Single-bit imaging method and system based on two-dimensional spectrum reconstruction of dequantized convolutional network
Through the method based on dequantization convolution network, the problems of high-complexity calculation and sampling rate improvement in two-dimensional spectrum reconstruction of single-bit signal are solved, and efficient single-bit SAR imaging is achieved, simplifying the system architecture and improving the imaging quality.
Patent Information
- Application Number
- CN202310208195.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-24
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2043-02-24
AI Technical Summary
The prior art cannot realize high-complexity calculations and improve sampling rates in single-bit signal two-dimensional spectrum reconstruction, resulting in increased system architecture complexity and data processing pressure.
Using a method based on dequantization convolutional network, the linear frequency modulation signal is obtained for demodulation and sampling, and a single-bit complex signal is obtained, and the Fourier transform and the trained dequantization convolutional network are used to predict Doppler spectrum. , realize high-order harmonic suppression and signal two-dimensional spectrum reconstruction.
Implement high-complexity calculations at high sampling rates, reduce system architecture complexity, improve imaging quality, and simplify the single-bit SAR imaging process.
Smart Images

Figure CN116520322B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of signal processing technology, and in particular to a single-bit imaging method and system based on dequantized convolutional network two-dimensional spectrum reconstruction. Background Art
[0002] The reconstruction of the two-dimensional spectrum of a single-bit signal can be applied to target positioning, instantaneous SAR (SAR: Synthetic Aperture Radar, Chinese name is synthetic aperture radar, which is an active ground observation system that can be installed on aircraft, satellites, spacecraft and other flying platforms to observe the ground all day and all weather and has a certain surface penetration ability) imaging and other application scenarios. It is a new radar working mode based on single-bit sampling. Single-bit sampling is a highly nonlinear expression method for symbol quantization of radar echo signals. Different from the traditional analog to digital converter (ADC) that performs uniform threshold comparison on SAR echo signals, the single-bit comparator compares the amplitude of the echo signal with the preset threshold once, retains the sign bit of the comparison result and returns it to the memory. Compared with traditional radar technology, single-bit radar technology can not only significantly reduce the cost of data acquisition, storage, transmission and processing, but also has the advantages of simple architecture, low cost and easy miniaturization. Existing single-bit reconstruction technologies are mainly divided into two categories: one is sparse signal recovery based on optimization theory to characterize the scattering characteristics of the target scene to be measured and complete tasks such as SAR imaging and target positioning; the other is harmonic suppression and separation based on threshold strategy to reconstruct the effective echo signal of the baseband.
[0003] Under the prior condition that the signal is sparse, the first single-bit compressed sensing reconstruction algorithm (Renormalized Fixed-Point Iteration, RFPI) was proposed. In the framework of compressed sensing, the single-bit measurement matrix is equivalent to the perception matrix, and the single-bit SAR imaging is also converted into an underdetermined problem (underconstrained problems are also called underdetermined problems, which have infinite solutions) of restoring the original signal under the condition of known measurement matrix and single-bit SAR echo signal. In order to solve such difficult problems, a series of reconstruction methods, such as greedy pursuit, linear programming, hard threshold iteration, etc., have been proposed and successfully applied to the field of single-bit SAR imaging. Specifically, Matching Sign Pursuit (MSP) selects a compact support set (Compact set) through a greedy algorithm and reconstructs the signal within the support set. The 1-bit Gradient Support Pursuit (1-bit GraSP) algorithm selects the support set according to the iterative gradient information, tries to find a more complete single-bit quantized observation space, and achieves high-quality SAR imaging. In addition, some people have proposed iterative hard thresholding algorithms (Binary Iterative Hard Thresholding, BIHT and BIHT2), and given the conditions for stable reconstruction of single-bit compressed sensing. These methods have also improved the performance of single-bit SAR imaging algorithms to a certain extent, but they cannot avoid the huge pressure on platform data processing caused by the introduction of sparse optimization methods. The actual test scene does not always meet the prior assumptions of sparse imaging. At the same time, the observation dictionary of single-bit compressed sensing is a linear representation and cannot fully represent the mapping space of single-bit quantization effects. Therefore, single-bit compressed sensing theory has certain limitations in SAR imaging applications.
[0004] In recent years, some scholars have introduced different threshold strategies to enhance the gain of SAR echo signals, compare the received signal with the preset reference signal, and retain the diversity of the signal spectrum at the cost of increasing the sampling rate. A wider spectrum can better suppress high-order harmonics to complete the SAR imaging task. Existing single-bit threshold strategies include: Single Frequency Thresholding (SFT) effectively retains the relative amplitude information lost during 1-bit sampling by introducing a periodic signal whose initial phase amplitude changes with time as a reference signal, while avoiding the introduction of noise-like interference, effectively improving the quality of SAR imaging under single-bit sampling. Shift Frequency (SF) uses the frequency shift characteristics of the reference signal to shift the baseband signal into the passband of the filter, and uses as much effective information as possible for imaging. Random Threshold (RT) alleviates problems such as defocusing caused by harmonics in single-bit SAR imaging by changing the distribution probability of high-order harmonic amplitudes. Although the threshold-based single-bit SAR imaging algorithm has good imaging effects, this type of method needs to increase the sampling rate of the single-bit signal to avoid harmonic aliasing, which brings about the problem of high-precision storage of the quantization threshold and cannot effectively simplify the system architecture; in addition, the setting of the threshold signal parameters also requires accurate prior knowledge and manual modulation, which increases the complexity of the single-bit SAR imaging algorithm to a certain extent.
