial function of modern airborne radar systems. It plays a vital role in battlefield reconnaissance, ground attack, emergency obstacle avoidance, airdrops, and blind landings. Furthermore, forward-looking imaging is crucial for terminal guidance of missiles, offering a key technical solution to the current problem of guided missiles requiring "tilted aiming."

[0003] In the field of radar imaging, synthetic aperture radar (SAR) is currently recognized as the most successful technology. Its two-dimensional imaging resolution has exceeded the centimeter level, approaching the level of optical imaging. Radar transmits high-frequency radio waves through an antenna. When the radio waves encounter an object, they are reflected and then captured by the radar system's receiving antenna. By processing and analyzing the reflected waves, a detailed image of the target object or area can be generated. However, when the imaging area is located in front of the moving platform, the Doppler frequency difference between scattering points in different azimuths approaches zero, making it difficult to distinguish scattering points located in different azimuths from the Doppler history, and it is impossible to obtain a forward-looking image. Therefore, for existing imaging radar systems, the radar image quality within a range of plus or minus 5 degrees in the platform's flight direction is severely degraded, becoming the so-called imaging "blind zone."

[0004] Real beam imaging is an imaging method used in early radars. A signal processor accumulates received echo signals and then performs coordinate transformation to produce a ground image of the area swept by the antenna beam. This method's azimuth resolution is completely limited by the beamwidth, making it impossible to produce high-resolution images.

[0005] Deconvolution forward-looking imaging technology treats the echo signal acquired by a scanning radar as the convolution of the antenna pattern and the target scattering coefficient in azimuth. Therefore, deconvolution can be used to invert the ground scene from the echo signal, thereby improving image resolution. This process involves a variety of complex signal processing algorithms, such as compressed sensing, Bayesian methods, Wiener filtering, and regularization. These algorithms can, to a certain extent, reduce or eliminate the errors introduced by convolution, restoring a more realistic target image. However, the imaging resolution is limited and the computational complexity is high.

[0006] Monopulse imaging technology incorporates sum-and-difference beam monopulse technology into radar imaging, utilizing angle measurement to improve image quality. Compared to traditional real-beam imaging, monopulse imaging offers certain advantages in improving imaging resolution. However, due to limitations in spatial degrees of freedom, this technology struggles to simultaneously distinguish multiple targets within a beam, and imaging quality needs further improvement.

[0007] While existing radar forward-looking imaging algorithms have advanced the forward-looking resolution of airborne radars at various stages, overall, these imaging radar systems are difficult to implement, making them unsuitable for airborne radar platforms under combat conditions, and their resolution improvement is limited. Because the models used in deep learning and radar forward-looking imaging are essentially deconvolution problems, a deep learning-based forward-looking imaging method has been proposed. Summary of the Invention

[0008] Purpose of the invention: The present invention provides an airborne array radar forward imaging method, device and medium based on an improved U-Net, which significantly increases the imaging resolution and saves time, making it suitable for real-time airborne processing.

[0009] Technical solution: The present invention provides an airborne array radar forward imaging method based on an improved U-Net, which specifically includes the following steps:

[0010] (1) Construct a forward-looking echo signal model for airborne array radar, thereby generating range-pulse two-dimensional forward-looking echo data from high-resolution SAR images, and producing a data set with a one-to-one correspondence between echoes and images;

[0011] (2) Constructing an improved U-Net network model, including an encoder, a decoder, a shallow convolution module and an asymmetric feature fusion module; the encoder includes three encoding blocks, each encoding block includes two convolution layers and a maximum pooling layer, which is used to extract multi-scale low-level features; the decoder includes three decoding blocks, each decoding block includes a deconvolution layer and two convolution layers, which is used to gradually restore image resolution and enhance high-level feature expression; the shallow convolution module uses two sets of convolution layers stacked, followed by a convolution layer to compensate for shallow information; the asymmetric feature fusion module realizes information exchange between different scales;

[0012] (3) The forward-looking echo generated in (1) is input into the encoder, and the decoder obtains the forward-looking image, which is compared with the high-resolution image in the dataset. The network weights are optimized and updated through the iterative return of the loss function;

[0013] (4) The echo to be processed is input into the network to obtain high-resolution forward-looking imaging.

