Radar range profile super-resolution method and system based on deep learning, terminal and storage medium

Abstract

The invention discloses a radar range profile super-resolution method and system based on deep learning, a terminal and a storage medium. The method comprises the following steps: acquiring simulated echo information reflected by a target; processing the simulated echo information to obtain baseband information; obtaining a frequency domain spectrum according to the baseband information; and constructing a deep learning neural network model, inputting the frequency domain spectrum into the deep learning neural network model, and outputting a super-resolution radar range profile. The method comprises the following steps: carrying out frequency mixing and baseband processing on an analog echo signal, extracting a baseband signal containing target distance information, carrying out fast Fourier transform on the baseband signal to obtain a low-resolution range profile frequency spectrum, inputting the range profile frequency spectrum into a physical-data dual-drive deep neural network, enhancing the feature extraction capability through a residual network module of a model, and obtaining a high-resolution range profile frequency spectrum. The resolution and noise immunity of the multi-target range profile under the condition of low signal-to-noise ratio are improved, and the dense target omission ratio is reduced.

Description

Deep learning-based radar range image super-resolution method, system, terminal, and storage medium Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a radar range image super-resolution method, system, terminal, and computer-readable storage medium based on deep learning. Background Technology

[0002] Currently, super-resolution methods are mainly represented by Capon, MUSIC, and ESPRIT. Although these methods have good super-resolution performance under medium to high signal-to-noise ratio (SNR) conditions, their performance deteriorates sharply as the SNR decreases. Furthermore, the MUSIC and ESPRIT methods require a known number of targets for decomposition of the signal and noise subspaces. Classical periodograms, as a maximum likelihood parameter estimation method, often outperform many complex parameter estimation methods at low SNRs. However, due to the mutual influence between the sidelobes of multiple components, the spectrum deviates from the true component positions, and the resolution cannot break through the classical theoretical resolution limit.

[0003] In related technologies, modern spectral estimation methods suffer from low resolution and poor noise resistance due to spectral heavy tailing; classical periodograms have wide main lobes and high side lobes, and the side lobes interfere with each other, resulting in poor resolution of dense targets. Therefore, the super-resolution methods used in related technologies have low resolution at low signal-to-noise ratios.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a radar range image super-resolution method, system, terminal, and computer-readable storage medium based on deep learning, aiming to solve the problem of low resolution of existing super-resolution methods at low signal-to-noise ratios.

[0006] To achieve the above objectives, the present invention provides a super-resolution radar range image method based on deep learning. The super-resolution radar range image method based on deep learning includes the following steps: acquiring simulated echo information reflected by the target; processing the simulated echo information to obtain baseband information; obtaining the frequency domain spectrum based on the baseband information; constructing a deep learning neural network model; inputting the frequency domain spectrum into the deep learning neural network model; and outputting a super-resolution radar range image.

[0007] Optionally, in the deep learning-based radar range image super-resolution method, the simulated echo information includes target echo signals and noise clutter signals; acquiring the simulated echo information reflected by the target specifically includes: after the radar transmitter generates a target electrical signal, converting the target electrical signal into electromagnetic waves radiated in space through a transmitting antenna; receiving a portion of the electromagnetic waves reflected from the target object through a receiving antenna, and converting the portion of the electromagnetic waves into simulated electrical signals; and performing front-end processing on the simulated electrical signals to obtain the target echo signal and noise clutter signals.

[0008] Optionally, in the deep learning-based radar range image super-resolution method, the baseband information is a baseband signal; the processing of the analog echo information to obtain the baseband information specifically includes: converting the analog echo information into a digital echo signal; obtaining an intermediate frequency signal based on the digital echo signal; and performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal.

[0009] Optionally, in the deep learning-based radar range image super-resolution method, the step of converting the analog echo information into a digital echo signal specifically involves converting the analog echo information into a digital echo signal using an analog-to-digital converter, wherein the analog echo information is a high-frequency analog signal and the digital echo signal is a high-frequency digital signal.

[0010] Optionally, the deep learning-based radar range image super-resolution method, wherein obtaining the intermediate frequency signal from the digital echo signal specifically includes: acquiring the radar's transmitted signal and using the transmitted signal as a reference signal; mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal.

[0011] Optionally, in the deep learning-based radar range image super-resolution method, the intermediate frequency signal includes a sum frequency component and a difference frequency component; the step of mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal specifically involves multiplying the digital echo signal with the reference signal to obtain the sum frequency component and the difference frequency component of the digital echo signal and the reference signal.

[0012] Optionally, in the deep learning-based radar range image super-resolution method, the baseband signal is a difference frequency component; the step of performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal specifically involves filtering the intermediate frequency signal through a low-frequency filter to obtain a low-frequency difference frequency component.

[0013] Optionally, the deep learning-based radar range image super-resolution method, wherein obtaining the frequency domain spectrum based on the baseband information specifically includes: windowing the baseband signal to obtain a time domain signal data block; and performing a fast Fourier transform on the time domain signal data block to obtain the frequency domain spectrum.

[0014] Optionally, in the deep learning-based radar range image super-resolution method, the frequency domain spectrum is a spectrogram; the step of performing a fast Fourier transform on the time domain signal data block to obtain the frequency domain spectrum specifically includes: performing a fast Fourier transform on the time domain signal data block to calculate a complex array; extracting the corresponding frequency and amplitude from the complex array; and obtaining the spectrogram corresponding to the baseband signal based on the frequency and amplitude.

[0015] Optionally, the deep learning-based radar range image super-resolution method includes a deep learning neural network model comprising an input layer, a feature extraction module, a residual learning module, and a reconstruction layer.

[0016] Optionally, the deep learning-based radar range image super-resolution method, wherein constructing the deep learning neural network model specifically includes: obtaining the training spectrum and the corresponding ground truth labels, constructing a network model to be trained; training the network model to be trained according to the training spectrum and the ground truth labels to obtain the deep learning neural network model.

[0017] Optionally, the deep learning-based radar range image super-resolution method, wherein training the network model to be trained based on the training spectrum and the ground truth label to obtain a deep learning neural network model specifically includes: inputting the training spectrum into the network model to be trained to obtain a predicted spectrum; defining a loss function and calculating the error between the predicted spectrum and the ground truth label based on the loss function; and updating the parameters of the network model to be trained based on the error to obtain a deep learning neural network model.

[0018] Optionally, the super-resolution radar range image method based on deep learning, wherein the step of inputting the frequency domain spectrum into the deep learning neural network model and outputting the super-resolution radar range image specifically includes: the input layer inputting the frequency domain spectrum into the feature extraction module; the feature extraction module extracting feature data from the frequency domain spectrum; the residual learning module correcting the frequency domain spectrum to obtain corrected data; and the reconstruction layer obtaining the super-resolution radar range image based on the feature data and the corrected data.

[0019] Furthermore, to achieve the above objectives, the present invention also provides a radar range image super-resolution system based on deep learning, wherein the radar range image super-resolution system based on deep learning includes: a signal receiving module for acquiring simulated echo information reflected by a target; a baseband processing module for processing the simulated echo information to obtain baseband information; a spectrum estimation module for obtaining a frequency domain spectrum based on the baseband information; and a neural network model construction and spectrum output module for constructing a deep learning neural network model, inputting the frequency domain spectrum into the deep learning neural network model, and outputting a super-resolution radar range image.

[0020] Optionally, in the deep learning-based radar range image super-resolution system, the signal receiving module includes: an electromagnetic wave transmitting unit, used to convert the target electrical signal into electromagnetic waves radiated in space via a transmitting antenna after the radar transmitter generates the target electrical signal; a reflected echo receiving unit, used to receive a portion of the electromagnetic waves reflected from the target object via a receiving antenna, and convert the portion of the electromagnetic waves into analog electrical signals; and an analog signal front-end processing unit, used to perform front-end processing on the analog electrical signals to obtain target echo signals and noise clutter signals.

[0021] Optionally, in the deep learning-based radar range image super-resolution system, the baseband processing module includes: an analog-to-digital conversion unit for converting the analog echo information into a digital echo signal; a mixing unit for obtaining an intermediate frequency signal based on the digital echo signal; and a low-pass filtering unit for performing low-pass filtering on the intermediate frequency signal to obtain a baseband signal.