[0005] Therefore the prior art still needs to be improved and enhanced. Summary of the invention
[0006] The main purpose of the present invention is to provide a single-bit imaging method and system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, aiming to solve the problem in the prior art that it is impossible to achieve high-complexity calculation requirements and increase the sampling rate when reconstructing the two-dimensional spectrum of a single-bit signal.
[0007] In order to achieve the above object, the present invention adopts the following technical solutions:
[0008] A single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network comprising:
[0009] Acquire a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal;
[0010] According to the single-bit complex signal and a preset pulse repetition period, a single-bit complex echo data matrix is acquired, and the single-bit complex echo data matrix is subjected to Fourier transform processing to obtain a first range Doppler spectrum;
[0011] Inputting the first range Doppler spectrum into a trained dequantized convolutional network to obtain a second range Doppler spectrum;
[0012] The second range Doppler spectrum is subjected to coordinate transformation to obtain an instantaneous synthetic aperture radar imaging image.
[0013] In the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the steps of acquiring a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, and sampling the analog signal to obtain a single-bit complex signal also include:
[0014] Preset the parameters of the transmitting radar according to the size of the scene to be tested; wherein the parameters include radar bandwidth, frequency modulation, time width and repetition rate;
[0015] The transmitting radar transmits a linear frequency modulation signal according to the parameters.
[0016] In the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the step of acquiring a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, and sampling the analog signal to obtain a single-bit complex signal specifically includes:
[0017] Performing orthogonal demodulation processing on the linear frequency modulation signal to obtain the analog signal;
[0018] The analog signal is single-bit sampled according to a preset linear frequency modulation signal sampling rate to obtain a single-bit complex signal of a first dimension; wherein the first dimension is N×1.
[0019] In the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the single-bit complex echo data matrix is obtained according to the single-bit complex signal and a preset pulse repetition period, and the single-bit complex echo data matrix is subjected to Fourier transform processing to obtain a first range Doppler spectrum, which specifically includes:
[0020] When the single-bit complex signal is sampled and processed at the pulse repetition period, a single-bit complex echo data matrix of a second dimension is obtained; wherein the second dimension is N×M;
[0021] The single-bit complex echo data matrix is subjected to Fourier transform processing in the column direction and the row direction respectively to obtain the first range-Doppler spectrum; wherein the first range-Doppler spectrum is a range-Doppler spectrum of single-bit quantized echo data.
[0022] In the single-bit imaging method based on two-dimensional spectrum reconstruction using a dequantized convolutional network, the first range Doppler spectrum is input into a trained dequantized convolutional network to obtain a second range Doppler spectrum, and the method also includes:
[0023] Presetting the structural parameters of the convolution layer, and constructing the convolution layer based on the structural parameters; wherein the structural parameters include: the size of the convolution kernel, the stride size, and the boundary expansion method;
[0024] Splicing the cascade layer into the convolution layer to construct a harmonic feature estimation module;
[0025] According to the point-by-point multiplication layer and the point-by-point addition layer, a reconstructed spectrogram generation module is constructed;
[0026] Constructing and generating the dequantized convolutional network according to the harmonic feature estimation module and the reconstructed spectrum generation module;
[0027] The dequantized convolutional network is trained and learned according to preset training parameters to obtain a trained dequantized convolutional network.
[0028] In the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the dequantized convolutional network is trained and learned according to preset training parameters to obtain a trained dequantized convolutional network, which specifically includes:
[0029] respectively setting a first training parameter in an initial training phase and a second training parameter in a fine tuning phase;
[0030] In the initial training phase, performing a first iterative update operation on the parameter matrix to be learned using the first training parameters;
[0031] When the error value of the adjacent loss functions is less than the preset difference, the second training parameter is used to perform a second iterative update operation on the parameter matrix to be learned after the first iterative update to obtain a trained dequantized convolutional network.
[0032] In the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the first training parameters include: the number of initial training times, the initial learning rate in the initial stage, and the number of initial training iterations; the second training parameters include: the number of fine-tuning training times, the initial learning rate in the fine-tuning stage, and the number of fine-tuning training iterations.
[0033] A single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network comprising:
[0034] A demodulation sampling module is used to obtain a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal;
[0035] A first range Doppler spectrum acquisition module, used to acquire a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, and perform Fourier transform processing on the single-bit complex echo data matrix to obtain a first range Doppler spectrum;
[0036] A second range Doppler spectrum acquisition module, used for inputting the first range Doppler spectrum into the trained dequantized convolutional network to obtain a second range Doppler spectrum;
[0037] The coordinate conversion module is used to perform coordinate conversion on the second range Doppler spectrum to obtain an instantaneous synthetic aperture radar imaging image.
[0038] A controller comprises: a memory, a processor, a display, and a single-bit imaging program based on two-dimensional spectrum reconstruction of a dequantized convolutional network, which is stored in the memory and can be run on the processor. When the single-bit imaging program based on two-dimensional spectrum reconstruction of a dequantized convolutional network is executed by the processor, the steps of the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network as described above are implemented.