[0014] Furthermore, the implementation process of step (1) is as follows:

[0015] A three-dimensional mathematical model of the forward scanning echo signal of an airborne array radar under the conditions of channel error, motion error, and noise is constructed. The radar transmits a linear frequency modulation signal while scanning the entire forward-looking scene. The two-dimensional signal received by the radar undergoes down-conversion, range pulse compression, and range migration correction, and is expressed as discrete echoes in range and azimuth directions as follows:

[0016]

[0017] Where τ and t represent fast-time variables and slow-time variables, respectively; L represents the unit length in range; K represents the unit length in azimuth; B represents the signal bandwidth; c represents the speed of light; λ represents the carrier wavelength; and σ(·) represents the range R l and azimuth θ k The scattering coefficient of the scatterer at the position, h(·) function represents the antenna pattern weight at that position, sinc( · ) function characterizes the signal waveform; by compensating for the delayed phase of the slant range and considering that the Doppler frequency gradient remains approximately constant within the forward scanning range, and ignoring the influence of the Doppler frequency shift, the above formula is simplified to:

[0018]

[0019] Thus, the forward-looking echo can be generated by the convolution of the ground scattering coefficient and the antenna pattern.

[0020] Furthermore, the coding block in step (2) takes echoes of different scales as input, and uses the scaled-down features and the downsampled echoes to complement information; specifically, the scaled-down features extracted by the coding block in the previous layer of the network are received, and then combined with the features extracted by the downsampled echoes through a convolution layer with a stride of 2, and a feature attention module is used to actively emphasize the previous scale features; the combined features are passed through a 3×3 convolution layer, and the output includes complementary deblurring information; and then a residual connection is performed for further refinement.

[0021] Furthermore, the residual connection is composed of 20 stacked residual blocks, and each residual block consists of two 3×3 convolutional layers.

[0022] Furthermore, the shallow convolution module in step (2) extracts features from the downsampled echo, and after fusing with the output of the coding block of the previous layer, it is passed to the coding block of the current layer. Specifically, two groups of convolution layers of 3×3 and 1×1 are stacked, and then this output is combined with the input of the shallow convolution module, and connected to the 1×1 convolution layer for feature refinement, thereby enhancing the ability to extract detailed information such as texture and edges in the radar echo.

[0023] Furthermore, the asymmetric feature fusion module in step (2) takes the output of all encoding blocks as input, extracts features using 3×3 and 1×1 convolution kernels, and then performs channel attention weighted fusion; the output of the asymmetric feature fusion module is passed to the corresponding decoding block.

[0024] Furthermore, the output result of the decoder in step (2) is subjected to a layer of convolution to map the feature map to an image, and then jump-connected with the original echo to obtain a high-resolution front view image.

[0025] Furthermore, the implementation process of step (3) is as follows:

[0026] Using the Pytorch deep learning platform, the optimization function is Adam, and the basic learning rate is set to 1×10 -4 , gamma is 0.5, and GPU is used for accelerated training; paired forward-looking echoes and high-resolution SAR images are used as data sets, the echoes are input into the network, and the reconstructed high-resolution forward-looking images are obtained, which are compared with the original high-resolution SAR images to obtain the multi-scale content loss function and the multi-scale frequency reconstruction loss function, and back-propagation is performed to adjust the network weight parameters.

[0027] The present invention provides a storage medium having a computer program stored thereon, which, when executed by at least one processor, implements the steps of the improved U-Net-based airborne array radar forward imaging method.

[0028] A device according to the present invention includes a memory and a processor, wherein:

[0029] a memory for storing computer programs capable of running on the processor;

[0030] The processor is configured to execute the steps of the airborne array radar forward imaging method based on the improved U-Net when running the computer program.

[0031] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are: the present invention uses simulated data sets to solve the problem of scarcity of high-resolution forward-looking labels, thereby driving deep learning training; uses a multi-input and multi-output encoding and decoding U-shaped network to construct a complex nonlinear mapping relationship between the echo and the forward-looking image, and introduces multi-scale loss to improve the generation effect; compared with traditional forward-looking imaging methods, it can extract forward-looking echo features to a greater extent and restore high-resolution forward-looking images, and it is time-saving and suitable for real-time airborne processing, exploring new technical approaches in the field of radar imaging and obtaining two-dimensional high-resolution imaging images. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1This is a flow chart of the airborne array radar forward imaging method based on the improved U-Net;

[0033] Figure 2 To improve the U-Net network structure diagram;

[0034] Figure 3 Schematic diagram of shallow convolutional layer, feature attention module and asymmetric feature fusion module;

[0035] Figure 4 is a point target image used to generate forward-looking echoes;

[0036] Figure 5 Forward-looking echo map of point targets;

[0037] Figure 6 This is the forward imaging result image of the point target;

[0038] Figure 7 is a scene 1 diagram for generating forward-looking echoes;

[0039] Figure 8 This is the forward echo map of scene 1;

[0040] Figure 9 This is the result of imaging the forward echo of simulation scene 1 using the method proposed in the present invention;

[0041] Figure 10 is a scene 2 diagram for generating forward-looking echoes;

[0042] Figure 11 This is the forward echo image of scene 2;

[0043] Figure 12 This is the result of imaging the forward echo of simulation scene 2 using the method proposed in the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below with reference to the accompanying drawings.

