A Single-Bit Synthetic Aperture Radar Imaging Reconstruction Training Method and Device Based on an Improved Loss Function

Abstract

The present application discloses a single-bit synthetic aperture radar imaging reconstruction training method and device based on an improved loss function. The training method includes obtaining a training sample set, where the training sample set includes high-precision synthetic aperture radar images and single-bit synthetic aperture radar images; inputting the training samples of the training sample set into a preset network model, and outputting a denoised synthetic aperture radar image through the preset network model; determining an image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image; training the preset network model and preset weight coefficients based on the image loss value to obtain an image reconstruction network. The present application improves the denoising effect and image quality of synthetic aperture radar images by jointly training based on single-bit quantization training samples and high-precision images and the image reconstruction network learning image features.

Description

Technical Field

[0001] This application relates to the field of radar imaging technology, and particularly to a single-bit synthetic aperture radar imaging reconstruction training method and device based on an improved loss function. Background Art

[0002] Synthetic Aperture Radar (SAR) is a high-resolution microwave imaging technology that is widely used in remote sensing monitoring, military reconnaissance, disaster emergency and other fields. SAR technology can provide clear images under various adverse weather conditions, with the advantages of all-day and all-weather operation. However, existing synthetic aperture radar imaging technologies face some technical problems in practical applications.

[0003] First, speckle noise and high-order harmonic aliasing in synthetic aperture radar images are the main technical problems. The noise and aliasing phenomena will seriously affect the clarity and contrast of the images, reducing the practical value of the images. Traditional synthetic aperture radar image processing methods mainly rely on complex and high-cost hardware devices to ensure image quality through high sampling rates, but this method is difficult to implement on resource-constrained small platforms (such as unmanned aerial vehicles and small satellites).

[0004] Second, with the increasing demand for high-quality radar images, traditional imaging algorithms also face huge challenges in terms of data volume, storage and processing efficiency. High-resolution imaging requires a large bandwidth, resulting in a huge amount of data, increasing the burden of transmission and processing. This contradiction is particularly prominent in the application of small platforms that require real-time performance. Summary of the Invention

[0005] The technical problem to be solved by this application is that existing synthetic aperture radar image processing methods cannot meet the requirements of low consumption and high clarity. In view of the deficiencies of the prior art, a single-bit synthetic aperture radar imaging reconstruction training method and device based on an improved loss function are provided to solve the problems of low clarity and contrast of existing single-bit radar images and the inability of traditional imaging algorithms to meet the actual needs.

[0006] To solve the above technical problems, in the first aspect of the embodiments of this application, a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function is provided. The method includes:

[0007] Obtain a training sample set, where the training sample set includes high-precision synthetic aperture radar images and single-bit synthetic aperture radar images;

[0008] Input the training samples of the training sample set into a preset network model, and output a denoised synthetic aperture radar image through the preset network model;

[0009] Determine an image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image;

[0010] Train a preset network model and preset weight coefficients based on the image loss value to obtain an image reconstruction network.

[0011] In an implementable manner of this embodiment, the process of obtaining the training sample set specifically includes:

[0012] Obtain a radar echo signal, perform sampling processing on the echo signal based on a preset sampling rate to obtain a sampled echo signal, and perform quantization processing on the sampled echo signal to obtain radar echo data, where the radar echo data is echo data after single-bit quantization;

[0013] Process the radar echo data based on a preset synthetic aperture radar imaging algorithm to obtain a single-bit synthetic aperture radar image, where the single-bit synthetic aperture radar image is a high-precision synthetic aperture radar image and a single-bit synthetic aperture radar image of a preset size;

[0014] Perform block processing on the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image to obtain a high-precision synthetic aperture radar image and a single-bit synthetic aperture radar image of a preset size;

[0015] Perform overlapping block processing on the small high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image of the preset size to obtain a first synthetic aperture radar image dataset;

[0016] Perform classification processing on the first synthetic aperture radar image dataset based on a preset scenario to obtain a training sample set.

[0017] In an implementable manner of this embodiment, the k estimation module of the preset network model includes a multi-scale feature fusion network and a residual neural network.

[0018] In an implementable manner of this embodiment, the process of determining the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image specifically includes:

[0019] Determine the mean square error loss of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image;

[0020] Determine the contrast loss of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image;

[0021] The mean square error loss and the contrast loss are weighted and summed based on a preset weight coefficient to obtain an image loss value.

[0022] In an implementable manner of this embodiment, the process of determining the mean square error loss of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image specifically includes:

[0023]

[0024] Where M represents the number of pixel rows of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image, N represents the number of pixel columns of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image, I represents the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image, and K represents the denoised synthetic aperture radar image.

[0025] In an implementable manner of this embodiment, the process of determining the contrast loss of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image specifically includes:

[0026]

[0027] Where H represents the high-precision synthetic aperture radar image, S represents the single-bit synthetic aperture radar image, and the contrast

[0028]

[0029] Where M and N respectively represent the number of rows and columns of the synthetic aperture radar image gray matrix, represents the pixel average value of the synthetic aperture radar image, represents the pixel average value of the m-th row and n-th column in the synthetic aperture radar image gray matrix.

[0030] In an implementable manner of this embodiment, before determining the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image, the method further includes:

[0031] Performing restoration and recombination processing on the denoised synthetic aperture radar image, the high-precision synthetic aperture radar image, and the single-bit synthetic aperture radar image to obtain the denoised synthetic aperture radar image, the high-precision synthetic aperture radar image, and the single-bit synthetic aperture radar image with a preset size, and the preset size is larger than the size of the image before restoration and recombination.

