Depth unfolding driven adaptive dither modulation single-bit millimeter wave imaging method

Abstract

A depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method includes: constructing a pairwise training dataset based on root mean square normalization and extracting one-dimensional statistical features; inputting the one-dimensional statistical features into the constructed adaptive jitter parameter prediction sub-network to generate adaptive jitter amplitude coefficients, and performing differentiable single-bit quantization with a pass-through estimator mechanism in the differentiable layer to obtain complex single-bit quantized echo data; constructing a deep unfolding network by combining Gaussian space energy prior and adaptive soft thresholding, inputting the complex single-bit quantized echo data into the deep unfolding network, and outputting a reconstructed image; performing end-to-end fine-tuning and joint training of the overall network, including the adaptive jitter parameter prediction sub-network, the differentiable layer, and the deep unfolding network, using the pairwise training dataset based on gradient isolation mechanism and joint loss function; acquiring echo features from the real scene, inputting them into the trained and converged network model, and outputting a three-dimensional millimeter-wave reconstructed image.

Description

Technical Field

[0001] This invention relates to the field of radar imaging and low-bit quantization signal processing technology, specifically to a depth-deployment-driven adaptive jitter modulation single-bit millimeter-wave imaging method. Background Technology

[0002] Near-field millimeter-wave 3D imaging technology has high application value in fields such as public security inspection, non-destructive testing, and target recognition. To obtain good image quality, existing millimeter-wave imaging systems typically employ high-speed, high-resolution analog-to-digital converters (ADCs) to acquire the amplitude and phase information of the echo signal. However, high-precision ADCs often come with high hardware costs, power consumption, and data transmission burdens, which to some extent limits the application of millimeter-wave imaging systems in portable, lightweight, and low-power scenarios.

[0003] To reduce sampling and storage overhead, single-bit quantization sampling technology has been introduced into radar imaging systems. Single-bit sampling retains only the symbolic information of the echo signal relative to a comparison threshold, enabling data compression with lower hardware complexity. However, traditional zero-threshold single-bit hard quantization is a nonlinear truncation process, which easily leads to the loss of echo amplitude information and introduces higher harmonics, intermodulation components, and structured quantization errors. During near-field three-dimensional focusing, these nonlinear errors may accumulate with the imaging operator, causing problems such as structured artifacts, dynamic range compression, and reduced discernibility of weak scattering regions in the reconstructed image.

[0004] Dither modulation is a feasible method to mitigate nonlinear distortion in low-bit quantization. The basic idea is to inject controlled random perturbations into the forward echo channel of single-bit quantization, causing the partially deterministic nonlinear errors generated by hard truncation to exhibit randomized characteristics, thereby reducing the coherent accumulation of harmonic components and structured artifacts during imaging. For single-bit millimeter-wave imaging, the dither intensity setting directly affects the quantization error structure and the quality of the reconstructed image.

[0005] In practical near-field millimeter-wave imaging scenarios, radar echo amplitude is significantly affected by factors such as target reflectivity, range attenuation, target attitude, system gain, noise, and clutter, exhibiting clear scene dependence. If jitter parameters remain constant, it becomes difficult to adapt to different targets and different echo amplitude ranges. When jitter intensity is low, its effect on reducing single-bit hard-truncation nonlinear errors is limited, and harmonic artifacts may still exist in the image. When jitter intensity is high, although it helps reduce some coherence distortion, it may also introduce additional random perturbations, increasing image noise and affecting the reconstruction of weakly scattering targets.

[0006] Therefore, fixed-parameter jitter mechanisms present a trade-off between harmonic artifact suppression and noise floor control, making it difficult to adapt to variations in the dynamic range of echoes in complex scenarios. On the other hand, relying on unquantized high-precision echo calculations of standard deviation or variance characteristics during the actual imaging phase diminishes the advantages of low hardware complexity and low data volume of single-bit sampling systems. Therefore, it is necessary to propose a single-bit millimeter-wave imaging method that utilizes lightweight single-bit pre-sampling information to estimate the coarse scale of the echo and adaptively configures the jitter modulation parameters accordingly.

[0007] Furthermore, traditional iterative reconstruction algorithms typically rely on manually setting the iteration step size, regularization parameters, and threshold strategies, resulting in limited ability to distinguish between target and background regions when single-bit quantization errors are significant. The depth unfolding method maps the physical iterative algorithm to a trainable network with a fixed number of layers, introducing a data-driven parameter learning mechanism while preserving the constraints of the imaging physical model. Combining adaptive jitter modulation with depth unfolding reconstruction helps improve the recovery of harmonic artifacts, background clutter, and weak scattering details in single-bit millimeter-wave imaging. Summary of the Invention

[0008] To address the challenges of existing single-bit millimeter-wave imaging methods that rely on fixed jitter parameters to simultaneously suppress harmonic artifacts and control noise floor, and the unsuitability of relying on unquantized high-precision echo calculation statistical features in real-world applications, this invention provides a depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method.