[0045] like Figure 1 As shown, the present invention proposes an airborne array radar forward imaging method based on an improved U-Net, comprising the following steps:

[0046] Step 1: Build a forward-looking echo signal model for airborne array radars. To address the scarcity of high-resolution forward-looking labeled data, this model is used to simulate existing high-resolution SAR images and generate corresponding two-dimensional range-pulse forward-looking echo data, thereby constructing a dataset with a one-to-one correspondence between imaging echoes and images. This method generates training samples on a large scale through simulation, enabling deep learning modeling driven by simulation data, effectively improving the feasibility and generalization capabilities of network training.

[0047] A three-dimensional mathematical model of the forward-looking scanning echo signal of an airborne array radar under the conditions of channel error, motion error, and noise is constructed. Based on this model, a high-precision two-dimensional data echo simulation algorithm is proposed. As the radar scans the entire forward-looking scene, it emits a linear frequency modulation signal. The two-dimensional signal received by the radar undergoes down-conversion, range pulse compression, and range shift correction. The discrete echoes in range and azimuth can be expressed as:

[0048]

[0049] Where τ and t represent fast-time variables and slow-time variables, respectively. L represents the unit length in range, K represents the unit length in azimuth, B represents the signal bandwidth, c represents the speed of light, and λ represents the carrier wavelength. The σ(·) function represents the distance R l and azimuth θ k The scattering coefficient of the scatterer at the position, h(·) function represents the antenna pattern weight at that position, sinc( · ) function characterizes the signal waveform. By compensating for the delayed phase of the slant range, taking into account that the Doppler frequency gradient remains approximately constant within the forward scanning range, and ignoring the influence of the Doppler frequency shift, the above formula can be simplified to:

[0050]

[0051] This shows that forward-looking echoes can be generated by convolving the ground scattering coefficient with the antenna pattern. Using existing high-resolution airborne SAR measured data, simulated data for forward-looking echoes from airborne array radars can be inverted. High-resolution SAR images obtained from the measured data are used as label images to develop a forward-looking imaging dataset with a one-to-one correspondence between samples and labels, enabling simulation-driven deep learning training.

[0052] Step 2: Build Figure 2 、 Figure 3The improved U-Net network model shown in the figure includes an encoder, a decoder, a shallow convolutional module (SCM), and an asymmetric feature fusion module (AFF). The encoder and decoder are composed of three encoding blocks (EBs) and three decoding blocks (DBs), respectively. The encoder consists of three encoding blocks (EBs), each of which includes two convolutional layers and a max pooling layer to extract multi-scale low-level features. The decoder consists of three decoding blocks (DBs), each of which includes a deconvolution layer and two convolutional layers to gradually restore image resolution and enhance high-level feature representation. The shallow convolution module consists of two stacked 3×3 and 1×1 convolutional layers, followed by a 1×1 convolutional layer for shallow layer information compensation. The asymmetric feature fusion module uses 3×3 and 1×1 convolution kernels to extract features and then performs channel-wise attention-weighted fusion. Compared with the traditional U-Net, the addition of the shallow convolution module enhances the ability to extract detailed information such as texture and edges in radar echo images. The asymmetric feature fusion module achieves accurate integration of cross-scale features, improving the azimuth ambiguity caused by beam broadening in the forward imaging process. The introduction of skip connections and attention mechanisms for channel weighting can avoid redundant feature interference.

[0053] The encoder uses radar echoes of different sizes as input and consists of three encoding blocks. It then uses the features after the upper layer is reduced and the echoes downsampled by the shallow convolution module to complement each other. This method can effectively handle the blur of various images. The shallow convolution module is used to extract features from the downsampled echo, and its output is represented as Specifically, two sets of convolutional layers of 3×3 and 1×1 are stacked, and then the output is combined with the input of the shallow convolution module, and finally a 1×1 convolutional layer is used to refine the connection features to obtain To make the previous layer encoding block output The feature size and output of the shallow convolution module Consistent fusion, Then use a convolutional layer with a stride of 2 to get the output The fusion of the two utilizes a feature attention module (FAM) to actively emphasize the previous scale features and learn the important features of the spatial channel features from the shallow convolution module, which helps to improve the PSNR index. and Multiply by elements; then, pass the multiplied output into a 3×3 convolution layer; finally, add the output of the previous convolution layer to the Residual connections are used for further refinement. Each residual block consists of two 3×3 convolutional layers, and 20 such residual blocks are stacked.