[0022] Optionally, in the deep learning-based radar range image super-resolution system, the spectrum estimation module includes: a data windowing unit for windowing the baseband signal to obtain a time-domain signal data block; and a fast Fourier transform processing unit for performing a fast Fourier transform on the time-domain signal data block to obtain a frequency-domain spectrum.

[0023] Optionally, the deep learning-based radar range image super-resolution system includes a neural network model construction and spectrum output module comprising: a network model construction unit for acquiring training spectra and corresponding ground truth labels to construct a network model to be trained; a network model training unit for training the network model to be trained based on the training spectra and the ground truth labels to obtain a deep learning neural network model; a spectrum input unit for the input layer to input the frequency domain spectrum to the feature extraction module; a feature extraction unit for the feature extraction module to extract feature data from the frequency domain spectrum; a spectrum correction unit for the residual learning module to correct the frequency domain spectrum to obtain corrected data; and a feature reconstruction unit for the reconstruction layer to obtain a super-resolution radar range image based on the feature data and the corrected data.

[0024] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a deep learning-based radar range super-resolution program stored in the memory and executable on the processor, wherein when the deep learning-based radar range super-resolution program is executed by the processor, it implements the steps of the deep learning-based radar range super-resolution method as described above.

[0025] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a deep learning-based radar range image super-resolution program, which, when executed by a processor, implements the steps of the deep learning-based radar range image super-resolution method as described above.

[0026] In this invention, simulated echo information reflected from a target is acquired; the simulated echo information is processed to obtain baseband information; based on the baseband information, a frequency domain spectrum is obtained; a deep learning neural network model is constructed, and the frequency domain spectrum is input into the deep learning neural network model to output a super-resolution radar range image. This invention extracts a baseband signal containing target range information by mixing and processing the simulated echo signal, performs a Fast Fourier Transform on the baseband signal to obtain a low-resolution range image spectrum, and then inputs the low-resolution range image spectrum into a physical-data dual-driven deep neural network (the core module includes a residual network). The network integrates the radar signal physical model (such as the echo generation mechanism and noise characteristics) with data-driven feature learning, and learns the mapping from the low-resolution spectrum to the high-resolution range image through training, outputting a super-resolution radar range image. This application enhances feature extraction capabilities through the residual network module, improves the resolution and noise resistance of multi-target range images under low signal-to-noise ratio conditions, and reduces the false negative rate of dense targets. Attached Figure Description

[0027] Figure 1 is a flowchart of a preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 2 is a flowchart of the specific implementation process of step S10 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 3 is a flowchart of the specific implementation process of step S20 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 4 is a flowchart of the specific implementation process of step S22 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 5 is a flowchart of the specific implementation process of step S30 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 6 is a flowchart of the specific implementation process of step S32 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 7 is a flowchart of the specific implementation process of step S40 in the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 8 is a flowchart of the preferred embodiment of the radar range image super-resolution method based on deep learning of the present invention; Figure 9 is a block diagram of the deep learning neural network in a preferred embodiment of the radar range image super-resolution method based on deep learning; Figure 10 is a flowchart of the specific implementation process of step S42 in a preferred embodiment of the radar range image super-resolution method based on deep learning; Figure 11 is a schematic diagram of the comparison results of multiple frequency component false negative rate curves in a preferred embodiment of the radar range image super-resolution method based on deep learning; Figure 12 is a schematic diagram of the radar range image super-resolution comparison results in a preferred embodiment of the radar range image super-resolution method based on deep learning; Figure 13 is a schematic diagram of one structure of the radar range image super-resolution system based on deep learning; Figure 14 is a schematic diagram of another structure of the radar range image super-resolution system based on deep learning; Figure 15 is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and effects of this application clearer and more explicit, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only possible technical implementations of this application and not all possible implementations. Based on the embodiments in this application, those skilled in the art can obtain other embodiments without creative effort, and these embodiments are also within the protection scope of this application.

[0029] First, let's introduce the terms used in the embodiments of this application: Capon, the Capon method, is a modern spectral estimation method, also known as the Minimum Variance Distortionless Response (MVDR) method.

[0030] MUSIC, or Multiple Signal Classification, is a modern spectral estimation method based on subspace decomposition. It achieves high-resolution direction finding or spectral estimation by separating the signal subspace and the noise subspace.

[0031] ESPRIT, Estimation of Signal Parameters via Rotational Invariance Techniques, is a modern spectral estimation method that utilizes the rotational invariance of the signal subspace to estimate signal parameters.

[0032] An ADC, or Analog-to-Digital Converter, is a device or circuit that converts analog signals into digital signals and is a key component of digital signal processing.

[0033] FFT, or Fast Fourier Transform, is an efficient algorithm for the Discrete Fourier Transform (DFT) used to convert time-domain signals into frequency-domain signals. It is a core step in radar signal processing for obtaining range profiles.

[0034] dB, decibel, is a logarithmic unit representing the signal strength or power ratio, and is often used to describe the signal-to-noise ratio.

[0035] Among related technologies, modern spectral estimation methods, represented by Capon, MUSIC (Multi-Signal Classification), and ESPRIT, rely on signal-noise subspace decomposition. They can achieve super-resolution by breaking the Rayleigh limit at medium to high signal-to-noise ratios (SNR), but require prior knowledge of the number of targets (e.g., MUSIC and ESPRIT). Furthermore, at low SNRs, subspace decomposition errors lead to a sharp deterioration in performance. Modern spectral estimation suffers from reduced resolution due to the spectral "tailing" effect (high sidelobe energy and slow decay); it relies on accurate prior knowledge of the number of targets, which is difficult to satisfy in real-world scenarios; and at low SNRs, subspace decomposition fails, resulting in a sharp performance decline.

[0036] The classical periodogram method, as a maximum likelihood parameter estimation method, outperforms complex methods in noise resistance at low signal-to-noise ratios. However, it is limited by the main lobe width and high sidelobe characteristics. In dense multi-target scenes, the sidelobes interfere with each other, causing the spectrum to deviate from the true position, and it cannot break through the theoretical resolution limit (such as the Rayleigh limit). The classical periodogram method has a wide main lobe (the resolution is limited by the time-bandwidth product of the Fourier transform); the sidelobes are high and the sidelobes of multiple targets overlap, resulting in a shift in the spectral peak, making it easy to miss or falsely detect dense targets.

[0037] This application can solve the problems of low resolution and poor noise resistance in the prior art, so as to improve the resolution of multi-target range images.

[0038] This application presents a deep learning-based super-resolution method for radar range images. By constructing a physical-data dual-driven deep neural network and combining residual network modules with domain knowledge, a deep learning network containing dense residual modules is used to achieve time-frequency domain transformation of the signal, thereby improving the radar range image resolution under low signal-to-noise ratio conditions and solving the problem of low range image resolution under low signal-to-noise ratio conditions. The deep learning network architecture designed in this application reduces the target false detection rate by more than 10% compared with traditional methods in dense target scenarios.

[0039] In this application, the radar acquires analog echo signals via a receiving antenna, which are first converted into digital echo signals by an ADC (analog-to-digital converter). Using the radar's known transmitted signal as a reference signal, the digital echo signal is mixed (i.e., multiplied) with this reference signal to obtain an intermediate frequency (IF) signal. The IF signal is then low-pass filtered to obtain the baseband signal. A Fast Fourier Transform (FFT) is then performed on the baseband signal to obtain its spectrum. This spectrum is then used as input to a neural network, which receives the super-resolution spectrum, i.e., the super-resolution radar range profile.

[0040] The radar range image super-resolution method based on deep learning according to a preferred embodiment of the present invention is shown in Figure 1. The radar range image super-resolution method based on deep learning includes the following steps: Step S10, acquiring simulated echo information reflected by the target.

[0041] Specifically, the simulated echo information includes the target echo signal and noise clutter signal.