[0039] A computer-readable storage medium stores a single-bit imaging program based on two-dimensional spectrum reconstruction of a dequantized convolutional network. When the single-bit imaging program based on two-dimensional spectrum reconstruction of a dequantized convolutional network is executed by a processor, the steps of the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network as described above are implemented.
[0040] Compared with the prior art, the present invention provides a single-bit imaging method and system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the method comprising: obtaining a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, sampling the analog signal to obtain a single-bit complex signal; obtaining a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, Fourier transforming the single-bit complex echo data matrix to obtain a first range Doppler spectrum; inputting the first range Doppler spectrum into a trained dequantized convolutional network to obtain a second range Doppler spectrum; performing coordinate transformation on the second range Doppler spectrum to obtain an imaging image of an instantaneous synthetic aperture radar. After demodulating, sampling and Fourier transforming the linear frequency modulation signal to obtain the first range Doppler spectrum, the first range Doppler spectrum is output to the trained dequantized convolutional network for prediction to obtain a second range Doppler spectrum, thereby realizing high-complexity calculation at a high sampling rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A flowchart of a preferred embodiment of the single-bit imaging method based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0042] Figure 2 A flowchart of step S100 in a preferred embodiment of the single-bit imaging method based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0043] Figure 3 A flowchart of step S300 in a preferred embodiment of the single-bit imaging method based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0044] Figure 4 A schematic diagram of node splitting and node aggregation provided by the present invention;
[0045] Figure 5 A flowchart of step S400 in a preferred embodiment of the single-bit imaging method based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0046] Figure 6 A schematic diagram of computing graph segmentation in the data parallel mode provided by the present invention;
[0047] Figure 7 A schematic diagram of computing graph segmentation in the pipeline parallel mode provided by the present invention;
[0048] Figure 8 The principle block diagram of the single-bit imaging system based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0049] Fig. 9 The architecture relationship diagram of the single-bit imaging system based on dequantized convolutional network two-dimensional spectrum reconstruction provided by the present invention;
[0050] Fig.10 A schematic diagram of the operating environment of a preferred embodiment of the controller provided by the present invention;
[0051] Fig.11 Schematic diagram of the operating environment of a preferred embodiment of the controller of the present invention. DETAILED DESCRIPTION
[0052] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0053] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the 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 groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.
[0054] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.
[0055] In order to facilitate understanding of the embodiments of the present application, relevant elements involved in the embodiments of the present invention are first introduced here.
[0056] The dequantized convolutional network module is an adaptive single-bit range Doppler (RD) recovery module that can characterize the characteristics of high-order harmonics from any input single-bit radar echo signal. As a typical deep discriminative network module, the convolutional neural network (CNNs) has rekindled the craze for a new generation of artificial intelligence in recent years. CNNs are deep network structures inspired by the structure of the visual cortex of the brain, and are mainly composed of convolutional layers and downsampling layers. The multidimensional convolutional layer can filter samples with different functions, and the downsampling layer (pooling) sparsely samples the filtered information by simulating the visual system to achieve compressed information acquisition; finally, the posterior probability is calculated based on the data category described by the information, and the discriminative ability of the model is provided based on this.
[0057] The present invention provides a single-bit imaging method and system based on two-dimensional spectrum reconstruction of a dequantized convolutional network. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network is a two-dimensional spectrum reconstruction technology of instantaneous single-bit signals based on a dequantized convolutional network, which can be used for single-bit instantaneous SAR imaging, single-bit target positioning, and single-bit signal spectrum estimation.
[0058] In the present invention, after demodulating the preset linear frequency modulation signal, the obtained analog signal is sampled to obtain a single-bit complex signal, and then a single-bit complex echo data matrix is obtained according to the single-bit complex signal and the preset pulse repetition period, and then Fourier transform processing is performed to obtain a first range Doppler spectrum. The first range Doppler spectrum is input into the trained dequantization convolution network to obtain a second range Doppler spectrum. The second range Doppler spectrum is then coordinate-converted to obtain an imaging image of an instantaneous synthetic aperture radar. In the present invention, by describing the characteristics of high-order harmonics associated with a single bit in a two-dimensional spectrum domain, a harmonic estimation module and a reconstructed spectrum generation module in a dequantization convolution network are used to achieve high-order harmonic suppression, and complete the task of reconstructing the two-dimensional spectrum of the signal, thereby achieving high-complexity calculations at a high sampling rate.
[0059] The following describes the design of a single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network through specific exemplary embodiments. It should be noted that the following embodiments are only used to explain the technical solution of the invention and are not specifically limited:
[0060] See also Figure 1 The present invention provides a single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network, and the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network comprises:
[0061] S100: Acquire a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal.
[0062] Specifically, after a preset linear frequency modulation signal is sent by a transmitting radar, the preset linear frequency modulation signal is first demodulated to obtain the analog signal (analog signal), and then the analog signal is sampled to obtain the single-bit complex signal, so that the single-bit complex signal can be further processed.
[0063] In the present invention, the linear frequency modulation signal is demodulated and sampled to obtain the single-bit complex signal, so as to be used for describing the two-bit spectrum domain.