[0009] The basic idea of ​​this invention is as follows: In the training phase, stable amplitude statistical features are extracted using high-precision complex echo data, and an end-to-end mapping relationship is constructed between the jitter parameter prediction module, the differentiable single-bit quantization layer, and the deep unfolding reconstruction network. In the real application phase, lightweight over-threshold statistical information is obtained through fixed-threshold single-bit presampling. This statistical information is used to estimate the coarse amplitude scale of the current scene echo and predict the jitter modulation parameters accordingly. Subsequently, in the formal acquisition phase, controlled jitter is injected into the echo channel and single-bit quantization is performed. The obtained single-bit observation data is input into the trained deep unfolding reconstruction network to obtain the millimeter-wave imaging results.

[0010] The parametric jitter modulation in this invention refers to adaptively determining the amplitude parameters of the jitter signal based on the characteristics of the echo amplitude scale, and injecting the jitter signal into the pre-quantization echo channel, so that the partial deterministic nonlinear error introduced by single-bit hard quantization presents randomized characteristics, thereby reducing the influence of harmonic components and structured artifacts in the imaging process.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method includes the following steps:

[0013] S1. Construct a pairwise training dataset based on root mean square normalization and extract one-dimensional statistical features;

[0014] S2. The adaptive jitter parameter prediction subnetwork constructed by inputting one-dimensional statistical features generates adaptive jitter amplitude coefficients, and performs differentiable single-bit quantization with a pass-through estimator mechanism in the differentiable quantization layer to obtain complex single-bit quantized echo data.

[0015] S3. A deep unfolding network is constructed by combining Gaussian spatial energy prior and adaptive soft thresholding. Complex single-bit quantized echo data is input into the deep unfolding network and the reconstructed image is output.

[0016] S4. Based on the gradient isolation mechanism and joint loss function, the overall network, including the adaptive jitter parameter prediction sub-network, differentiable layer and deep unfolding network, is fine-tuned and jointly trained using paired training datasets.

[0017] S5. Obtain the echo features of the real scene, input them into the trained and converged network model, and output the 3D millimeter wave reconstructed image.

[0018] Preferably, S1 includes:

[0019] The radar antenna is controlled to scan and sample on a planar rectangular grid. The antenna receives the raw echo and obtains a three-dimensional complex echo tensor. The global root mean square value of the three-dimensional complex echo tensor is calculated and normalized to obtain a normalized complex echo. The standard deviations of the real and imaginary parts of the normalized complex echo are calculated in the global space. The obtained standard deviations of the real and imaginary parts are concatenated to form a one-dimensional statistical feature. A pairwise training dataset containing the one-dimensional statistical feature, the corresponding normalized complex echo, and the corresponding reference image is constructed.

[0020] As a preferred embodiment, S1 also includes:

[0021] A single-bit presampled normalized complex echo was performed using a fixed threshold. The threshold occupancy rate was statistically analyzed, and an offline mapping relationship between the threshold occupancy rate and the echo amplitude scale was established.

[0022] Preferably, in S2, the adaptive jitter parameter prediction subnetwork is composed of a multi-layer sensing mechanism, which includes a linear mapping layer, a nonlinear activation layer and an output mapping layer in sequence; the adaptive jitter amplitude coefficient corresponds to the current echo amplitude dispersion state.

[0023] Preferably, S2 includes:

[0024] Multiply the uniform random matrix by the adaptive jitter amplitude coefficient and then multiply by the real part standard deviation and the imaginary part standard deviation respectively to construct the real part random jitter matrix and the imaginary part random jitter matrix respectively.

[0025] The real part random jitter matrix and the imaginary part random jitter matrix are injected into the real part channel and the imaginary part channel of the normalized complex echo data, respectively, to obtain the complex signal after jitter injection;

[0026] The real and imaginary parts of the complex signal after jitter are quantized separately using differentiable single-bit quantization. The differentiable single-bit quantization outputs of the real and imaginary parts are combined to obtain complex single-bit quantized echo data.