[0042] It should be noted that simulated echo information includes target echo signals, noise, clutter, and interference. The target echo signal is the most desirable part, carrying crucial target information. Delay is the time difference between signal transmission and reception, directly corresponding to the target's distance. Doppler shift occurs due to the target's motion relative to the radar; the frequency of the echo signal changes slightly, corresponding to the target's speed. Amplitude is the intensity of the echo signal, roughly reflecting the target's size and reflection characteristics. Noise is mainly thermal noise generated by the thermal motion of electronic components and noise introduced by the receiver itself; it is random and ubiquitous, a major factor limiting radar detection capabilities, especially under low signal-to-noise ratio conditions. Clutter refers to reflections from non-target objects, such as the ground, buildings, mountains, raindrops, and ocean waves. These reflected signals can be very strong and can mask the signal of the target of true interest. Interference comes from signals from other radars, communication equipment, or intentional interference sources, contaminating the echo signal. In this case, the simulated echo information is a high-frequency analog signal, with a frequency close to the radar's transmission frequency. Directly processing this high-frequency signal is technically very difficult and costly.

[0043] Radar detects target distance, speed, and angle by emitting electromagnetic waves and receiving the echo signals reflected from the target. The principle is broken down into three stages: signal generation, target reflection, and echo reception and processing. The radar workflow includes the following steps: (1) Signal generation: Radar usually emits linear frequency modulated (LFM) signals or pulse signals. When the signal type is LFM, the frequency changes linearly with time and has a large bandwidth, which can improve the range resolution. When the signal type is pulse signal: short-duration high-frequency pulse, the target distance is calculated by measuring the echo delay. Transmission method: the signal is radiated into space through the transmitting antenna to form electromagnetic wave propagation. (2) Target reflection: when the electromagnetic wave encounters the target, some of the energy is reflected, forming an echo signal. (3) Receiving antenna: captures the simulated echo signal (including noise) reflected by the target.

[0044] As shown in Figure 2, step S10 specifically includes: step S11, after the radar transmitter generates the target electrical signal, the target electrical signal is converted into electromagnetic waves radiated in space through the transmitting antenna.

[0045] Specifically, step S11 is implemented as follows: the transmitter of the radar system generates a specific electrical signal (such as a linear frequency modulated signal), which is converted into electromagnetic waves by the transmitting antenna and radiated into space.

[0046] Understandably, this is the starting point of the entire detection process. Without transmission, there is no subsequent reflection and reception. The characteristics of the transmitted signal (such as frequency, bandwidth, and waveform) directly determine the radar's final performance (such as resolution and detection range).

[0047] Step S12: Receive a portion of the electromagnetic waves reflected from the target object using a receiving antenna, and convert the portion of the electromagnetic waves into an analog electrical signal.

[0048] Specifically, step S12 is implemented as follows: (1) The emitted electromagnetic wave propagates in space. When it encounters a target object (such as a car), part of the energy is absorbed by the target, part is scattered, and the other part is reflected back to the direction of the radar system.

[0049] Understandably, this is a process of information modulation. The target's physical properties, such as distance, velocity, and shape, are "modulated" into the reflected echo, causing it to carry crucial information about the target. For example, the target's distance determines the echo delay time, and the target's velocity determines the Doppler shift of the echo.

[0050] (2) The receiving antenna of the radar system (which may be shared with the transmitting antenna, i.e., the transmitting and receiving are co-located) is specially designed to capture the weak electromagnetic waves reflected back from the target and convert them back into weak analog electrical signals.

[0051] Understandably, this is a core aspect of information acquisition. The directivity, gain, and noise characteristics of the receiving antenna determine its ability to capture weak echo signals, directly affecting the radar's detection range and sensitivity.

[0052] Step S13: Perform front-end processing on the analog electrical signal to obtain the target echo signal and noise clutter signal.

[0053] Specifically, step S13 is implemented as follows: The captured analog echo signal is usually very weak and mixed with noise, requiring immediate front-end processing. This includes: low-noise amplification: amplifying the weak echo signal to a sufficient amplitude while introducing very low noise to improve the signal-to-noise ratio. Filtering: filtering out out-of-band noise and interference, retaining the signal frequency band containing target information.

[0054] Understandably, maximizing signal quality before it enters the digital processing (ADC) stage lays the foundation for subsequent digitization and precise processing. The quality of this processing step directly determines the performance ceiling of the entire radar system.

[0055] From a physics and signal processing perspective, noise and clutter are both signals. Echo signal: A physical quantity of electromagnetic waves that changes over time, carrying information such as the target's distance and speed. It is a signal. Noise: Random and irregular changes in electromagnetic waves (or voltage). It is also a physical quantity that changes over time. Although it doesn't carry the information we want, it is itself a description of a physical phenomenon, therefore it is also a signal. Clutter: Reflections from non-target objects such as the ground, buildings, and clouds. It is also a physical quantity of electromagnetic waves that changes over time, but the information it carries is not the target information we are interested in. Therefore, it is also a signal, called interference signal or unwanted signal.

[0056] It's important to note that transmission is a prerequisite for reception. The raw signal captured by the receiving antenna is often of poor quality. Processing is necessary to purify and amplify this signal, making it more suitable for subsequent digital processing. The information flow relationships are as follows: Information embedding: The target's physical information (distance, speed, etc.) is encoded into the delay and frequency of the reflected echo during electromagnetic wave propagation and interaction with the target; Information transmission: The receiving antenna captures this encoded analog signal during the reception of the reflected echo, but it is a "raw" and "contaminated" information carrier; Information pre-extraction: The amplification and filtering in the analog signal front-end processing can be seen as the first purification of the information, removing a large amount of irrelevant noise and making the target information more prominent. The system cascading relationship is as follows: The entire process is a cascaded system. The output of the previous step is the input of the next step. The quality of the transmitted signal affects the signal-to-noise ratio of the reflected echo, and the performance of the receiving antenna determines the initial quality of the signal entering the front-end processing. The processing effect of the front-end processing directly determines the quality of the final output analog echo information, which will directly affect the final effect of subsequent ADC conversion, mixing, FFT, and even neural network processing. Therefore, step S10 is not a single action, but a complete physical process that includes transmission, propagation, reception, and preprocessing. Its core logic is: by transmitting a known signal, information is loaded by using the target's modulation effect on the signal, and then the information-carrying signal is captured and preliminarily optimized through reception and front-end processing, ultimately providing a high-quality analog signal source for subsequent digital signal processing and intelligent algorithms.

[0057] Step S20: Process the simulated echo information to obtain baseband information.

[0058] Specifically, the baseband information refers to the baseband signal. The main purpose of step S20 is not to remove noise and clutter, but to perform frequency down-conversion. It processes the entire analog echo information, which contains the desired target signal, but also inextricably includes noise, clutter, and interference.

[0059] As shown in Figure 3, step S20 specifically includes: step S21, converting the analog echo information into a digital echo signal.

[0060] Specifically, step S21 is as follows: Step S211: Convert the analog echo information into a digital echo signal using an analog-to-digital converter, wherein the analog echo information is a high-frequency analog signal and the digital echo signal is a high-frequency digital signal.

[0061] Specifically, step S211 involves converting the continuous, analog echo signal into a discrete, digital signal, which means transforming the continuous voltage waveform into a digital sequence. Since computers and digital signal processors (DSPs, GPUs) can only process digital signals, this is the necessary step from the analog world to the digital world. Thus, analog echo information (high-frequency analog signal) is input, and digital echo signal (high-frequency digital signal) is output.

[0062] Step S22: Obtain the intermediate frequency signal based on the digital echo signal.

[0063] Specifically, the digital echo signal is multiplied by a local oscillator (lo) signal. The frequency of this lo signal is very close to the frequency of the radar's transmitted signal.

[0064] As shown in Figure 4, step S22 specifically includes: step S221, acquiring the radar's transmission signal and using the transmission signal as a reference signal.

[0065] Step S222: Mix the digital echo signal with the reference signal to obtain an intermediate frequency signal.

[0066] Specifically, the intermediate frequency signal includes a sum frequency component and a difference frequency component.