[0064] For further information, see Figure 2 The step S100, obtaining a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, sampling the analog signal to obtain a single-bit complex signal, also includes:
[0065] A10. Preset the parameters of the transmitting radar according to the size of the scene to be measured; wherein the parameters include radar bandwidth, modulation frequency, time width and repetition rate;
[0066] A20. The transmitting radar transmits a linear frequency modulation signal according to the parameters.
[0067] Specifically, the steps of using the transmitting radar to send a preset linear frequency modulation signal are as follows:
[0068] First, according to the size of the scene to be tested, the radar bandwidth, the modulation frequency, the time width, the repetition rate and other parameters are pre-set. For example, the radar bandwidth can be set to 4GHz, the modulation frequency can be set to 10e13Hz / s, the time width can be set to 0.2ms, and the repetition rate can be set to 5000Hz. Then, the transmitting radar transmits a linear frequency modulation signal at a fixed repetition rate.
[0069] For further information, see Figure 3 The step S100, obtaining a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, and sampling the analog signal to obtain a single-bit complex signal, specifically includes:
[0070] S110, performing orthogonal demodulation processing on the linear frequency modulation signal to obtain the analog signal;
[0071] S120 , performing single-bit sampling on the analog signal according to a preset linear frequency modulation signal sampling rate to obtain a single-bit complex signal of a first dimension; wherein the first dimension is N×1.
[0072] Specifically, after the preset linear frequency modulation signal is sent by the transmitting radar, the linear frequency modulation signal is orthogonally demodulated to obtain two analog signals I and Q, that is, the analog signal is obtained, wherein the analog signal includes two analog signals I and Q. Then, according to the preset linear frequency modulation signal sampling rate f s Single-bit sampling is performed on the I and Q signals respectively to obtain a single-bit complex signal with a dimension of the first dimension N×1.
[0073] For further information, please refer to Figure 1 , S200, according to the single-bit complex signal and the preset pulse repetition period, obtain a single-bit complex echo data matrix, and perform Fourier transform processing on the single-bit complex echo data matrix to obtain a first range-Doppler spectrum; wherein the first range-Doppler spectrum is a range-Doppler spectrum of single-bit quantized echo data.
[0074] Specifically, after the analog signal is sampled to obtain the single-bit complex signal, the single-bit complex signal is subjected to a preset M pulse repetition period to obtain the single-bit complex echo data matrix S oThen, the single-bit complex echo data matrix is subjected to Fourier transform processing to obtain the first range-Doppler spectrum (range-Doppler spectrum of single-bit quantized echo data) Z o .
[0075] In the present invention, the pulse repetition period and Fourier transform are performed on the single-bit complex signal, thereby effectively describing the characteristics of the high-order harmonics associated with the single bit in the two-dimensional spectrum domain.
[0076] For further information, see Figure 4 The step S200, obtaining a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, and performing Fourier transform processing on the single-bit complex echo data matrix to obtain a first range Doppler spectrum, specifically includes:
[0077] S210, after the single-bit complex signal is sampled and processed at the pulse repetition period, a single-bit complex echo data matrix of a second dimension is obtained; wherein the second dimension is N×M;
[0078] S220, performing Fourier transform processing on the single-bit complex echo data matrix in the column direction and the row direction respectively to obtain the first range Doppler spectrum.
[0079] Specifically, after obtaining the single-bit complex signal, the single-bit complex signal is subjected to the pulse repetition period to obtain a single-bit complex echo data matrix S with a second dimension of N×M. o Then, the single-bit complex echo data matrix S o Fourier transform processing is performed in the column direction and the row direction respectively to obtain a range-Doppler spectrum (first range-Doppler spectrum) Zo of single-bit quantized echo data.
[0080] For further information, please refer to Figure 1 , S300, inputting the first range Doppler spectrum into the trained dequantized convolutional network to obtain a second range Doppler spectrum.
[0081] Specifically, after Fourier transform processing is performed on the single-bit complex echo data matrix to obtain the first range Doppler spectrum, the first range Doppler spectrum is input into the trained dequantized convolutional network for prediction, and the second range Doppler spectrum is calculated by solving the following minimization problem, that is, the reconstructed two-dimensional spectrum (range-Doppler spectrum) is calculated.
[0082] In the present invention, the first range Doppler spectrum is input into a trained dequantized convolutional network for prediction to obtain the second range Doppler spectrum, thereby achieving suppression of high-order harmonics and effectively improving the imaging quality of the reconstructed two-dimensional spectrum.
[0083] For further information, see Figure 5 , the S300, inputting the first range Doppler spectrum into the trained dequantized convolutional network to obtain a second range Doppler spectrum, and the above also includes:
[0084] B10, presetting structural parameters of the convolution layer, and constructing the convolution layer based on the structural parameters; wherein the structural parameters include: the size of the convolution kernel, the stride size, and the boundary expansion method;
[0085] B20, splicing the cascade layer into the convolution layer to construct a harmonic feature estimation module;
[0086] B30, constructing a reconstructed spectrogram generation module according to the point-by-point multiplication layer and the point-by-point addition layer;
[0087] B40, constructing and generating the dequantized convolutional network according to the harmonic feature estimation module and the reconstructed spectrum generation module;
[0088] B50. Train and learn the dequantized convolutional network according to preset training parameters to obtain a trained dequantized convolutional network.