[0054] The asymmetric feature fusion module realizes the information exchange between different scales. It takes the output of all encoding blocks as input, resizes them, concatenates the resized features, and uses a 1×1 and a 3×3 convolution layer to combine multi-scale features. The output of the asymmetric feature fusion module will be passed to the corresponding decoding block. as well as Upsampling (↑) and downsampling (↓) are used to facilitate the fusion of features of different scales.

[0055] The decoder is also composed of three decoding blocks. First, the output of the asymmetric feature fusion module is It is concatenated with the output of the next layer’s decoding block. Then, the concatenated output is connected to a 1×1 convolution layer for feature fusion. Next, the output of the previous step is fed into a stacked residual block, which has the same structure as the residual block in the encoding block. A transposed convolution is then connected to achieve an upsampling effect. This operation is done so that after it is input into the previous layer’s decoding block, it can be combined with the output of the asymmetric feature fusion module. Finally, a 3×3 convolutional layer with 3 output channels is connected to the output end of the residual block to output clear images of different scales.

[0056] Step 3: Input the forward-looking echo generated in step 1 into the encoder, and the decoder obtains the forward-looking imaging, which is compared with the high-resolution image in the dataset. The network weights are optimized and updated through the iterative feedback of the loss function.

[0057] Using the Pytorch deep learning platform, the optimization function is Adam, and the basic learning rate is set to 1×10 -4 , gamma is 0.5, and GPU is used for accelerated training; paired forward-looking echoes and high-resolution SAR images are used as data sets, the echoes are input into the network, and the reconstructed high-resolution forward-looking images are obtained, which are compared with the original high-resolution SAR images to obtain the multi-scale content loss function and the multi-scale frequency reconstruction loss function, and back-propagation is performed to adjust the network weight parameters.

[0058] The network loss function consists of two parts. The first is the multi-scale content loss function used by most multi-scale deblurring networks:

[0059]

[0060] Where K is the number of network layers, t k is the number of all elements, divided by t kThe purpose is to perform normalization. The second is the multi-scale frequency reconstruction (MSFR) loss function, which helps to recover the lost high-frequency parts, reduce the difference in frequency space, and help improve model performance. The multi-scale frequency reconstruction loss measures the L1 distance between the multi-scale real image and the network imaging in the frequency domain:

[0061]

[0062] Where F represents the fast Fourier transform, which is used to convert the image signal into the frequency domain. The final loss function is as follows:

[0063] L total =L cont +λL MSFR

[0064] Step 4: Input the echo to be processed into the network to obtain high-resolution forward-looking imaging.

[0065] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by at least one processor, implements the steps of the airborne array radar forward-looking imaging method based on the improved U-Net as described above.

[0066] The present invention also provides a device, including a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the airborne array radar forward imaging method based on the improved U-Net when running the computer program.

[0067] Input the forward-looking echo to be restored into the network to obtain the generated high-resolution forward-looking image. Select the simulated point target as the ground scene. Figure 4 As shown, the radar forward-looking echo is generated as follows Figure 5 As shown, Figure 6 This is the forward imaging result of the point target. It can be seen that all point targets can be identified, and the imaging contours of the point targets are clear, and the azimuth focus is no longer blurred. Figure 7 、 Figure 10 are high-resolution SAR images of ground scenes 1 and 2, Figure 8 、 Figure 11 The forward-looking echo generated by this method has a low angular resolution and cannot distinguish point targets in azimuth, and the contour features of surface targets are blurred. Figure 9 、 Figure 12 This is the imaging result of the proposed method. It can be seen that the azimuth resolution is improved, strong scattering points are clear, the outline of the surface target is distinct, and the texture details of the image are well restored. This proves that the algorithm has better imaging effects and is more timely than traditional methods.