[0032] The second aspect of this application provides a method for using single-bit synthetic aperture radar imaging reconstruction based on an improved loss function, where the method includes:

[0033] Obtain synthetic aperture radar image data;

[0034] Input the synthetic aperture radar image data into a single-bit synthetic aperture radar imaging reconstruction network based on an improved loss function for enhanced radar imaging processing to obtain high-quality synthetic aperture radar image data.

[0035] The third aspect of this application provides a computer-readable storage medium, where the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps in any of the above-mentioned single-bit synthetic aperture radar imaging reconstruction training methods based on an improved loss function.

[0036] The fourth aspect of this application provides a terminal device, which includes: a processor, a memory, and a communication bus; a computer-readable program executable by the processor is stored on the memory;

[0037] The communication bus realizes the connection and communication between the processor and the memory;

[0038] When the processor executes the computer-readable program, it implements the steps in any of the above-mentioned single-bit synthetic aperture radar imaging reconstruction training methods based on an improved loss function.

[0039] Beneficial effects:

[0040] This application reduces the data volume by performing single-bit quantization processing on the radar echo signal, avoids the problem of high-order harmonic aliasing, realizes high-quality imaging at a low sampling rate, overcomes the technical problems of complex data processing, heavy storage and transmission burdens in traditional methods, and improves the clarity and contrast of the image;

[0041] Secondly, based on three no-reference evaluation metrics of contrast, information entropy, and equivalent number of looks, this application guides the network training process, overcomes the interference caused by speckle noise in synthetic aperture radar images, thereby improving the noise reduction performance of the image reconstruction network, and significantly improving the quality of the synthetic aperture radar image after noise reduction, with clearer details;

[0042] Thirdly, by improving the K estimation module of the AOD-Net network into a multi-scale feature fusion network and introducing a residual neural network to enhance the noise reduction ability and image reconstruction effect of the network, this application further improves the quality of the single-bit synthetic aperture radar image, thereby achieving the effect of high-performance radar imaging on a resource-constrained small platform. Description of the Drawings

[0043] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative labor, other accompanying drawings can also be obtained based on these drawings.

[0044] Figure 1 This is a flowchart of a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0045] Figure 2 This is a flowchart of a single-bit synthetic aperture radar imaging reconstruction network based on an improved loss function provided by the present application.

[0046] Figure 3 This is a flowchart of the loss function optimization of a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0047] Figure 4 This is a comparison chart of the noise reduction results of the AOD-Net, DnCNN, and SRN-DeblurNet networks in different scenarios in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0048] Figure 5 This is a speckle noise reduction map of the AOD-Net network in the port scenario in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0049] Figure 6 This is a structural diagram of the AOD-Net network in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0050] Figure 7 This is a structural diagram of the K estimation module of the AOD-Net network in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0051] Figure 8 This is a diagram of the K estimation module of the AOD-Net network with a multi-scale feature fusion DCNN network in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0052] Figure 9 This is a diagram of the K estimation module of the AOD-Net network with a multi-scale feature fusion DCNN network and a residual network ResNet in a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by the present application.

[0053] Figure 10 In the single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by this application, the AOD-Net network and the DnCNN network reconstruct the single-bit synthetic aperture radar image based on the denoising result.

[0054] Figure 11 In the single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by this application, it is a comparison chart of the denoising results of two improved AOD-Net networks in the sea area.

[0055] Figure 12 In the single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by this application, it is a comparison chart of the denoising results of the AOD-Net and the improved AOD-Net networks.

[0056] Figure 13 In the single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by this application, it is a comparison chart of the denoising results of the AOD-Net and the improved AOD-Net networks in College B of University A.

[0057] Figure 14 In the single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function provided by this application, it is a comparison chart of the results of restoring and reconstructing the single-bit synthetic aperture radar image in Campus a of University A.

[0058] Figure 15 It is the structural schematic diagram of the terminal device provided by this application. Specific implementation manners

[0059] This application provides a single-bit synthetic aperture radar imaging reconstruction training method and device based on an improved loss function. To make the purpose, technical solution and effect of this application clearer and more definite, the following further describes this application in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only used to explain this application and are not used to limit this application.

[0060] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of this application means the presence of the stated features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any unit and all combinations of one or more related listed items.

[0061] Those skilled in the art can understand that, unless otherwise defined, all terms used herein (including technical terms and scientific terms) have the same meaning as the general understanding of those of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless specifically defined as here.

[0062] It should be understood that the sequence numbers and magnitudes of the steps in this embodiment do not mean the order of execution, and the execution order of each process is determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of this application.

[0063] Synthetic Aperture Radar (Synthetic Aperture Radar) is a system that uses microwave imaging technology for remote sensing detection. By synthesizing a longer radar antenna aperture to achieve high-resolution images, it can provide clear images under various adverse weather conditions.

[0064] Speckle noise is a common type of noise in radar images. The speckle noise is generated by the interference of reflected waves, resulting in speckle-like noise in the image, which affects the clarity and contrast of the image.

[0065] Single-bit quantization is a data quantization method that reduces the amount of data and the burden of storage and transmission by quantifying multi-bit data into single-bit data.

[0066] Reference-free evaluation metrics are mainly metrics for image quality evaluation. They can evaluate the image quality without referring to a standard image. Common evaluation metrics include contrast, information entropy, and equivalent number of looks.

[0067] Neural network denoising is a method that uses a neural network model to perform denoising processing on images and improve the clarity and quality of images.

[0068] The multi-scale feature fusion network (DCNN) is a neural network structure that enhances the model's expressive ability by fusing features of different scales.

[0069] The residual neural network (ResNet) is a deep neural network structure that introduces residual connections to solve the vanishing gradient problem in deep networks and improve the training effect and performance of the network.