[0089] Specifically, before using the trained dequantized convolutional network to predict the second range Doppler spectrum, the construction and training process of the dequantized convolutional network is divided into two steps. The first step is to construct the dequantized convolutional network as follows:
[0090] Among them, the dequantization convolution network is composed of the harmonic feature estimation module and the reconstructed spectrum generation module.
[0091] The harmonic feature estimation module is composed of a convolutional layer and a concatenated layer. The structural diagram of the convolutional layer and the concatenated layer is as follows: Figure 6 shown.
[0092] Then, the structural parameters of the convolution layer (the size of the convolution kernel, the stride size, and the boundary expansion method) are set in advance, and the convolution layer is constructed using the structural parameters, such as Figure 6As shown in the figure, conv1, conv2, conv3, conv4, conv5 and conv6 all represent neural network layers with convolutional structures. The structural parameters of the convolutional layers are: the convolution kernel size (the receptive field of the convolution operation) is 1, 3, 5, 7, 9, 1, respectively, the stride size is set to 1 by default, and the boundary expansion method is set, for example, the boundary is expanded in a standard way (zero padding operation is performed on the image boundary) to ensure that the spatial size of the output data is consistent with the input. In addition, concat1, concat2 and concat3 all represent the cascade layer of intermediate features (the features are spliced by skipping layers to obtain better local feature extraction functions).
[0093] Then, the cascade layer is spliced into the convolutional layer to construct the harmonic feature estimation module, which can approximate the characteristics of high-order harmonics. Using massive data sets for training can realize the characterization of the aliasing and deformation characteristics of high-order harmonics in the spectral domain, which is one of the key steps in single-bit instantaneous SAR (synthetic aperture radar) imaging.
[0094] Furthermore, the reconstructed spectrogram generation module is constructed by using the point-by-point multiplication layer and the point-by-point addition layer. The dequantized convolutional network is constructed and generated according to the harmonic feature estimation module and the reconstructed spectrogram generation module. Finally, the dequantized convolutional network is trained and learned according to preset training parameters to obtain a trained dequantized convolutional network.
[0095] For further information, see Figure 7 , the B50, training and learning the dequantized convolutional network according to preset training parameters to obtain a trained dequantized convolutional network, specifically includes:
[0096] B51, respectively setting the first training parameters in the initial training phase and the second training parameters in the fine-tuning phase;
[0097] B52. In the initial training phase, using the first training parameters to perform a first iterative update operation on the learning parameter matrix;
[0098] B53. When the error value of the adjacent loss functions is less than the preset difference, the second training parameter is used to perform a second iterative update operation on the parameter matrix to be learned after the first iterative update to obtain a trained dequantized convolutional network.
[0099] Among them, the first training parameters include: the number of initial training times, the initial learning rate in the initial stage and the number of initial training iterations; the second training parameters include: the number of fine-tuning training times, the initial learning rate in the fine-tuning stage and the number of fine-tuning training iterations.
[0100] Specifically, before training the original dequantized convolutional network, similarly, the linear frequency modulation signal is first orthogonally demodulated to obtain an analog signal, and the analog signal is single-bit sampled using a preset linear frequency modulation signal sampling rate to obtain a single-bit complex signal of the first dimension.
[0101] After the single-bit complex signal of the first dimension has passed through the pulse repetition period M, a multi-bit complex echo data matrix S of the second dimension of N×M is obtained. dc . And the multi-bit complex echo data matrix S dc After quantization, the highest bit of each element is taken to obtain the single-bit complex echo data matrix S of the second dimension: o .
[0102] Then, the single-bit complex echo data matrix S o Fourier transform processing is performed in the column direction and the row direction respectively to obtain the range-Doppler spectrum (first range-Doppler spectrum) Z of the single-bit quantized echo data o . And multi-bit complex echo data matrix S dc Fourier transform processing is performed in the column direction and the row direction respectively to obtain the range-Doppler spectrum of multi-bit quantized echo data in = represents the operation of Fourier transform on the matrix in the column direction and the row direction. Finally, the first range Doppler spectrum and the third range Doppler spectrum are used to construct the training set image pair {(Z dc ) i ,(Z o ) i}∈I, where I represents the total number of datasets and i is the number of image pairs in the dataset.
[0103] Next, the second step is to train the original dequantized convolutional network as follows: The learning process of the weight matrix Θ is as follows:
[0104] First, the first training parameters in the initial training phase and the second training parameters in the fine-tuning phase are set respectively, that is, the initial training times t=1, the initial learning rate β in the initial phase t =0.001, the initial training iteration number is set to epoch1=100. And the fine-tuning training number T=1, the initial learning rate γ0 of the fine-tuning stage is set to 0.0001, and the fine-tuning training iteration number is set to epoch2=50.
[0105] Then, in the initial training phase, the first training parameters are used to perform a first iterative update operation on the parameter matrix to be learned.