[0027] Preferably, in S2, the differentiable single-bit quantization processing includes:

[0028] Forward propagation uses a hard sign function to truncate the output, while backward propagation uses a hyperbolic tangent smoothing function as a gradient substitute. The output of differentiable single-bit quantization is the sum of the difference between the hard quantization result and the soft quantization result after gradient truncation and the soft quantization result.

[0029] Preferably, S3 includes:

[0030] Range compression processing is performed on complex single-bit quantized echo data to extract two-dimensional observation slices at the reference depth;

[0031] An imaging operator based on an electromagnetic wave propagation model performs an initial inversion on a two-dimensional observation slice to obtain an initial scattering coefficient estimate.

[0032] The initial scattering coefficient estimate is input into a depth-unfolded network containing K-level cascade nodes for iterative reconstruction. After K-level node iteration, the reconstructed image is output.

[0033] As a preferred approach, in S3, each node of the deep unfolded network performs the following operations:

[0034] Based on the forward observation operator and the imaging operator, data consistency is updated with a learnable step size to obtain the intermediate state image. The amplitude of the intermediate state image is spatially convolved with a Gaussian kernel with a learnable variance parameter to obtain a local spatial energy density prior map. A learnable baseline threshold is introduced, and a dynamic inverse proportional threshold mapping map is obtained based on the local spatial energy density prior map. The intermediate state image is then truncated using a complex-domain nonlinear soft thresholding method with the dynamic inverse proportional threshold mapping map to obtain the reconstructed image of the current node.

[0035] Preferably, S4 includes:

[0036] The mean square error of the amplitudes of the reconstructed image and the reference image after normalization to the maximum value is used as the shape fidelity loss, where the normalization denominator is subjected to gradient isolation processing; the global mean of the amplitude of the reconstructed image is used as the amplitude sparsity penalty loss; a total loss function is constructed based on the shape fidelity loss and the amplitude sparsity penalty loss, and the overall network is jointly optimized.

[0037] As a preferred embodiment, S5 includes:

[0038] The echo of the target under test is presampled with a fixed threshold single bit, and the threshold occupancy rate is calculated. A coarse amplitude scale estimate is obtained through offline mapping. Normalized scale features are constructed using the coarse amplitude scale estimate and input into the trained adaptive jitter parameter prediction sub-network to output the optimal jitter coefficient. A physical jitter signal is generated based on the optimal jitter coefficient and injected into the echo channel. The formal single-bit imaging data is obtained through a single-bit analog-to-digital converter. The single-bit imaging data is processed by distance compression and then input into the trained depth unfolding network to output a three-dimensional millimeter-wave reconstructed image.

[0039] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0040] This invention deeply integrates feature extraction with adaptive dithering, breaking the trade-off between harmonics and noise. Root mean square normalization and standard deviation extraction are used as the pre-driving kernels for adaptive dithering, allowing the predicted unified adaptive dithering intensity to dynamically adjust according to changes in the target's true characteristics. This effectively disrupts harmonic coherence accumulation caused by single-bit hard truncation while strictly suppressing the unwarranted rise in the system noise floor. For back-end sparse reconstruction, this invention introduces a Gaussian space prior filtering mechanism, enabling spatial awareness during the reconstruction process. By constructing an inverse proportional threshold, it strongly removes background clutter artifacts while imposing a minimal shrinkage penalty on the effective target region. This design benefits from the superior quantization data provided by the front-end adaptive dithering, enabling the back-end reconstruction network to more accurately distinguish between the target and the background, thus solving the problems of blurred target edges and loss of weak scattering points that are common in traditional sparse reconstruction algorithms.

[0041] In summary, the synergistic effect of various technical features ultimately achieved high-fidelity image restoration, significantly improving the adaptability and reconstruction quality of single-bit millimeter-wave imaging. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the method steps in Embodiment 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;

[0044] Figure 3 This is an optical diagram of the experimental scene in Embodiment 1 of the present invention. Detailed Implementation

[0045] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0046] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0047] To address the issues of strong nonlinear distortion and structured artifacts caused by zero-threshold hard truncation in existing single-bit millimeter-wave radar imaging, as well as the technical shortcomings of traditional fixed-parameter jitter mechanisms that are unable to simultaneously achieve harmonic suppression and noise floor rise control and cannot adapt to drastic changes in dynamic range in complex scenarios, this invention proposes a single-bit millimeter-wave imaging method with adaptive jitter parameter prediction.