[0070] Synthetic aperture radar (SAR) technology is widely used in remote sensing monitoring, military reconnaissance, disaster emergency and other fields. Its advantages of high-resolution imaging ability and all-weather, all-day working make it of great value in various application scenarios. However, existing synthetic aperture radar imaging technologies face several technical problems in practical applications.

[0071] First of all, the speckle noise and high-order harmonic aliasing problems commonly existing in synthetic aperture radar images seriously affect the clarity and contrast of the images. Although some traditional methods attempt to solve these problems through high sampling rates and complex hardware devices, the high cost and huge data processing burden make these methods difficult to implement on small platforms (such as drones and small satellites).

[0072] Secondly, with the increasing demand for high-quality radar images, existing imaging algorithms also face huge challenges in terms of data volume, storage and processing efficiency. High-resolution imaging requires a large bandwidth, resulting in a huge amount of data and increasing the burden of transmission and processing.

[0073] In response to the above problems, some improvements have been made in the existing technology, such as introducing more efficient data compression and transmission algorithms, and using more advanced hardware devices to improve processing capabilities. However, these improvements still have the following problems:

[0074] 1. High-order harmonic aliasing and speckle noise are still difficult to completely eliminate;

[0075] 2. The problems of large data volume, low storage and processing efficiency have not been fundamentally solved;

[0076] 3. Complex and expensive hardware devices are difficult to apply on small platforms.

[0077] To solve the above problems, in the embodiments of the present application, a single-bit synthetic aperture radar imaging reconstruction training method and apparatus based on an improved loss function are first disclosed. The training method includes obtaining a training sample set, where the training sample set includes high-precision synthetic aperture radar images and single-bit synthetic aperture radar images; inputting the training samples of the training sample set into a preset network model, and outputting a denoised synthetic aperture radar image through the preset network model; determining an image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image; training the preset network model and preset weight coefficients based on the image loss value to obtain an image reconstruction network. The training method improves the denoising effect and image quality of synthetic aperture radar images by jointly training based on single-bit quantization training samples and high-precision images, so as to meet the requirements for high-quality radar imaging in practical applications.

[0078] The following further describes the content of the application by describing the embodiments in conjunction with the accompanying drawings.

[0079] This embodiment provides a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function, as Figure 1 and Figure 2 shown. The method includes:

[0080] S10. Obtain a training sample set, where the training sample set includes high-precision synthetic aperture radar images and single-bit synthetic aperture radar images.

[0081] The high-precision synthetic aperture radar image refers to a high-resolution image obtained by synthetic aperture radar (SAR) technology, and the single-bit synthetic aperture radar image is an image generated by processing the SAR echo signal through single-bit quantization technology.

[0082] Specifically, the process of obtaining the training sample set specifically includes:

[0083] S110. Obtain a radar echo signal, perform sampling processing on the echo signal based on a preset sampling rate to obtain a sampled echo signal, and perform quantization processing on the sampled echo signal to obtain radar echo data, where the radar echo data is the echo data after single-bit quantization.

[0084] First, the synthetic aperture radar (SAR) system first transmits a chirp signal on a radar platform (such as an unmanned aerial vehicle or a small satellite). The signal is sent to the target area through an antenna. When the radar signal encounters a ground target, it will be reflected. The reflected signal (referred to as the radar echo signal) is captured and received by the receiving antenna of the radar system.

[0085] The received radar echo signal is sampled according to a preset sampling rate. The preset sampling rate determines the number of signal samples collected per second, and the preset sampling rate is usually determined based on the bandwidth and resolution requirements of the system. The radar echo signal collected through the preset sampling rate is converted into discrete digital signals, forming a set of time series data.

[0086] The sampled radar echo signal is first preliminarily quantized by a multi-bit analog-to-digital converter (ADC). Multi-bit quantization usually uses quantization precisions of 8 bits, 12 bits, or 16 bits to ensure that the details of the signal are retained.

[0087] Based on multi-bit quantization, the sign bits of the quantization data are extracted, and the data of each sampling point is represented as a single bit. Thereby reducing the data volume, and each sampling point is represented by only 1 bit for positive and negative signs.

[0088] The finally generated radar echo data is the data after single-bit quantization. The data after single-bit quantization compresses the amount of information, reduces the burden of subsequent storage and transmission, and provides the possibility for efficient processing on small platforms.

[0089] S111. Process the radar echo data based on a preset synthetic aperture radar imaging algorithm to obtain a single-bit synthetic aperture radar image, and the single-bit synthetic aperture radar image is a high-precision synthetic aperture radar image and a single-bit synthetic aperture radar image of a preset size.

[0090] First, preprocess the collected single-bit radar echo data. The preprocessing operations include correcting the geometric distortion caused by platform movement, correcting the time delay and amplitude change of the echo signal. The preprocessed echo data can be stored in a preset format.

[0091] The preset synthetic aperture radar image algorithms can include the backprojection algorithm (BP) and the range-Doppler algorithm (RD), etc. The backprojection algorithm generates the final image by projecting each echo signal along its propagation path onto the imaging plane and accumulating the contributions of all echoes. The backprojection algorithm has high imaging accuracy and is suitable for scenarios with high-resolution requirements. The range-Doppler algorithm generates an image by performing range and Doppler processing on each echo signal and converting the frequency domain information into spatial domain information. The range-Doppler algorithm has a fast calculation speed and is suitable for scenarios with real-time imaging requirements.

[0092] Using full-bit data processed by the backprojection algorithm or the range-Doppler algorithm can generate high-resolution and high-precision synthetic aperture radar images;

[0093] The data after single-bit quantization is processed through the same algorithm to generate a high-precision synthetic aperture radar image and a single-bit synthetic aperture radar image. The sizes of both the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image are M×N.