[0106] Then, the pseudocode for the specific operations during the initial training is as follows:
[0107] Initialize the parameter matrix Θ0 to be learned as a random matrix, and initialize t = 1;
[0108] Execute the loop when t < epoch1;
[0109] Update the parameter matrix:
[0110] Update the loop count: t = t + 1;
[0111] Calculate the error between the two training loss functions: When the error is small enough (usually there is a small empirical value), reduce the learning rate β t to 0.0001;
[0112] End the loop and save the trained network parameter matrix (dequantized convolutional network) Θ t .
[0113] Among them, the loss function in the initial training stage is expressed as: Among them, represents the loss function the calculation of the derivative of Θ; I represents the total number of data sets.
[0114] Initial training stage (when the number of training times t is less than the initial training times epoch1) and network parameter fine-tuning stage (when the number of training times T is less than the fine-tuning times epoch2). Θ t is the network weight after t iterations, is the result of the derivative of the loss function of the initial training with respect to the network parameters, and after updating the network parameters Θ t in the first step, the number of iterations is incremented by 1 (t = t + 1).
[0115] Calculate the error value of the loss function between adjacent ones. When the error value is less than the preset difference, that is, calculate the errors of the two loss functions and When the error is small enough (usually there is a small empirical value), then reduce the learning rate β t to 0.0001, repeat this operation until the iteration reaches the number of training times, and save the trained network parameter Θ t , and enter the network parameter fine-tuning stage: use the second training parameter to perform a second iteration update operation on the parameter matrix to be learned after the first iteration update, and obtain the trained dequantized convolutional network.
[0116] Among them, the pseudocode for the specific operations during the network parameter fine-tuning training:
[0117] Utilize the network parameter matrix (dequantized convolutional network) Θ saved during initial training t Initialize the parameter matrix Θ to be learned T = Θ t , and initialize T = 1;
[0118] Execute the loop when T < epoch2;
[0119] Update the parameter matrix:
[0120] Update the loop count: T = T + 1;
[0121] Calculate the loss function: When the loss function stops decreasing, end the loop and save the trained network parameter matrix (dequantized convolutional network) Θ T .
[0122] Among them, the loss function in the fine-tuning stage is expressed as: Among them, represents the loss function The calculation of the derivative of Θ.
[0123] Then, after obtaining the trained dequantized convolutional network, input the first range-Doppler spectrogram into the weight matrix Θ of the trained dequantized convolutional network T to calculate the reconstructed two-dimensional spectrum (second range-Doppler spectrogram) by solving the following minimization problem
[0124] Among them, f resnet (·) represents the general features extracted by the residual network (ResNet-50); Z dc represents the range-Doppler spectrogram after two-dimensional Fourier transform of the multi-bit complex echo data matrix (range-Doppler spectrogram of multi-bit quantized echo data); Z o represents the range-Doppler spectrogram after two-dimensional Fourier transform of the single-bit complex echo data matrix (range-Doppler spectrogram of single-bit quantized echo data); represents the convolutional neural network obtained by training and learning, where Θ is the parameter matrix to be learned, and the input of the convolutional neural network is Z o ; represents the two-norm operation; α represents the balance coefficient.
[0125] Furthermore, please continue to refer to Figure 1 , S400. Perform coordinate transformation on the second range-Doppler spectrogram to obtain the imaging map of the instantaneous synthetic aperture radar.
[0126] Specifically, after obtaining the second range Doppler spectrum, the second range Doppler spectrum is subjected to coordinate transformation using the platform speed to obtain an instantaneous synthetic aperture radar imaging image. The coordinate transformation may be from a range Doppler coordinate system to a range azimuth coordinate system, and the transformation method may be through interpolation or other methods to obtain a SAR (instantaneous synthetic aperture radar) image corresponding to the data.
[0127] Among them, the beneficial effects of the present invention are: by transforming the representation domain of the SAR echo signal, the influence of the loss of the absolute amplitude information of the signal under single-bit sampling on SAR imaging is compensated. At the Nyquist sampling rate, the harmonic characteristics can be adaptively obtained according to the single-bit quantized de-skewed signal, and the instantaneous range Doppler can be restored by training the deep weight matrix, and then the single-bit SAR imaging task can be realized by using velocity interpolation.
[0128] The beneficial effects of the present invention can be further illustrated by the following simulation experiment, which is implemented based on the TensorFlow 2.x deep learning platform of Python3.
[0129] The parameters of the simulation data are as follows:
[0130] The radar is set to work in the positive side-view mode, with a beam angle θ of 120°, a radar carrier frequency of 77 GHz, a range sampling rate of 28 MHz, a pulse repetition frequency of 5000 Hz, and the number of echoes converted by FFT (FFT: Fast Fourier transform, Chinese name: Fast Fourier Transform) is 480.
[0131] Simulation results:
[0132] Figure 8 It is a two-dimensional instantaneous range-Doppler map of single-bit sampling (range-Doppler spectrum of single-bit quantized echo data), and the simulation target is the characteristics of multiple point targets in the scene to be tested. It can be seen that the background noise interference of the two-dimensional RD image is too large, and the signal-to-noise ratio of the scene information has changed, affecting the subsequent imaging tasks.
[0133] Fig. 9 This is the two-dimensional instantaneous range Doppler image (second range Doppler spectrum) restored after using the dequantization convolutional network. It can be seen that after the deep dequantization network module, the background noise is suppressed, the background is cleaner, the simulated target is clearer, most of the noise in the scene to be tested is eliminated, and the signal-to-noise ratio is restored to a certain extent.