[0067] Specifically, step S222 involves: step S2221, multiplying the digital echo signal with the reference signal to obtain the sum frequency component and difference frequency component of the digital echo signal and the reference signal.

[0068] Specifically, step S2221 is implemented by using the trigonometric function product-difference formula: cos(ω1t)*cos(ω2t)=1 / 2*[cos((ω1+ω2)t)+cos((ω1-ω2)t)]; ω1 is the angular frequency of the echo signal; ω2 is the angular frequency of the lo signal. After multiplication, two new frequency components are generated: the sum frequency (ω1+ω2) and the difference frequency (ω1-ω2). This is the key to the entire process. Frequency shift: Since the lo frequency (ω2) is extremely close to the echo frequency (ω1), their difference frequency (ω1-ω2) will be a very low frequency, close to 0 Hz. This low-frequency signal is the desired baseband signal. Reduced processing difficulty: Shifting signal processing from high frequency (GHz level) to low frequency (MHz or even kHz level) can greatly reduce the requirements of subsequent processing (such as filtering, FFT) on hardware sampling rate and computing power, thereby reducing costs and improving efficiency. It takes a digital echo signal and a local oscillator (lo) signal as input and outputs an intermediate frequency (IF) signal, which contains both high-frequency sum frequency components and low-frequency difference frequency components.

[0069] Step S23: Perform low-pass filtering on the intermediate frequency signal to obtain the baseband signal.

[0070] Specifically, the baseband signal is a difference frequency component.

[0071] It should be noted that: Target: The object being detected by radar (such as a car, building, etc.). Range: The straight-line distance from the target to the radar (determined by the round-trip time delay of electromagnetic waves). Velocity: The radial velocity of the target relative to the radar (caused by frequency shift due to the Doppler effect). The core task of radar is to deduce these two key parameters by analyzing the difference between the transmitted and echo signals.

[0072] Specifically, step S23 is as follows: Step S231, the intermediate frequency signal is filtered by a low-frequency filter to obtain the low-frequency difference frequency component.

[0073] It should be noted that when the radar's transmitted signal (such as a continuous wave) and the echo signal reflected from the target are multiplied (mixed) in the mixer, two frequency components are generated: a sum frequency component (high frequency): the sum of the transmitted signal frequency and the echo signal frequency; and a difference frequency component (low frequency): the difference between the transmitted signal frequency and the echo signal frequency. The frequency of the sum frequency component is much higher than that of the original signal (for example, if the transmitted signal is 24 GHz and the echo frequency is close to 24 GHz, the sum frequency may be close to 48 GHz). Since the cutoff frequency of the low-pass filter is designed to retain only the low-frequency difference frequency component (typically in the kHz to MHz range), the sum frequency component is filtered out and does not participate in subsequent processing.

[0074] Specifically, step S231 involves processing the intermediate frequency (IF) signal using a low-pass filter (LPF). After mixing, two frequency components (sum frequency and difference frequency) are generated. Only the low-frequency difference frequency component containing target information is relevant; the LPF filters this component, allowing only the low-frequency component to pass while completely filtering out the high-frequency sum frequency component. The input IF signal (containing both high and low frequency components) outputs a baseband signal (a clean low-frequency signal containing the target's distance and velocity information).

[0075] After mixing and low-pass filtering, the retained baseband signal is a low-frequency difference frequency signal. Its low-frequency characteristics (relative to the original RF signal) make subsequent digital processing (such as ADC sampling and FFT transformation) easier to implement, and it retains the target's range and velocity information. The difference frequency component is the carrier of information. High frequencies and frequencies that are too high are filtered out by the low-pass filter and do not carry useful information. The low-frequency difference frequency is determined by both the target's range (time delay) and velocity (Doppler shift), and is the core basis for radar to extract target parameters. By analyzing the spectrum of the baseband signal (difference frequency signal), the frequencies corresponding to the range and velocity can be separated, thereby achieving high-precision target detection.

[0076] Understandably, a high-frequency analog echo is like an article written in a very complex and obscure foreign language (high-frequency carrier wave), filled with a lot of useless information (noise and clutter), making it very difficult to understand directly. ADC conversion is like typing this foreign language article into an electronic version; although still in a foreign language, it's now digitally editable. Mixing is like hiring a top-notch translator (local oscillator, LO) to translate this complex foreign language article into a very simple Esperanto (baseband signal), but during the translation process, echoes of the original text (sum frequency components) may be generated simultaneously. Low-pass filtering is like a skilled editor who only retains the translated, simple Esperanto article, completely removing the useless "echoes." The resulting baseband signal is like a clean article written in a simple language, highlighting the core content. It retains all the key information about the target (distance, velocity) while reducing the signal frequency to a very manageable range. This clean baseband signal is the ideal input for subsequent FFT and deep learning super-resolution applications.

[0077] Step S30: Obtain the frequency domain spectrum based on the baseband information.

[0078] Understandably, a time-domain signal (baseband signal) describes how a signal changes over time; a baseband signal is a complex waveform that varies with time. The frequency domain spectrum tells us which frequencies make up this complex waveform and the proportion of each frequency. The Fourier Transform (FT) is the mathematical tool for analyzing these components, and the Fast Fourier Transform (FFT) is the most efficient algorithm for calculating the results of this tool.

[0079] Specifically, the frequency domain spectrum is a spectrogram. The frequency domain spectrum is a distribution map of the frequency components of a baseband signal (time domain). It tells us which frequencies constitute the signal primarily within the analyzed time segment, and their respective intensities. It represents a perspective transformation from the time domain to the frequency domain. The input spectrum is low-resolution and has a low signal-to-noise ratio, while the super-resolution spectrum (super-resolution radar range image) output by the neural network is high-resolution and has a high signal-to-noise ratio.

[0080] As shown in Figure 5, step S30 specifically includes: step S31, windowing the baseband signal to obtain a time-domain signal data block.

[0081] Specifically, step S31 is implemented as follows: This application does not perform FFT on the entire long baseband signal all at once. Because the signal changes over time (for example, the difference frequency of an FMCW radar changes during triangular wave modulation), this application needs to analyze a stable segment within a short period of time. This process is called windowing. Specifically, a small segment (e.g., 1024 or 2048 sampling points) is truncated from the continuous baseband signal. To reduce spectral leakage caused by truncation (i.e., energy from one frequency leaks to adjacent frequencies, leading to inaccurate measurements), this segment of the signal is multiplied by a window function (such as a Hanning window or a Hamming window). The window function smoothly decreases to zero at both ends and is 1 in the middle, as if adding a smooth frame to this segment of the signal. A windowed, finite-length time-domain signal data block is output.

[0082] Step S32: Perform a fast Fourier transform on the time-domain signal data block to obtain the frequency-domain spectrum.

[0083] As shown in Figure 6, step S32 specifically includes: step S321, performing a fast Fourier transform on the time-domain signal data block to calculate a complex array.

[0084] Specifically, step S321 is implemented as follows: inputting this data block into the FFT algorithm, inputting n complex or real sampling points, representing the change of the amplitude of the time-domain signal over time (in radar, the baseband signal is usually i / q two-way, so it is a complex signal); during the calculation process, FFT decomposes these n time-domain points into n frequency components through a series of efficient complex multiplications and additions; outputting n complex numbers, the output of FFT is not a single value, but a complex array.

[0085] Step S322: Extract the corresponding frequency and amplitude from the complex number array.

[0086] Specifically, step S322 involves extracting meaningful information from the n complex numbers output by the FFT, each corresponding to a specific frequency bin. For the k-th complex number x_k = a + bi: Frequency: The frequency f_k corresponding to this complex number is f_k = k * (sampling rate / number of FFT points). Amplitude a_k = sqrt(a^2 + b^2), where the amplitude value represents the amount or strength of the frequency component f_k contained in the original signal. A large amplitude value means that this frequency is significant in the original signal.

[0087] Step S323: Obtain the spectrum diagram corresponding to the baseband signal based on the frequency and the amplitude.