[0094] S112. Perform block processing on the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image to obtain high-precision synthetic aperture radar images and single-bit synthetic aperture radar images of a preset size.

[0095] According to the computing power and training requirements of the network, preset the size of the blocks. For example, the block size can be set to r×r, where r represents the size of the block.

[0096] The steps of the block processing specifically include:

[0097] First, perform image size calculation: Calculate the width and height of the large-size image. Assume that the sizes of both the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image are r×r.

[0098] Second, perform block loop: Use the sliding window method to slide on the image with the set block size r×r and step size s. The step size s is usually set to be no greater than half of the block size to ensure sufficient overlap between blocks.

[0099] Third, perform block extraction: At each sliding position, extract the image part within the window to form an image block. Perform this operation on the high-precision synthetic aperture radar image and the single-bit synthetic aperture radar image respectively to obtain the corresponding image blocks.

[0100] Finally, perform edge processing: Process the edge regions that cannot be divided evenly during the sliding process to ensure that the edge parts can also form complete image blocks. Methods such as zero-padding or mirror extension can be used.

[0101] Considering that too small a block size may cause loss of image information, while too large a block size will increase the computational amount and make gradient transmission difficult. Therefore, it is necessary to reasonably set the block size. Determine the optimal block size through experiments according to the size of the existing synthetic aperture radar image and the number of required image samples. The specific selection depends on the resolution of the image and the computing power of the network.

[0102] S113. Perform overlapping block processing on the small high-precision synthetic aperture radar images and single-bit synthetic aperture radar images of the preset size to obtain a first synthetic aperture radar image dataset.

[0103] Before performing the overlapping block operation, small high-precision synthetic aperture radar image blocks and single-bit synthetic aperture radar image blocks of a preset size are obtained based on step S112. According to the block size r×r and the actual application requirements, the overlapping step s is set. The step s determines the size of the overlapping area. Generally, the step s is set to be no greater than half of the block size.

[0104] Using the sliding window method, overlapping block processing is performed on the high-precision synthetic aperture radar image blocks and single-bit synthetic aperture radar image blocks. Specifically:

[0105] Starting from the upper left corner of the image block, slide according to the set overlapping step s;

[0106] At each sliding position, the image part within the window is extracted to form an overlapping image block. This operation is performed separately for the high-precision synthetic aperture radar image blocks and single-bit synthetic aperture radar image blocks to obtain the corresponding overlapping image blocks;

[0107] Repeat the above operation until the entire image area is covered to ensure that all areas are processed.

[0108] It should be noted that if the overlapping step s is too small, there will be too much information crossover between adjacent blocks, increasing redundant information; if the step s is too large, the edge effect will be obvious and information will be lost. Therefore, it is necessary to reasonably set the overlapping step to balance information crossover and data utilization rate. Generally speaking, the common overlapping step is The specific selection depends on the image block size and the application scenario.

[0109] S114. Classify the first synthetic aperture radar image dataset based on a preset scenario to obtain a training sample set.

[0110] According to the actual application requirements, different scenarios are preset. For example, common scenarios include libraries, teaching buildings, stadiums, and Wenshan Lake, etc. Each scenario has different physical characteristics and environmental conditions, and can provide diverse training data. An independent label is assigned to each preset scenario.

[0111] From the previous overlapping block processing step, the first synthetic aperture radar image dataset is obtained. The dataset contains pairs of high-precision synthetic aperture radar image blocks and single-bit synthetic aperture radar image blocks. The dataset is classified according to the scenario label to which the image block belongs. For example, if the scenario corresponding to the image block is a library, then the image block is assigned to the category of the library scenario.

[0112] According to the ratio of 8:1:1, the image dataset under each scenario is divided into a training set, a validation set, and a test set. The specific steps are as follows:

[0113] Training set: It contains 80% of the image patches and is used to train the neural network.

[0114] Validation set: It contains 10% of the image patches and is used to verify the network performance and adjust hyperparameters during the training process.

[0115] Test set: It contains 10% of the image patches and is used to finally evaluate the generalization ability and performance of the network. The scenarios in the test set need to be different from those in the training set and the validation set to ensure the generalization ability of the model and the accuracy of new data testing.

[0116] For example, assuming the scenarios include library, teaching building, stadium and Wenshan Lake, the specific operation steps are as follows:

[0117] Scene label assignment:

[0118] Library: Label is 1,

[0119] Teaching building: Label is 2,

[0120] Stadium: Label is 3,

[0121] Wenshan Lake: Label is 4,

[0122] Image dataset classification:

[0123] For each image patch, assign the corresponding label according to the scene it belongs to. For example, the image patches in the library scene are assigned label 1, and the image patches in the teaching building scene are assigned label 2, and so on.

[0124] Dataset division:

[0125] Divide the image patches in each scene according to the ratio of 8:1:1.

[0126] For example, if there are 1000 image patches in the library scene, then 800 image patches are assigned to the training set, 100 image patches are assigned to the validation set, and 100 image patches are assigned to the test set.

[0127] S20. Input the training samples of the training sample set into a preset network model, and output the denoised synthetic aperture radar image through the preset network model.

[0128] Specifically, the k estimation module of the preset network model includes a multi-scale feature fusion network and a residual neural network.

[0129] Obtain the classified training sample set from step S10. The training samples include high-precision synthetic aperture radar image patches and single-bit synthetic aperture radar image patch pairs. Before inputting the training samples of the training sample set into the preset network model, preprocessing operations such as normalization processing can also be performed on the training samples to meet the input requirements of the network model.