[0134] For further information, see Fig.10The present invention provides a single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, and the single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network comprises:
[0135] The demodulation sampling module 1 is used to obtain a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal;
[0136] A first range Doppler spectrum acquisition module 2 is used to acquire a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, and to perform Fourier transform processing on the single-bit complex echo data matrix to obtain a first range Doppler spectrum;
[0137] A second range Doppler spectrum acquisition module 3, used for inputting the first range Doppler spectrum into the trained dequantized convolutional network to obtain a second range Doppler spectrum;
[0138] The coordinate conversion module 4 is used to perform coordinate conversion on the second range Doppler spectrum to obtain an instantaneous synthetic aperture radar imaging image.
[0139] Specifically, since the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network is applied to the single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, then, similarly, the single-bit imaging system based on two-dimensional spectrum reconstruction of a dequantized convolutional network is implemented as follows:
[0140] First, a preset linear frequency modulation signal is sent by a transmitting radar, and the preset linear frequency modulation signal is demodulated to obtain the analog signal, and then the analog signal is sampled to obtain the single-bit complex signal.
[0141] Secondly, after the single-bit complex signal is subjected to a preset M pulse repetition period, the single-bit complex echo data matrix S is obtained. o Furthermore, for the single-bit complex echo data matrix S o Perform Fourier transform processing to obtain the first range Doppler spectrum (range-Doppler spectrum of single-bit quantized echo data) Z o .
[0142] Furthermore, the first range Doppler spectrum is input into the trained dequantized convolutional network for prediction, and the second range Doppler spectrum is calculated by solving the following minimization problem, that is, the reconstructed two-dimensional spectrum (range-Doppler spectrum) is calculated. Finally, the second range Doppler spectrum is transformed into coordinates using the platform speed to obtain an instantaneous synthetic aperture radar imaging image.
[0143] Furthermore, if Fig.11 As shown, based on the above-mentioned single-bit imaging method based on two-dimensional spectrum reconstruction of dequantized convolutional network, the present invention also provides a controller accordingly, and the controller includes a processor 10, a memory 20 and a display 30. Fig.11 Only some of the components of the controller are shown, but it should be understood that implementation of all of the components shown is not required, and more or fewer components may be implemented instead.
[0144] In some embodiments, the memory 20 may be an internal storage unit of the controller, such as a hard disk or memory of the controller. In other embodiments, the memory 20 may also be an external storage device of the controller, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the controller. Further, the memory 20 may also include both an internal storage unit of the controller and an external storage device. The memory 20 is used to store application software and various types of data installed in the controller, such as the program code of the installed controller. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a single-bit imaging program 40 based on two-dimensional spectrum reconstruction of a dequantized convolutional network is stored on the memory 20, and the single-bit imaging program 40 based on two-dimensional spectrum reconstruction of a dequantized convolutional network can be executed by the processor 10, thereby realizing the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network in the present application.
[0145] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network.
[0146] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the controller and to display a visual user interface. The components 10-30 of the controller communicate with each other via a system bus.
[0147] In one embodiment, the steps implemented when the processor 10 executes the single-bit imaging program 40 based on dequantization convolutional network two-dimensional spectrum reconstruction in the memory 20 can refer to the corresponding description in the above-mentioned single-bit imaging method based on dequantization convolutional network two-dimensional spectrum reconstruction, and will not be repeated here.
[0148] Furthermore, the present invention provides a computer-readable storage medium, which stores a single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction, and when the single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction is executed by a processor, the steps of the single-bit imaging method based on dequantization convolutional network two-dimensional spectrum reconstruction as described above are implemented; since the steps of the single-bit imaging method based on dequantization convolutional network two-dimensional spectrum reconstruction have been described in detail above, they will not be repeated here.
[0149] In summary, the present invention provides a single-bit imaging method and system based on two-dimensional spectrum reconstruction of a dequantized convolutional network, the method comprising: obtaining a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, sampling the analog signal to obtain a single-bit complex signal; obtaining a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, Fourier transforming the single-bit complex echo data matrix to obtain a first range Doppler spectrum; inputting the first range Doppler spectrum into a trained dequantized convolutional network to obtain a second range Doppler spectrum; performing coordinate transformation on the second range Doppler spectrum to obtain an imaging image of an instantaneous synthetic aperture radar. After demodulating, sampling and Fourier transforming the linear frequency modulation signal to obtain the first range Doppler spectrum, the first range Doppler spectrum is output to the trained dequantized convolutional network for prediction to obtain a second range Doppler spectrum, thereby achieving high-complexity calculation at a high sampling rate.
[0150] It is understandable that those skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention, and all these changes or substitutions should fall within the protection scope of the claims attached to the present invention.
Claims
1. A single-bit imaging method based on two-dimensional spectrum reconstruction using a dequantized convolutional network. It is characterized in that The single-bit imaging method based on dequantized convolutional network two-dimensional spectrum reconstruction includes: Acquire a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal; According to the single-bit complex signal and a preset pulse repetition period, a single-bit complex echo data matrix is acquired, and the single-bit complex echo data matrix is subjected to Fourier transform processing to obtain a first range Doppler spectrum; Inputting the first range Doppler spectrum into a trained dequantized convolutional network to obtain a second range Doppler spectrum; The second range Doppler spectrum is subjected to coordinate transformation to obtain an instantaneous synthetic aperture radar imaging image.
2. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network according to claim 1, It is characterized in that The step of acquiring a linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, and sampling the analog signal to obtain a single-bit complex signal also includes: Preset the parameters of the transmitting radar according to the size of the scene to be tested; wherein the parameters include radar bandwidth, frequency modulation, time width and repetition rate; The transmitting radar transmits a linear frequency modulation signal according to the parameters.
3. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network according to claim 1, It is characterized in that The acquiring of the linear frequency modulation signal, demodulating the linear frequency modulation signal to obtain an analog signal, and sampling the analog signal to obtain a single-bit complex signal specifically includes: Performing orthogonal demodulation processing on the linear frequency modulation signal to obtain the analog signal; The analog signal is single-bit sampled according to a preset linear frequency modulation signal sampling rate to obtain a single-bit complex signal of a first dimension; wherein the first dimension is N×1.
4. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network according to claim 1, It is characterized in that The step of acquiring a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, and performing Fourier transform processing on the single-bit complex echo data matrix to obtain a first range Doppler spectrum specifically includes: When the single-bit complex signal is sampled and processed at the pulse repetition period, a single-bit complex echo data matrix of a second dimension is obtained; wherein the second dimension is N×M; The single-bit complex echo data matrix is subjected to Fourier transform processing in the column direction and the row direction respectively to obtain the first range-Doppler spectrum; wherein the first range-Doppler spectrum is a range-Doppler spectrum of single-bit quantized echo data.
5. The single-bit imaging method based on two-dimensional spectrum reconstruction of dequantized convolutional network according to claim 1, It is characterized in that The step of inputting the first range Doppler spectrum into a trained dequantized convolutional network to obtain a second range Doppler spectrum further includes: Presetting the structural parameters of the convolution layer, and constructing the convolution layer based on the structural parameters; wherein the structural parameters include: the size of the convolution kernel, the stride size, and the boundary expansion method; Splicing the cascade layer into the convolution layer to construct a harmonic feature estimation module; According to the point-by-point multiplication layer and the point-by-point addition layer, a reconstructed spectrogram generation module is constructed; Constructing and generating the dequantized convolutional network according to the harmonic feature estimation module and the reconstructed spectrum generation module; The dequantized convolutional network is trained and learned according to preset training parameters to obtain a trained dequantized convolutional network.
6. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network according to claim 5, It is characterized in that The step of training the dequantized convolutional network according to preset training parameters to obtain a trained dequantized convolutional network specifically includes: respectively setting a first training parameter in an initial training phase and a second training parameter in a fine tuning phase; In the initial training phase, performing a first iterative update operation on the parameter matrix to be learned using the first training parameters; When the error value of the adjacent loss functions is less than the preset difference, the second training parameter is used to perform a second iterative update operation on the parameter matrix to be learned after the first iterative update to obtain a trained dequantized convolutional network.
7. The single-bit imaging method based on two-dimensional spectrum reconstruction of a dequantized convolutional network according to claim 6, It is characterized in that The first training parameters include: the number of initial training times, the initial learning rate in the initial stage, and the number of initial training iterations; the second training parameters include: the number of fine-tuning training times, the initial learning rate in the fine-tuning stage, and the number of fine-tuning training iterations.
8. A single-bit imaging system based on two-dimensional spectrum reconstruction using a dequantized convolutional network. It is characterized in that The single-bit imaging system based on dequantized convolutional network two-dimensional spectrum reconstruction includes: A demodulation sampling module is used to obtain a linear frequency modulation signal, demodulate the linear frequency modulation signal to obtain an analog signal, and sample the analog signal to obtain a single-bit complex signal; A first range Doppler spectrum acquisition module, used to acquire a single-bit complex echo data matrix according to the single-bit complex signal and a preset pulse repetition period, and perform Fourier transform processing on the single-bit complex echo data matrix to obtain a first range Doppler spectrum; A second range Doppler spectrum acquisition module, used for inputting the first range Doppler spectrum into the trained dequantized convolutional network to obtain a second range Doppler spectrum; The coordinate conversion module is used to perform coordinate conversion on the second range Doppler spectrum to obtain an instantaneous synthetic aperture radar imaging image.
9. A controller, It is characterized in that The controller includes: a memory, a processor, a display, and a single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction stored in the memory and executable on the processor. When the single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction is executed by the processor, the steps of the single-bit imaging method based on dequantization convolutional network two-dimensional spectrum reconstruction as described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction. When the single-bit imaging program based on dequantization convolutional network two-dimensional spectrum reconstruction is executed by a processor, the steps of the single-bit imaging method based on dequantization convolutional network two-dimensional spectrum reconstruction as described in any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Single-bit band-pass sampling synthetic aperture radar imaging method
CN112965066A
Deception jamming suppression method and system based on single-channel SAR single imaging
CN113311431A
System and method for processing data in weather radar
US20080001808A1
Acoustinc sensing nodes and related systems and methods
US20210099196A1