[0088] Specifically, step S323 is implemented as follows: Now we have the necessary elements for plotting the spectrum: x-axis (horizontal axis): frequency, with the frequency f_k corresponding to each complex number as the x-axis coordinate; y-axis (vertical axis): amplitude (or power, power = amplitude²), with the amplitude a_k calculated for each complex number as the y-axis coordinate. Connecting these points (f_k, a_k) yields the frequency domain spectrum, which is a curve with peaks appearing at the frequencies that dominate the original baseband signal.

[0089] As shown in Figures 7, 8 and 9, step S40 specifically includes: step S41, obtaining the training spectrum and the corresponding ground truth labels, and constructing the network model to be trained.

[0090] Specifically, step S41 is implemented as follows: (1) Generate input data (x-low resolution spectrum): Through actual experimental measurement or high-precision electromagnetic simulation software, simulate and generate a large number of radar echoes of different target scenarios (such as multiple point targets, different distances, and different intensities); artificially add Gaussian white noise to these echo signals to simulate low signal-to-noise ratio conditions; use the above-mentioned physical signal processing chain (adc->mixing->low-pass filtering->fft) to process these noisy signals and obtain the low-resolution range image spectrum, which is the input x of the network.

[0091] Generating ground truth labels (y-high resolution spectrum): This is crucial for data-driven learning. A perfect high-resolution distance image is needed as the target for the network's learning.

[0092] Method a (simulation-driven): In the simulation, a larger bandwidth or a longer coherent accumulation time is used to generate an ideal, noise-free, high-resolution range image spectrum. This high-resolution spectrum is the label y.

[0093] Method b (Measurement-Driven): In actual measurements, it is difficult to obtain a perfect true value. An alternative approach is to use measurement data with extremely high signal-to-noise ratio and extremely high bandwidth as an approximate true value, or to use traditional super-resolution algorithms (such as Capon and Music, but they fail at low signal-to-noise ratios) to process high signal-to-noise ratio data to generate labels.

[0094] Data pairing and augmentation: The generated data are matched one-to-one to form (x, y) data pairs. Data augmentation is then performed to improve the model's generalization ability. For example, adding noise of different intensities to the input x or making small random offsets to the target position. This step essentially creates a supervised learning task by intentionally generating a bad input (low signal-to-noise ratio, low resolution) while having a good answer (high resolution), and then letting the network learn the complex mapping relationship between the two.

[0095] (2) Define a deep neural network model (the network model to be trained), and design a function f(x;θ), where x is the input and θ is the network parameters (weights and biases) to be learned, so that its output f(x;θ) is as close as possible to the true value y. The input layer receives a low-resolution spectrum x; multiple convolutional layers and residual blocks constitute the core of the network, used to extract deep features from x and learn complex nonlinear transformations; the output layer outputs a high-resolution spectrum with the same dimension as the true value y. Incorporating domain knowledge (physics-driven): Architecture design: The design of the residual network itself draws on the residual learning idea in the physical inversion problem (learning a corrected quantity rather than a completely new quantity); Initial prior: The initial layers or weights of some layers of the network can be initialized by traditional matched filtering or deconvolution algorithms, allowing the network to learn from the shoulders of giants from the beginning, rather than randomly exploring from scratch.

[0096] Understandably, a network model is like a highly complex, adjustable filter. Its millions of parameters θ are like countless knobs that can be rotated, and the goal of training is to find a set of optimal knob angles so that the filter can turn bad inputs into good outputs.

[0097] Step S42: Train the network model to be trained according to the training spectrum and the ground truth label to obtain a deep learning neural network model.

[0098] Specifically, the construction of the deep learning neural network model includes an input layer, a feature extraction module, a residual learning module, and a reconstruction layer.

[0099] As shown in Figure 10, step S42 specifically includes: step S421, inputting the training spectrum into the network model to be trained to obtain the predicted spectrum.

[0100] Specifically, step S421 involves inputting a batch of training data (low-resolution spectrum) into the network and performing forward propagation to calculate the network output (predicted high-resolution spectrum). That is, inputting a low-resolution spectrum (i.e., the training spectrum) into the network will result in the network outputting a current super-resolution distance image (i.e., the predicted spectrum).

[0101] The process involves calculating the loss by comparing the network's output with the prepared "standard answer" (high-resolution label) and calculating the difference (loss). Common loss functions are mean squared error or mean absolute error. Backpropagation, based on the loss value, uses gradient descent to calculate the contribution (gradient) of each network parameter (weights and biases) to the total error from the output layer back to the input layer. Parameter updates involve using an optimizer (such as Adam) to fine-tune all parameters based on the gradients, ensuring that the network's output is closer to the label and the loss is smaller in the next forward propagation. Iteration involves repeating this process thousands of times until the network's loss no longer decreases significantly across the entire training set. At this point, the network has learned how to reconstruct a high-resolution range image from a low-resolution spectrum.

[0102] Input x: Obtained by performing an FFT on a noisy, low-bandwidth simulated / measured signal, referred to as a low-resolution range profile (spectral form). Ground truth label y: Obtained by performing an FFT on a noiseless, high-bandwidth simulated signal (or a high-quality range profile obtained through other methods), referred to as a high-resolution range profile (spectral form). Network prediction f(x;θ): The network attempts to reconstruct y from x. Loss function: Calculates the difference between f(x;θ) and y (e.g., MSE), comparing two spectral vectors, i.e., two range profile vectors.

[0103] Step S422: Define a loss function and calculate the error between the predicted spectrum and the true label based on the loss function.

[0104] Step S422 specifically involves quantifying the error between the network's current prediction f(x;θ) and the true value y. Specifically, the base loss typically uses mean squared error (mse) or mean absolute error (mae). Physical constraints are incorporated (the core of the dual-drive approach): a composite loss function is designed, adding a physical constraint term, loss_physics, to the mse. loss_total = loss_mse + λ * loss_physics; loss_physics can be: non-negativity constraint: the amplitude of the distance image should not be negative; energy conservation constraint: the total energy of the network output should be approximately equal to the total energy of the input; data consistency constraint: the network output f(x;θ) is then substituted into a simplified physical model (such as a Fourier transform) to see if it is consistent with the original input x, forcing the network output to conform to physical laws. λ is a hyperparameter used to balance the importance of the data fitting term and the physical constraint term.

[0105] Understandably, the loss function acts as a guide for training. It not only tells the network "your prediction is wrong," but also tells it "your prediction violates physical laws" through physical constraints, thereby greatly narrowing the network's search space and guiding it to find a solution that conforms to physical reality more quickly.

[0106] The process involves calculating the loss using a defined loss function to determine the error between the predicted and true values. Backpropagation calculates the gradient of the loss with respect to the parameters of each network layer. The optimizer updates all weights and biases of the network based on the gradient. Validation and tuning are performed after each training epoch, using a validation set (data not used in training) to evaluate model performance, monitor for overfitting, and adjust hyperparameters such as the learning rate and batch size accordingly.

[0107] Step S423: Update the parameters of the network model to be trained according to the error to obtain a deep learning neural network model.

[0108] Specifically, step S423 is implemented as follows: Forward propagation: A batch of input data x is fed into the network, and the output f(x;θ) is calculated. Loss calculation: The error between f(x;θ) and y is calculated using the total loss function. Backpropagation: Using the chain rule of calculus, the gradient (derivative) of the loss function with respect to each network parameter θ is calculated. Each gradient indicates the direction and adjustment of each parameter to reduce the error. Parameter update: All parameters θ are updated using an optimizer (such as Adam) based on the gradient information. Iterative loop: The steps are repeated until the model's performance on the validation set no longer improves (i.e., the model converges).

[0109] Understandably, physical preprocessing provides structured input to the DNN: the traditional radar signal processing chain (mixing, filtering, FFT) is not discarded, but rather acts as a feature extractor, transforming the raw, verbose time-domain signal into a frequency-domain representation (range profile) with explicit physical meaning. This provides the DNN with a strong prior, allowing it to focus on the specific task of super-resolution in the frequency domain instead of learning everything from raw ADC sampling, significantly reducing the learning difficulty.

[0110] DNNs, as complex mapping functions, address the issue that physical models perform poorly at low signal-to-noise ratios. DNNs leverage their powerful nonlinear fitting capabilities to learn complex mappings from noisy and bandwidth-limited spectra to ideal high-resolution spectra. The residual structure ensures the learnability of this mapping within deep networks.