[0130] The preset network model may be a denoising network such as DnCNN, a dehazing network such as AOD-Net, and a deblurring network such as SRN-DeblurNet. In the specific implementation of this application, based on the AOD-Network, its K estimation module is improved, and a multi-scale feature fusion network (DCNN) and a residual neural network (ResNet) are adopted to enhance the denoising performance of the model. By introducing a multi-scale feature fusion network into the K estimation module, the model's ability to capture features of different scales is enhanced, the denoising effect is improved, and by introducing residual connections, the problem of gradient disappearance in deep networks is solved, and the training effect and performance of the network are improved.

[0131] Finally, the preset network model is applied to the test set for inference, and the denoised synthetic aperture radar image is output.

[0132] The network structure of the AOD-Network specifically includes:

[0133] Input layer: Receives high-precision synthetic aperture radar image blocks and single-bit synthetic aperture radar image blocks.

[0134] Multi-scale feature fusion DCNN layer: Extracts multi-scale features to enhance the model's capture of information at different scales.

[0135] ResNet layer: Introduces residual connections to improve the training effect of the network.

[0136] Output layer: Outputs the denoised synthetic aperture radar image.

[0137] S30. Determine the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image.

[0138] Specifically, before determining the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image, the method further includes:

[0139] S310. Perform restoration and recombination processing on the denoised synthetic aperture radar image, the high-precision synthetic aperture radar image, and the single-bit synthetic aperture radar image to obtain the denoised synthetic aperture radar image, the high-precision synthetic aperture radar image, and the single-bit synthetic aperture radar image of a preset size, and the preset size is larger than the size of the image before restoration and recombination.

[0140] The recombination and restoration operation specifically includes:

[0141] Set the expected size of the recombined image;

[0142] Determine the position of each image block in the recombined image according to the position information of the image block in the original image;

[0143] In the order of positions, splice the image blocks into a complete large-sized image. This operation is performed on the high-precision synthetic aperture radar (SAR) image blocks, single-bit SAR image blocks, and the denoised SAR image blocks respectively to obtain the corresponding large-sized images;

[0144] Among them, to process the edge effects that may occur during the splicing process and ensure the continuity and consistency between the image blocks, methods such as smoothing processing or transition region compensation can be adopted.

[0145] The preset size of the recombined image is usually larger than the size of the image blocks before restoration and recombination, so as to better evaluate the denoising effect of the network model. For the image block regions that cannot be completely covered due to insufficient size during the splicing process, cropping or zero-padding processing is performed to ensure that the size of the recombined image meets the preset requirements.

[0146] For example, through the restoration and recombination operation, a high-precision SAR image, a single-bit SAR image pair, and a denoised SAR image with a size of w×h are obtained. Among them, w and h respectively represent the width and height of the large-sized SAR image after recombination. Since the SAR image samples with a size less than r×r are removed, so w<M, h<N

[0147] Specifically, the process of determining the image loss value based on the denoised SAR image and the corresponding high-precision SAR image and single-bit SAR image specifically includes:

[0148] S320. Determine the mean square error loss of the high-precision SAR image, single-bit SAR image, and denoised SAR image;

[0149] Specifically, the process of determining the mean square error loss of the high-precision SAR image, single-bit SAR image, and denoised SAR image specifically includes:

[0150]

[0151] Among them, M represents the number of pixel rows of the high-precision SAR image, single-bit SAR image, and denoised SAR image, N represents the number of pixel columns of the high-precision SAR image, single-bit SAR image, and denoised SAR image, I represents the high-precision SAR image and single-bit SAR image, and K represents the denoised SAR image.

[0152] S321. Determine the contrast loss of the high-precision SAR image, single-bit SAR image, and denoised SAR image;

[0153] Specifically, the process of determining the contrast loss of the high-precision synthetic aperture radar image, the single-bit synthetic aperture radar image, and the denoised synthetic aperture radar image specifically includes:

[0154]

[0155] Among them, H represents the high-precision synthetic aperture radar image, S represents the single-bit synthetic aperture radar image, and the contrast

[0156]

[0157] Among them, M and N respectively represent the number of pixel rows and the number of pixel columns of the synthetic aperture radar image gray matrix. represents the pixel average value of the synthetic aperture radar image. represents the pixel average value of the m-th row and the n-th column in the synthetic aperture radar image gray matrix.

[0158] S322. Perform weighted summation on the mean square error loss and the contrast loss based on a preset weight coefficient to obtain an image loss value.

[0159] As Figure 3 shown, introduce a preset weight coefficient λ, which represents the weight coefficient of the contrast loss function. Among them, the initial value of λ is 0.1. The λ adjusts the weight coefficient in a decreasing manner at a certain step according to the denoising performance of the network training.

[0160] Perform weighted summation on the mean square error loss function and the contrast loss function to obtain a composite loss function:

[0161] L = L MSE + λL CON

[0162] Among them, L represents the image loss value after weighted summation.

[0163] S40. Train the preset network model and the preset weight coefficient based on the image loss value to obtain an image reconstruction network.

[0164] The training process specifically includes:

[0165] Initialize the network model:

[0166] Select a suitable neural network model as the preset network model. For example, the denoising network DnCNN, the defogging network AOD-Net, or the deblurring network SRN-DeblurNet. In the embodiments of the present application, AOD-Net is taken as an example for description.

[0167] Initialize the weight parameters of the network model, usually using the random initialization method or the pre-trained model for initialization.

[0168] Set an initial learning rate to control the step size of each parameter update; set the number of samples used in each iteration. The choice of batch size affects the stability of training and memory usage. Set the total number of training iterations. Usually, the total number of training iterations is adjusted according to the scale of the dataset and the complexity of the model.

[0169] Based on the training sample set obtained after classification processing in step S20, perform data augmentation on the training samples in the training sample set, such as random cropping, flipping, and rotation, to increase data diversity and improve the generalization ability of the model.