[0068] The present invention has been described in detail above with reference to specific embodiments. However, these descriptions should not be construed as limiting the present invention. Those skilled in the art will appreciate that various equivalent substitutions, modifications, or improvements may be made to the technical solutions and implementations of the present invention without departing from the spirit and scope of the present invention, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A forward-looking imaging method for airborne array radar based on improved U-Net, characterized in that: The following steps are involved: (1) Construct a forward-looking echo signal model for airborne array radar, thereby generating range-pulse two-dimensional forward-looking echo data from high-resolution SAR images, and producing a data set with a one-to-one correspondence between echoes and images; (2) Construct an improved U-Net network model, including an encoder, a decoder, a shallow convolution module, and an asymmetric feature fusion module; The encoder includes three encoding blocks, each of which includes two convolutional layers and a maximum pooling layer for extracting multi-scale low-level features; The decoder includes three decoding blocks, each of which includes a deconvolution layer and two convolution layers, which are used to gradually restore the image resolution and enhance the high-level feature expression; The shallow convolution module uses two sets of convolutional layers stacked together, followed by a convolutional layer to compensate for shallow information; the asymmetric feature fusion module realizes information exchange between different scales; (3) The forward-looking echo generated in (1) is input into the encoder, and the decoder obtains the forward-looking image, which is compared with the high-resolution image in the dataset. The network weights are optimized and updated through the iterative return of the loss function; (4) The echo to be processed is input into the network to obtain high-resolution forward-looking imaging.

2. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: The implementation process of step (1) is as follows: A three-dimensional mathematical model of the forward scanning echo signal of an airborne array radar under the conditions of channel error, motion error, and noise is constructed. The radar transmits a linear frequency modulation signal while scanning the entire forward-looking scene. The two-dimensional signal received by the radar undergoes down-conversion, range pulse compression, and range migration correction, and is expressed as discrete echoes in range and azimuth directions as follows: Where τ and t represent fast-time variables and slow-time variables, respectively; L represents the unit length in range; K represents the unit length in azimuth; B represents the signal bandwidth; c represents the speed of light; λ represents the carrier wavelength; and σ(·) represents the range R l and azimuth θ k The scattering coefficient of the scatterer at location , the h() function represents the antenna pattern weight at that location, and the sinc(·) function characterizes the signal waveform. By compensating for the delayed phase of the slant range and considering that the Doppler frequency gradient remains approximately constant within the forward scanning range, the influence of the Doppler frequency shift is ignored, and the above formula is simplified to: Thus, the forward-looking echo can be generated by the convolution of the ground scattering coefficient and the antenna pattern.

3. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: In step (2), the coding block takes echoes of different scales as input and uses the scaled-down features and the downsampled echoes to complement information. Specifically, the scaled-down features extracted by the coding block in the previous network layer are received, and then combined with the features extracted by the downsampled echoes through a convolution layer with a stride of 2, and a feature attention module is used to actively emphasize the previous scale features. The combined features are passed through a 3×3 convolution layer, and the output includes complementary deblurring information. Then, a residual connection is performed for further refinement.

4. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: The residual connection is composed of 20 stacked residual blocks, each of which consists of two 3×3 convolutional layers.

5. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: In step (2), the shallow convolution module extracts features from the downsampled echo, and after fusing with the output of the coding block of the previous layer, it is passed to the coding block of the current layer. Specifically, two groups of convolution layers, 3×3 and 1×1, are stacked. Then, this output is combined with the input of the shallow convolution module, and the 1×1 convolution layer is connected to perform feature refinement, thereby enhancing the ability to extract detailed information such as texture and edges in the radar echo.

6. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: The asymmetric feature fusion module in step (2) takes the output of all coding blocks as input, extracts features using 3×3 and 1×1 convolution kernels, and then performs channel attention weighted fusion; The output of the asymmetric feature fusion module is passed to the corresponding decoding block.

7. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: The output result of the decoder in step (2) is mapped to an image through a layer of convolution, and then jump-connected with the original echo to obtain a high-resolution front view image.

8. The airborne array radar forward imaging method based on improved U-Net according to claim 1, characterized in that: The implementation process of step (3) is as follows: Using the Pytorch deep learning platform, the optimization function is Adam, and the basic learning rate is set to 1×10 -4 , gamma is 0.5, and GPU is used for accelerated training; Paired forward-looking echoes and high-resolution SAR images are used as data sets. The echoes are input into the network to obtain the reconstructed high-resolution forward-looking image, which is compared with the original high-resolution SAR image to obtain the multi-scale content loss function and the multi-scale frequency reconstruction loss function. Backpropagation is performed to adjust the network weight parameters.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of the airborne array radar forward imaging method based on the improved U-Net as described in any one of claims 1 to 8.

10. A device, characterized in that: comprising a memory and a processor, wherein: a memory for storing computer programs capable of running on the processor; A processor, configured to, when running the computer program, execute the steps of the airborne array radar forward imaging method based on the improved U-Net according to any one of claims 1 to 8.

Images