[0111] Dual-drive fusion: Physical drive is reflected in the design of the signal processing chain, the network structure's imitation of physical processes (e.g., residual learning is similar to deconvolution), and the loss function that may incorporate physical constraints. Data drive is reflected in the network's autonomous learning from large amounts of data on how best to compensate for information loss caused by noise and bandwidth limitations.

[0112] The training and building cycle: Model building (architecture design) and training (parameter optimization) are a tightly coupled, iterative process. After the initial model is trained, its performance is fed back to the designer, which may trigger adjustments to the model structure (such as increasing depth or changing the number of residual blocks), and then it is trained again until satisfactory performance is achieved.

[0113] Ultimately, this trained model can be deployed in a real-world system, taking low signal-to-noise ratio FFT spectrum as input in real time and quickly outputting a clear, high-resolution range profile, thereby improving the radar's target recognition and resolution capabilities.

[0114] It's important to note that the output f(x;θ) of the neural network is a super-resolution spectrum. This spectrum is formally identical to the input spectrum x (e.g., both are 1x1024 vectors), but its content has undergone denoising and super-resolution processing, making it clearer and sharper. The resulting super-resolution radar range profile is also a spectrum; in the context of this model, the super-resolution radar range profile is this super-resolution spectrum f(x;θ). The range profile and spectrum are two names for the same thing, or two representations of the same data. In frequency-domain-based radar systems, the range profile and spectrum are equivalent.

[0115] Specifically, in the network architecture of a deep learning neural network model: Input layer: Receives the preprocessed spectral vector. Encoding / downsampling path (feature extraction module): Uses multiple one-dimensional convolutional layers and pooling layers; Convolutional layers: Like filters, they slide across the spectrum, automatically learning and extracting useful local feature patterns (e.g., the spectral peaks of a real target typically have a specific shape and width, while noise is cluttered); Pooling layers: Gradually reduce the spatial dimension of the data (i.e., the number of distance units), extracting more abstract, global features while reducing computational cost. Residual module: Inserts residual blocks between convolutional layers. A residual block typically contains two convolutional layers with a shortcut connection that directly adds the input to the output. Decoding / Upsampling Path (Reconstruction Layer): Uses transposed convolutional layers or upsampling layers; its function is the opposite of the encoding path, gradually restoring the abstract feature map back to the original range image size. In this process, the network combines and refines the learned features, ultimately generating a high-resolution output, producing the final super-resolution radar range image, whose size may be larger than the input spectrum (e.g., the input is 256 points, and the output is 512 or 1024 points), thus achieving super-resolution.

[0116] It's important to note that residual modules are used because they address the vanishing gradient problem in deep networks, allowing for very deep training (tens or even hundreds of layers). Deeper networks mean they can learn more complex mappings. More importantly, shortcut connections allow the network to focus on learning residuals—the differences between the input and output. This makes it easier for the network to learn fine-tuning tasks like denoising and super-resolution, rather than learning a completely new and complex mapping from scratch.

[0117] Step S43: The input layer inputs the frequency domain spectrum to the feature extraction module.

[0118] Specifically, step S43 involves feeding the input spectrum x_in into the neural network, using the low-quality spectrum x_in obtained in the previous step as input, and then feeding it into the pre-trained deep neural network model. This network has been trained with massive amounts of data and has learned how to recover clear and clean images from blurry, noisy images.

[0119] Step S44: The feature extraction module extracts feature data from the frequency domain spectrum.

[0120] Specifically, step S44 involves a nonlinear transformation within the network, where x_in flows through a series of carefully designed layers (such as convolutional layers, residual blocks, activation functions, etc.). This can be understood as feature extraction: the first few layers of the network (the encoder part) act like a detective, analyzing the input x_in and identifying potential target peaks, random noise, and patterns of multiple overlapping targets.

[0121] Step S45: The residual learning module corrects the frequency domain spectrum to obtain corrected data.

[0122] Specifically, step S45 is implemented as follows: Learning the correction factor: Since the network structure includes a residual network module, its core task is not to generate a perfect spectrum from scratch, but to learn a correction factor or residual. That is, it calculates how x_in needs to be modified to become sharp. This correction factor contains two key pieces of information: denoising: the noise that needs to be subtracted; super-resolution: the high-frequency details that need to be added to make the target peak sharper.

[0123] Step S46: The reconstruction layer obtains a super-resolution radar range image based on the feature data and the correction data.

[0124] Specifically, step S46 is implemented as follows: feature reconstruction. The later layers of the network (decoder part) begin to gradually reconstruct a brand new, high-quality spectrum based on the extracted features and the calculated correction amount.

[0125] It should be noted that this application employs a dual-drive approach of physics and data. Throughout the process, domain knowledge (physics-driven) ensures that the network's output still conforms to the fundamental physical laws of radar signals and does not fabricate results. For example, it may constrain the output spectrum to be real and non-negative. Massive data training (data-driven) allows the network to learn the mapping patterns from fuzzy to clear in various complex scenarios, a capability that traditional physical models struggle to achieve.

[0126] Understandably, the input spectrum x is low-resolution and has a low signal-to-noise ratio, while the super-resolution spectrum f(x;θ) output by the neural network is high-resolution and has a high signal-to-noise ratio. Although both are in the form of "spectrum" (i.e., a complex or real vector), their intrinsic quality and the amount of information they contain are vastly different.

[0127] Traditional radar obtains range profiles through FFT, and its resolution is physically limited by the radar signal bandwidth (resolution ≈ c / (2 * bandwidth)). If the bandwidth of the radar's transmitted signal is limited, the spectral peaks after FFT will naturally be wide, much like taking a picture with a low-resolution camera—it's inherently difficult to capture details. Furthermore, channel noise, system nonlinearity, and other factors further degrade the spectral quality, reducing its signal-to-noise ratio (SNR). Imagine x as a raw photograph with very low resolution and filled with noise and blur (low SNR). You can barely make out a person in the photo, but facial features, expressions, and other details are completely indistinguishable.

[0128] Neural networks enhance spectral quality through two key capabilities: Denoising: Trained on massive amounts of data, the network learns the characteristics of a real signal (such as a sharp peak) and the characteristics of noise (random, irregular fluctuations). When it sees an input x, it acts like an experienced filter, precisely removing noise components and preserving the signal components. Super-resolution: This is an even more crucial step. The network doesn't just filter; it reconstructs. Leveraging domain knowledge: The physics-data dual-drive and domain knowledge incorporated into the model mean that the network not only looks at data but also understands the physical imaging mechanism of radar signals. It knows what an ideal point target should look like in the spectrum. Learning the inverse problem: Recovering high-resolution information from a signal with limited bandwidth is essentially an inverse problem. By learning from a large amount of low-resolution spectrum -> high-resolution spectrum pairing data, the neural network learns how to infer and reconstruct the most likely sharp peak behind a blurry peak. This process transcends the physical limitations of traditional FFT. Continuing with the photographic analogy, the neural network f(x;θ) is like a top-notch photo editor: it first removes all noise from the photo (denoising). Then, based on his deep understanding of human faces (domain knowledge), he redrawn the details of the blurred areas, such as making the blurred eyes bright and clear, and making the blurred outlines sharp and distinct (super-resolution). The final output photo f(x;θ), although still the same photo, has a much higher clarity and information content than the original photo x.

[0129] Therefore, although the input spectrum x and the output spectrum f(x;θ) have the same form (both are vectors), their intrinsic qualities are completely different. The neural network, through its powerful nonlinear modeling ability, extracts and creates high-quality outputs from low-quality inputs.

[0130] Referring to Figure 11, SNR is the signal-to-noise ratio and FNR is the false negative rate. In Figure 12, the red dashed line represents the true location of the target. The technical effects that this invention can bring are as follows: In dense target scenes, the target false detection rate is reduced by more than 10% compared with traditional methods (periodic graph, MUSIC, etc.).