[0170] According to the composite loss function (including mean square error loss and contrast loss) set in step S30, calculate the image loss value in each training iteration; measure the difference between the denoised synthetic aperture radar image and the high-precision synthetic aperture radar image; evaluate the brightness and darkness difference of the denoised synthetic aperture radar image to ensure that the image maintains the original contrast characteristics.

[0171] Based on the calculated loss value, calculate the gradient of the loss value with respect to the network parameters through the backpropagation algorithm; use an optimization algorithm (such as Stochastic Gradient Descent SGD, Adam, etc.) to update the network parameters, gradually reduce the loss value, and improve the denoising performance of the model.

[0172] Dynamically adjust the learning rate according to the change of the loss value during training. For example, a learning rate decay strategy can be adopted to gradually reduce the learning rate after training for a certain stage to improve the stability and convergence of the model; adjust the weight coefficients in the composite loss function according to the contribution degrees of the contrast loss and the mean square error loss during training, so that the network reaches a balance between different loss terms and optimizes the overall performance.

[0173] When the training reaches the preset number of iterations or the loss value converges to the set threshold, terminate the training process; save the trained network model and its weight parameters for subsequent use. Usually, it is saved as a model file (such as HDF5 or ONNX format) for easy loading and application.

[0174] In addition, the embodiments of this application also include the following simulation experiments:

[0175] First, the RADARSAT-1 spaceborne data in Region W was used for preliminary experimental analysis. The raw echo data was processed using the R-D algorithm to obtain a high-precision / single-bit synthetic aperture radar (SAR) image pair with a size of 9286×18432. After overlapping separation and scene classification, a spaceborne SAR dataset for the noise reduction network was obtained. Among them, the sample scenes included mountains, waters, cities, farmlands, etc. The noise reduction performances of the DnCNN with the improved loss function, the AOD-Net with the improved loss function, and the SRN-DeblurNet network with the improved loss function were evaluated based on three evaluation metrics: contrast, information entropy, and equivalent number of looks. And the quality of the restored and recombined SAR images was compared.

[0176] Figure 4 The noise reduction performances of each reconstruction network model under different experimental scenarios. From top to bottom are the river channel, city, port, airport, and farmland scenes. From left to right are the high-precision SAR image, single-bit SAR image, SAR image after noise reduction by the AOD-Net with the improved loss function, SAR image after noise reduction by the DnCNN with the improved loss function, and SAR image after noise reduction by the SRN-DeblurNet with the improved loss function for each scene. It can be seen from the image clarity and the degree of detail retention that the image after noise reduction by the DnCNN network with the improved loss function shows an oversaturation phenomenon, especially in the port area; the contrast of the image processed by the SRN-DeblurNet network with the improved loss function is partially improved, but obvious details are lost, such as the track at the airport basically disappearing after processing; while the AOD-Net network with the improved loss function can relatively completely retain the scene information on the premise of removing harmonic noise, and the image clarity is higher.

[0177] Figure 5 The result diagram of the AOD-Net network for reducing speckle noise. From Figure 4 it can be seen that the AOD-Net network with the improved loss function has a strong ability to reduce harmonic noise. At the same time, from the Figure 4 port scene, it can be seen that for the SAR image processed by the AOD-Net network with the improved loss function compared with the original high-precision SAR image, the sidelobe interference caused by speckle noise is significantly reduced. That is to say, the AOD-Net network with the improved loss function can not only reduce harmonic noise but also suppress the speckle noise interference inherent in the SAR image, improving the quality of the high-precision SAR image.

[0178] The following table shows the comparison of the evaluation index results of different network models. The noise reduction performance of different network models is evaluated based on contrast, information entropy, and equivalent number of looks. It can be seen from the table that the indicators of the synthetic aperture radar (SAR) images after noise reduction by the three networks are all better than those of the single-bit SAR images, and the image indicators processed by the AOD-Net with the improved loss function and the DnCNN network with the improved loss function are even better than those of the high-precision SAR images. Compared with the DnCNN network with the improved loss function, the image processed by the AOD-Net network with the improved loss function has higher contrast, lower information entropy and equivalent number of looks, more obvious scene details, and better image quality.

[0179]

[0180] Figure 6 represents the AOD-Net network structure, including the K estimation and image restoration modules. The K estimation module estimates the K(x) parameter using the foggy image, and the image restoration module adaptively estimates the defogged image using the K(x) parameter. Among them, the structure of the K estimation module is as Figure 7 shown. Figure 8 represents that the multi-scale feature fusion CNN of the K estimation module of the AOD-Net network is improved to a multi-scale feature fusion DCNN network, Figure 9 represents the K estimation module after introducing the residual network ResNet on the basis of the multi-scale feature fusion DCNN network.

[0181] Figure 10 is the comparison chart of the noise reduction results of different K estimation modules. From left to right are the high-precision SAR image, the SAR image processed by the original AOD-Net network, the improved AOD-Net processed by the multi-scale DCNN network as the K estimation module, and the improved AOD-Net processed by the DCNN network combined with the residual neural network ResNet as the K estimation module. It can be seen that the quality of the SAR image has been partially improved, especially Figure 11 as can be seen from, after replacing the multi-scale feature fusion CNN network of the original K estimation module with the multi-scale DCNN network, the noise reduction ability in the sea area has been improved to a certain extent.

[0182] The following table shows the comparison of the evaluation index results of different K estimation modules. It can be seen that after improving the K estimation module of the AOD-Net network to the multi-scale feature fusion DCNN, all indicators have been improved, the quality of the SAR image has been improved, and it also verifies the Figure 6 conclusion of the enhanced noise reduction ability in the sea area in, further reducing the loss of scene details.