[0131] Taking five target scattering points as an example, when the signal-to-noise ratio is 0dB, the false negative rate is reduced by more than 7% compared to the periodogram method and by more than 17% compared to the multi-signal classification (MUSIC) method. Taking five target scattering points as an example, when the signal-to-noise ratio is 10dB, the false negative rate is reduced by more than 12% compared to the periodogram method and by more than 16% compared to the MUSIC method.

[0132] Furthermore, as shown in Figure 13, based on the above-mentioned deep learning-based radar range image super-resolution method, the present invention also provides a deep learning-based radar range image super-resolution system, wherein the deep learning-based radar range image super-resolution system includes: a signal receiving module 50, used to acquire simulated echo information reflected by the target; a baseband processing module 60, used to process the simulated echo information to obtain baseband information; a spectrum estimation module 70, used to obtain the frequency domain spectrum based on the baseband information; and a neural network model construction and spectrum output module 80, used to construct a deep learning neural network model, input the frequency domain spectrum into the deep learning neural network model, and output a super-resolution radar range image.

[0133] As shown in Figure 14, another embodiment of the radar range image super-resolution system based on deep learning in this invention is described. In this embodiment, the signal receiving module 50 includes: an electromagnetic wave transmitting unit 501, which is connected to the radar transmitter to generate a target electrical signal and converts the target electrical signal into electromagnetic waves radiated in space through a transmitting antenna; a reflected echo receiving unit 502, which receives a portion of the electromagnetic waves reflected from the target object through a receiving antenna and converts the portion of the electromagnetic waves into analog electrical signals; and an analog signal front-end processing unit 503, which performs front-end processing on the analog electrical signals to obtain target echo signals and noise clutter signals.

[0134] In this embodiment, the baseband processing module 60 includes: an analog-to-digital conversion unit 601, used to convert the analog echo information into a digital echo signal; a mixing unit 602, used to obtain an intermediate frequency signal based on the digital echo signal; and a low-pass filtering unit 603, used to perform low-pass filtering on the intermediate frequency signal to obtain a baseband signal.

[0135] In this embodiment, the spectrum estimation module 70 includes: a data windowing unit 701, used to window the baseband signal to obtain a time-domain signal data block; and a fast Fourier transform processing unit 702, used to perform a fast Fourier transform on the time-domain signal data block to obtain a frequency-domain spectrum.

[0136] In this embodiment, the neural network model construction and spectrum output module 80 includes: a network model construction unit 801, used to acquire training spectrum and corresponding ground truth labels to construct a network model to be trained; a network model training unit 802, used to train the network model to be trained according to the training spectrum and the ground truth labels to obtain a deep learning neural network model; a spectrum input unit 803, used by the input layer to input the frequency domain spectrum to the feature extraction module; a feature extraction unit 804, used by the feature extraction module to extract feature data from the frequency domain spectrum; a spectrum correction unit 805, used by the residual learning module to correct the frequency domain spectrum to obtain corrected data; and a feature reconstruction unit 806, used by the reconstruction layer to obtain a super-resolution radar range image according to the feature data and the corrected data.

[0137] This invention designs different deep learning-based radar range image super-resolution systems, thereby improving the radar range image resolution under low signal-to-noise ratio conditions.

[0138] Furthermore, as shown in Figure 15, based on the aforementioned deep learning-based radar range image super-resolution method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 15 only shows some components of the terminal; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0139] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a deep learning-based radar range image super-resolution program 40, which can be executed by the processor 10 to implement the deep learning-based radar range image super-resolution method of this application.

[0140] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the deep learning-based radar range image super-resolution method.

[0141] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminal's processor 10, memory 20, and display 30 communicate with each other via a system bus.

[0142] In one embodiment, when the processor 10 executes the deep learning-based radar range image super-resolution program 40 in the memory 20, it performs the following steps: acquiring simulated echo information reflected by the target; processing the simulated echo information to obtain baseband information; obtaining the frequency domain spectrum based on the baseband information; constructing a deep learning neural network model, inputting the frequency domain spectrum into the deep learning neural network model, and outputting a super-resolution radar range image.

[0143] The simulated echo information includes target echo signals and noise clutter signals. Acquiring the simulated echo information reflected by the target specifically includes: after the radar transmitter generates a target electrical signal, converting the target electrical signal into electromagnetic waves radiated in space via a transmitting antenna; receiving a portion of the electromagnetic waves reflected from the target object via a receiving antenna, and converting the portion of the electromagnetic waves into simulated electrical signals; and performing front-end processing on the simulated electrical signals to obtain the target echo signal and noise clutter signals.

[0144] Wherein, the baseband information is a baseband signal; the process of processing the analog echo information to obtain the baseband information specifically includes: converting the analog echo information into a digital echo signal; obtaining an intermediate frequency signal based on the digital echo signal; and performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal.

[0145] Specifically, converting the analog echo information into a digital echo signal involves converting the analog echo information into a digital echo signal using an analog-to-digital converter, wherein the analog echo information is a high-frequency analog signal and the digital echo signal is a high-frequency digital signal.

[0146] Specifically, obtaining the intermediate frequency signal based on the digital echo signal includes: acquiring the radar's transmitted signal and using the transmitted signal as a reference signal; and mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal.

[0147] The intermediate frequency signal includes a sum frequency component and a difference frequency component; the process of mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal specifically involves multiplying the digital echo signal with the reference signal to obtain the sum frequency component and the difference frequency component of the digital echo signal and the reference signal.

[0148] Wherein, the baseband signal is a difference frequency component; the step of performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal specifically involves filtering the intermediate frequency signal through a low-frequency filter to obtain a low-frequency difference frequency component.

[0149] Specifically, obtaining the frequency domain spectrum based on the baseband information includes: windowing the baseband signal to obtain a time-domain signal data block; and performing a fast Fourier transform on the time-domain signal data block to obtain the frequency domain spectrum.

[0150] Wherein, the frequency domain spectrum is a spectrum diagram; the step of performing a fast Fourier transform on the time domain signal data block to obtain the frequency domain spectrum specifically includes: performing a fast Fourier transform on the time domain signal data block to calculate a complex array; extracting the corresponding frequency and amplitude from the complex array; and obtaining the spectrum diagram corresponding to the baseband signal based on the frequency and amplitude.

[0151] The deep learning neural network model includes an input layer, a feature extraction module, a residual learning module, and a reconstruction layer.

[0152] Specifically, constructing the deep learning neural network model includes: obtaining the training spectrum and the corresponding ground truth labels, constructing a network model to be trained; and training the network model to be trained based on the training spectrum and the ground truth labels to obtain the deep learning neural network model.

[0153] The step of training the network model to be trained based on the training spectrum and the ground truth label to obtain a deep learning neural network model specifically includes: inputting the training spectrum into the network model to be trained to obtain a predicted spectrum; defining a loss function and calculating the error between the predicted spectrum and the ground truth label based on the loss function; and updating the parameters of the network model to be trained based on the error to obtain a deep learning neural network model.

[0154] Specifically, the step of inputting the frequency domain spectrum into the deep learning neural network model and outputting a super-resolution radar range image includes: the input layer inputting the frequency domain spectrum into the feature extraction module; the feature extraction module extracting feature data from the frequency domain spectrum; the residual learning module correcting the frequency domain spectrum to obtain corrected data; and the reconstruction layer obtaining the super-resolution radar range image based on the feature data and the corrected data.

[0155] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a deep learning-based radar range image super-resolution program, which, when executed by a processor, implements the steps of the deep learning-based radar range image super-resolution method as described above.

[0156] In summary, this invention provides a super-resolution radar range image method, system, terminal, and computer-readable storage medium based on deep learning. The method includes: acquiring simulated echo information reflected by a target; processing the simulated echo information to obtain baseband information; obtaining a frequency domain spectrum based on the baseband information; constructing a deep learning neural network model; inputting the frequency domain spectrum into the deep learning neural network model; and outputting a super-resolution radar range image. This invention extracts a baseband signal containing target range information by mixing and processing the simulated echo signal with baseband data. A fast Fourier transform is performed on the baseband signal to obtain a low-resolution range image spectrum. This low-resolution range image spectrum is then input into a physical-data dual-driven deep neural network (the core module includes a residual network). The network integrates a radar signal physical model (such as echo generation mechanism and noise characteristics) with data-driven feature learning. Through training, it learns the mapping from the low-resolution spectrum to the high-resolution range image, outputting a super-resolution radar range image. This application enhances feature extraction capabilities through the residual network module, improving the resolution and noise resistance of multi-target range images under low signal-to-noise ratio conditions, and reducing the false negative rate of dense targets.

[0157] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0158] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0159] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A deep learning-based super-resolution method for radar range images, characterized in that, The deep learning-based super-resolution radar range image method includes: acquiring simulated echo information reflected by the target; processing the simulated echo information to obtain baseband information; obtaining the frequency domain spectrum based on the baseband information; constructing a deep learning neural network model; inputting the frequency domain spectrum into the deep learning neural network model; and outputting a super-resolution radar range image.

2. The deep learning-based radar range image super-resolution method according to claim 1, characterized in that, The simulated echo information includes target echo signals and noise clutter signals; the acquisition of simulated echo information reflected by the target specifically includes: after the radar transmitter generates a target electrical signal, converting the target electrical signal into electromagnetic waves radiated in space through a transmitting antenna; receiving a portion of the electromagnetic waves reflected from the target object through a receiving antenna, and converting the portion of the electromagnetic waves into simulated electrical signals; and performing front-end processing on the simulated electrical signals to obtain the target echo signal and noise clutter signals.

3. The deep learning-based radar range image super-resolution method according to claim 2, characterized in that, The baseband information is a baseband signal; the process of processing the analog echo information to obtain the baseband information specifically includes: converting the analog echo information into a digital echo signal; obtaining an intermediate frequency signal based on the digital echo signal; and performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal.

4. The deep learning-based radar range image super-resolution method according to claim 3, characterized in that, The step of converting the analog echo information into a digital echo signal specifically involves converting the analog echo information into a digital echo signal using an analog-to-digital converter, wherein the analog echo information is a high-frequency analog signal and the digital echo signal is a high-frequency digital signal.

5. The deep learning-based radar range image super-resolution method according to claim 3, characterized in that, The step of obtaining the intermediate frequency signal based on the digital echo signal specifically includes: acquiring the radar's transmitted signal and using the transmitted signal as a reference signal; and mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal.

6. The deep learning-based radar range image super-resolution method according to claim 5, characterized in that, The intermediate frequency signal includes a sum frequency component and a difference frequency component; the process of mixing the digital echo signal with the reference signal to obtain the intermediate frequency signal specifically involves multiplying the digital echo signal with the reference signal to obtain the sum frequency component and the difference frequency component of the digital echo signal and the reference signal.

7. The deep learning-based radar range image super-resolution method according to claim 6, characterized in that, The baseband signal is a difference frequency component; the step of performing low-pass filtering on the intermediate frequency signal to obtain the baseband signal specifically involves filtering the intermediate frequency signal through a low-frequency filter to obtain a low-frequency difference frequency component.

8. The deep learning-based radar range image super-resolution method according to claim 3, characterized in that, The step of obtaining the frequency domain spectrum based on the baseband information specifically includes: windowing the baseband signal to obtain a time domain signal data block; and performing a fast Fourier transform on the time domain signal data block to obtain the frequency domain spectrum.

9. The deep learning-based radar range image super-resolution method according to claim 8, characterized in that, The frequency domain spectrum is a spectrum diagram; the step of performing a fast Fourier transform on the time domain signal data block to obtain the frequency domain spectrum specifically includes: performing a fast Fourier transform on the time domain signal data block to calculate a complex array; extracting the corresponding frequency and amplitude from the complex array; and obtaining the spectrum diagram corresponding to the baseband signal based on the frequency and amplitude.

10. The deep learning-based radar range image super-resolution method according to claim 8, characterized in that, The constructed deep learning neural network model includes an input layer, a feature extraction module, a residual learning module, and a reconstruction layer.

11. The deep learning-based radar range image super-resolution method according to claim 10, characterized in that, The construction of the deep learning neural network model specifically includes: obtaining the training spectrum and the corresponding ground truth labels, constructing a network model to be trained; and training the network model to be trained based on the training spectrum and the ground truth labels to obtain the deep learning neural network model.

12. The deep learning-based radar range image super-resolution method according to claim 11, characterized in that, The step of training the network model to be trained based on the training spectrum and the ground truth label to obtain a deep learning neural network model specifically includes: inputting the training spectrum into the network model to be trained to obtain a predicted spectrum; defining a loss function and calculating the error between the predicted spectrum and the ground truth label based on the loss function; and updating the parameters of the network model to be trained based on the error to obtain a deep learning neural network model.

13. The deep learning-based radar range image super-resolution method according to claim 10, characterized in that, The step of inputting the frequency domain spectrum into the deep learning neural network model and outputting a super-resolution radar range image specifically includes: the input layer inputting the frequency domain spectrum into the feature extraction module; the feature extraction module extracting feature data from the frequency domain spectrum; the residual learning module correcting the frequency domain spectrum to obtain corrected data; and the reconstruction layer obtaining the super-resolution radar range image based on the feature data and the corrected data.

14. A radar range image super-resolution system based on deep learning, characterized in that, The deep learning-based radar range image super-resolution system includes: a signal receiving module for acquiring simulated echo information reflected by a target; a baseband processing module for processing the simulated echo information to obtain baseband information; a spectrum estimation module for obtaining a frequency domain spectrum based on the baseband information; and a neural network model construction and spectrum output module for constructing a deep learning neural network model, inputting the frequency domain spectrum into the deep learning neural network model, and outputting a super-resolution radar range image.

15. The deep learning-based radar range image super-resolution system according to claim 14, characterized in that, The signal receiving module includes: an electromagnetic wave transmitting unit, which is connected to the radar transmitter to generate a target electrical signal and converts the target electrical signal into electromagnetic waves radiated in space through a transmitting antenna; a reflected echo receiving unit, which receives part of the electromagnetic waves reflected from the target object through a receiving antenna and converts part of the electromagnetic waves into analog electrical signals; and an analog signal front-end processing unit, which performs front-end processing on the analog electrical signals to obtain target echo signals and noise clutter signals.

16. The deep learning-based radar range image super-resolution system according to claim 14, characterized in that, The baseband processing module includes: an analog-to-digital conversion unit for converting the analog echo information into a digital echo signal; a mixing unit for obtaining an intermediate frequency signal based on the digital echo signal; and a low-pass filtering unit for performing low-pass filtering on the intermediate frequency signal to obtain a baseband signal.

17. The deep learning-based radar range image super-resolution system according to claim 16, characterized in that, The spectrum estimation module includes: a data windowing unit for windowing the baseband signal to obtain a time-domain signal data block; and a fast Fourier transform processing unit for performing a fast Fourier transform on the time-domain signal data block to obtain a frequency-domain spectrum.

18. The deep learning-based radar range image super-resolution system according to claim 17, characterized in that, The neural network model construction and spectrum output module includes: a network model construction unit for acquiring training spectrum and corresponding ground truth labels to construct a network model to be trained; a network model training unit for training the network model to be trained based on the training spectrum and the ground truth labels to obtain a deep learning neural network model; a spectrum input unit for the input layer to input the frequency domain spectrum to the feature extraction module; a feature extraction unit for the feature extraction module to extract feature data from the frequency domain spectrum; a spectrum correction unit for the residual learning module to correct the frequency domain spectrum to obtain corrected data; and a feature reconstruction unit for the reconstruction layer to obtain a super-resolution radar range image based on the feature data and the corrected data.

19. A terminal, characterized in that, The terminal includes: a memory, a processor, and a deep learning-based radar range super-resolution program stored in the memory and executable on the processor. When the deep learning-based radar range super-resolution program is executed by the processor, it implements the steps of the deep learning-based radar range super-resolution method as described in any one of claims 1-13.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a deep learning-based radar range image super-resolution program, which, when executed by a processor, implements the steps of the deep learning-based radar range image super-resolution method as described in any one of claims 1-13.

Images