[0183]

[0184]

[0185] Next, a small unmanned aerial vehicle (UAV) equipped with a radar was used to collect synthetic aperture radar (SAR) data and perform single-bit imaging on different experimental scenarios within A University. The above-mentioned overlapping block division and statistical classification were carried out on high-precision / single-bit SAR image pairs to obtain a measured dataset of UAV-borne SAR for the noise reduction network. Among them, the experimental scenarios included College B, Wenshan Lake, Library, Gymnasium, Flower Beds, etc. According to the noise reduction performance of the above different network models on the RADARSAT-1 dataset, the AOD-Net network based on the improved loss function and the AOD-Net network with the improved K estimation module were used to process the measured dataset of UAV-borne SAR. Among them, the AOD-Net network with the improved K estimation module means that the K estimation module of the AOD-Net network is improved to a multi-scale feature fusion DCNN network, which will be uniformly described below.

[0186] Figure 12 For the noise reduction performance of the AOD-Net network with the improved loss function and the AOD-Net network with the improved K estimation module in different scenarios, it can be clearly seen that the contrast and clarity of the SAR images processed by the two networks are significantly improved compared with the high-precision images, and the detail features of the AOD-Net network with the improved K estimation module are clearer around the target, such as Figure 9 shown.

[0187] Figure 13 Figure for comparing the noise reduction results of the AOD-Net with the improved loss function and the AOD-Net with the improved K estimation module in College B. It can be seen that compared with the AOD-Net network with the improved loss function, the contour lines and edge features of the building targets in the SAR image processed by the AOD-Net network with the improved K estimation module are clearer, and the focusing effect is better.

[0188] Figure 14 Figure for the result of restoring and recombining several high-precision / single-bit SAR image blocks in Campus a of A University. It can be seen that although there is harmonic noise interference in single-bit quantization, a high-definition SAR image can still be obtained. After noise reduction by the network model, the image clarity and contrast can be significantly improved on the basis of less loss of target detail features, realizing high-quality imaging at a low sampling rate.

[0189] The following table shows the comparison of the evaluation index results of the synthetic aperture radar images processed by the AOD-Net with the improved loss function and the AOD-Net network with the improved K estimation module. As can be seen from Table 3, the synthetic aperture radar image denoised by the AOD-Net network with the improved K estimation module has higher contrast, lower information entropy and equivalent number of looks, reduces the interference of high-order harmonic noise, suppresses the influence of speckle noise, and improves the quality of the original high-precision synthetic aperture radar image, thus verifying the reliability of the denoising ability of the AOD-Net network with the improved K estimation module.

[0190]

[0191]

[0192] This embodiment provides a method for using single-bit synthetic aperture radar imaging reconstruction based on an improved loss function, as Figure 1 shown, the method includes:

[0193] H10. Obtain synthetic aperture radar image data;

[0194] H20. Input the synthetic aperture radar image data into a single-bit synthetic aperture radar imaging reconstruction network based on an improved loss function for enhanced radar imaging processing to obtain high-quality synthetic aperture radar image data.

[0195] Specifically, in step H10:

[0196] A synthetic aperture radar (SAR) system is started on a carrier platform (such as a satellite, drone, or aircraft) and begins to work;

[0197] The synthetic aperture radar system transmits a chirp signal, and the signal is reflected after encountering a ground object;

[0198] The reflected signal is captured by the receiving antenna of the radar system to form an echo signal;

[0199] The received echo signal is preprocessed to remove environmental noise and system noise;

[0200] The echo signal is dechirped to compensate for the frequency change caused by platform movement;

[0201] The de-noised and de-chirped echo signal is sampled at a preset sampling rate and quantized using a multi-bit analog-to-digital converter (usually 8 bits, 12 bits, or 16 bits) to obtain high-precision echo data;

[0202] Based on the high-precision quantization data, the sign bit is extracted and single-bit quantization processing is performed to generate single-bit echo data;

[0203] Generate synthetic aperture radar images based on synthetic aperture radar imaging algorithms: Use synthetic aperture radar imaging algorithms such as the back projection algorithm (BP) and the range Doppler algorithm (RD) to process high-precision quantization data and single-bit quantization data, and generate high-precision synthetic aperture radar images and single-bit synthetic aperture radar images;

[0204] In step H20:

[0205] Select a suitable single-bit synthetic aperture radar imaging reconstruction network model based on an improved loss function, such as the denoising network DnCNN, the defogging network AOD-Net, or the deblurring network SRN-DeblurNet. In a specific embodiment of the present application, the trained AOD-Net is used as the network reconstruction network. The trained AOD-Net K estimation module adopts a multi-scale feature fusion network (DCNN) and a residual neural network (ResNet) to enhance the noise reduction and image reconstruction performance of the model;

[0206] Preprocess the generated high-precision synthetic aperture radar images and single-bit synthetic aperture radar images according to the input format of the network model, such as normalization processing, and adjust the pixel values to the range of [0,1]; input the preprocessed high-precision synthetic aperture radar images and single-bit synthetic aperture radar images into the image reconstruction network.

[0207] Use the trained image reconstruction network to perform inference on the input synthetic aperture radar image data to generate a high-quality synthetic aperture radar image after noise reduction;

[0208] Evaluate the output high-quality synthetic aperture radar image to ensure that the clarity and contrast of the image meet the expected requirements; after obtaining the output high-quality synthetic aperture radar image, the output high-quality synthetic aperture radar image data can also be stored in a preset format for subsequent use and analysis.

[0209] Based on the above single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function, the present application also provides a terminal device, such as Figure 15 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communication interface 23 and a bus 24. Among them, the processor 20, the display screen 21, the memory 22, and the communication interface 23 can communicate with each other through the bus 24. The display screen 21 is set to display a user guidance interface preset in the initial setting mode. The communication interface 23 can transmit information. The processor 20 can call the logical instructions in the memory 22 to execute the method in the above embodiment.

[0210] In addition, when the logical instructions in the above-mentioned memory 22 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0211] As a computer-readable storage medium, the memory 22 can be configured to store software programs and computer-executable programs, such as the program instructions or modules corresponding to the methods in the embodiments of the present disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, that is, implements the methods in the above embodiments.

[0212] The memory 22 may include a program storage area and a data storage area. Among them, the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory 22 may include high-speed random access memory and may also include non-volatile memory. For example, various media that can store program codes such as USB flash drives, external hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs may also be transient storage media.

[0213] In addition, the specific processes of loading and executing multiple instructions by the instruction processor in the above storage medium and terminal device have been described in detail in the above methods and will not be repeated here one by one.

[0214] In summary, this embodiment provides a single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function. The method reduces the data volume through a single-bit quantization strategy, prevents high-order harmonic aliasing, and achieves high-quality imaging at a low sampling rate; by introducing three no-reference evaluation indicators of contrast, information entropy, and equivalent number of looks to guide network training and improve the noise reduction effect and image quality; using an improved AOD-Net network model, enhancing the noise reduction performance of the model through a multi-scale feature fusion network and a residual neural network, achieving high-quality image reconstruction, and based on data classification and block processing, ensuring the continuity and integrity between image blocks, improving the diversity and quality of training data, and optimizing the parameters of the network model through the weighted sum of mean square error loss and contrast loss to improve the overall noise reduction performance.

[0215] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function, characterized in that The method includes: Obtain a training sample set, where the training sample set includes high-precision synthetic aperture radar images and single-bit synthetic aperture radar images; Input the training samples of the training sample set into a preset network model, and output a denoised synthetic aperture radar image through the preset network model; Determine an image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image; Train the preset network model and preset weight coefficients based on the image loss value to obtain an image reconstruction network; The process of determining the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image specifically includes: Determine the mean square error loss of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image; Determine the contrast loss of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image; Perform weighted summation on the mean square error loss and contrast loss based on a preset weight coefficient to obtain an image loss value.

2. The single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function according to claim 1, wherein The process of obtaining the training sample set specifically includes: Obtain radar echo signals, perform sampling processing on the echo signals based on a preset sampling rate to obtain sampled echo signals, and perform quantization processing on the sampled echo signals to obtain radar echo data, where the radar echo data is single-bit quantized echo data; Process the radar echo data based on a preset synthetic aperture radar imaging algorithm to obtain a single-bit synthetic aperture radar image, where the single-bit synthetic aperture radar image is a high-precision synthetic aperture radar image and a single-bit synthetic aperture radar image of a preset size; Perform block processing on the high-precision synthetic aperture radar image and single-bit synthetic aperture radar image to obtain high-precision synthetic aperture radar images and single-bit synthetic aperture radar images of a preset size; Perform overlapping block processing on the small high-precision synthetic aperture radar images and single-bit synthetic aperture radar images of the preset size to obtain a first synthetic aperture radar image data set; Perform classification processing on the first synthetic aperture radar image data set based on a preset scenario to obtain a training sample set.

3. The single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function according to claim 1, wherein The k estimation module of the preset network model includes a multi-scale feature fusion network and a residual neural network.

4. The single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function according to claim 1, wherein The process of determining the mean square error loss of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image specifically includes: Where M represents the number of pixel rows of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image, N represents the number of pixel columns of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image, I represents the high-precision synthetic aperture radar image and single-bit synthetic aperture radar image, and K represents the denoised synthetic aperture radar image.

5. The single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function according to claim 1, characterized in that The process of determining the contrast loss of the high-precision synthetic aperture radar image, single-bit synthetic aperture radar image, and denoised synthetic aperture radar image specifically includes: where H represents the high-precision synthetic aperture radar image, S represents the single-bit synthetic aperture radar image, and the contrast CON is where M and N respectively represent the number of pixel rows and the number of pixel columns of the gray matrix of the synthetic aperture radar image, represents the pixel average value of the synthetic aperture radar image, represents the pixel average value of the m-th row and the n-th column in the gray matrix of the synthetic aperture radar image.

6. The single-bit synthetic aperture radar imaging reconstruction training method based on an improved loss function according to claim 1, characterized in that Before determining the image loss value based on the denoised synthetic aperture radar image and the corresponding high-precision synthetic aperture radar image and single-bit synthetic aperture radar image, the method further includes: Performing restoration and recombination processing on the denoised synthetic aperture radar image, high-precision synthetic aperture radar image, and single-bit synthetic aperture radar image to obtain denoised synthetic aperture radar images, high-precision synthetic aperture radar images, and single-bit synthetic aperture radar images of a preset size, where the preset size is larger than the size of the images before restoration and recombination.

7. A method for using single-bit synthetic aperture radar imaging reconstruction based on an improved loss function, characterized in that Applied to the image reconstruction network obtained by the single-bit synthetic aperture radar imaging reconstruction training method based on the improved loss function according to any one of claims 1-6, the method includes: Obtaining synthetic aperture radar image data; Inputting the synthetic aperture radar image data into the image reconstruction network obtained by the single-bit synthetic aperture radar imaging reconstruction training method based on the improved loss function for enhanced radar imaging processing to obtain high-quality synthetic aperture radar image data.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps in the single-bit synthetic aperture radar imaging reconstruction training method based on the improved loss function according to any one of claims 1-6.

9. A terminal device, characterized in that, Including: A processor and a memory; The memory stores a computer-readable program executable by the processor; When the processor executes the computer-readable program, it implements the steps in the single-bit synthetic aperture radar imaging reconstruction training method based on the improved loss function according to any one of claims 1-6.

Images