[0048] Example 1:

[0049] like Figure 1 , Figure 2 The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method shown includes the following steps:

[0050] Step 1: Construct a pairwise training dataset based on root mean square normalization and extract one-dimensional statistical features. The specific implementation process is as follows:

[0051] like Figure 3 The experimental scenario diagram shows the radar antenna being set at... A rectangular grid scan is performed on the plane, and the antenna position is denoted as... Move the target area along the depth Divide into equidistant slices, and denot the target scattering point position as follows: .

[0052] After receiving the raw echo and adding Gaussian white noise with a preset signal-to-noise ratio, the antenna obtains the three-dimensional complex echo tensor through discrete sampling. .in, They are respectively Number of scanning sampling points in the direction, This represents the number of sampling points in the fast time.

[0053] To ensure that the front-end jitter parameter prediction network can stably extract features and generalize to different scenarios, it is essential to first standardize the energy benchmark of the input signal. Specifically, this involves calculating the three-dimensional complex echo tensor. global root mean square value :

[0054]

[0055] In the formula, The total number of elements in the echo tensor. This is a very small constant. Subsequently, the signal is scaled using this root mean square value to obtain a power reference-aligned normalized complex echo. This operation effectively eliminates the interference of absolute energy level fluctuations on network feature perception, enabling the normalized echo to have a stable energy scale across scenarios, providing standardized input for subsequent extraction of objective statistical scatter features.

[0056] Then calculate the normalized complex echo. The standard deviation of the real and imaginary parts in the global space is used to obtain the standard deviation of the real part. and the standard deviation of the imaginary part The two are concatenated to form a one-dimensional statistical feature. It is used to reflect the true fluctuation state of the echo.

[0057] Meanwhile, to enable the acquisition of similar scale inputs in real-world applications without obtaining complete high-precision echoes, a fixed-threshold single-bit presampling process is simulated on the training data to establish an offline mapping relationship between the threshold occupancy rate and the echo amplitude scale. This mapping relationship can be implemented using a lightweight estimation model, which can then be used to estimate coarse scale features from the single-bit presampling results in subsequent real-world scenarios.

[0058] Finally, by combining the corresponding high-precision target scattering coefficient distribution as labels, a pairwise training dataset containing "feature-echo-reference image" is constructed.

[0059] Step 2: Input the one-dimensional statistical features into the adaptive jitter parameter prediction sub-network to generate jitter coefficients, and perform differentiable single-bit quantization with a pass-through estimator mechanism to obtain complex single-bit quantized echo data; the specific implementation process is as follows:

[0060] Step 21: Construct a feature-driven adaptive jitter parameter prediction subnetwork;

[0061] The one-dimensional statistical features extracted in step 1 The input is fed into an adaptive jitter parameter prediction subnetwork composed of a multilayer perceptron (MLP). This network internally consists of a cascaded linear mapping layer, a nonlinear activation layer (such as the ReLU function), and an output mapping layer composed of a sigmoid activation function and a constant scaling factor. Through forward inference, the network automatically outputs a globally adaptive jitter amplitude coefficient that matches the dynamic range of the current echo scene based on the energy distribution of the input features. (in ).

[0062] Step 22: Utilize the predicted adaptive jitter amplitude coefficient Using the extracted standard deviation features of the real and imaginary parts, uniform random jitter matrices are constructed for the real and imaginary channels, respectively. and The specific generation rules are as follows:

[0063]

[0064]

[0065] In the formula, To have the same dimension as the echo tensor to be processed and to independently obey A uniformly distributed matrix of random numbers.

[0066] Subsequently, the generated jitter matrix is ​​injected into the normalized complex echo. In the real and imaginary channels, a complex signal modulated by controlled jitter is obtained. :

[0067]

[0068]

[0069] Step 23: Traditional zero-threshold single-bit quantization is limited by its step truncation characteristic; the derivative is zero everywhere during backpropagation, which can cause gradient breakage in deep neural networks, making end-to-end training impossible. To overcome this problem, this embodiment embeds a pass-through estimator (STE) mechanism within the quantization layer. For the injected jittered real or imaginary part matrix (collectively denoted as the input tensor)... Its forward physical hard symbol truncation operation and reverse smooth activation operation They are defined as follows:

[0070]

[0071]

[0072] In the formula, For symbol extraction functions, The hyperparameters for controlling the steepness of the hyperbolic tangent function (set in this embodiment) ).

[0073] In the computational graph of a neural network, the final differentiable single-bit quantized output Build as:

[0074]

[0075] in, This indicates that a gradient truncation / isolation operation is performed on the tensor within the parentheses.

[0076] This differentiable architecture design ensures that the network strictly adheres to purely physical single-bit hard quantization during forward propagation, outputting... The discrete data is used; however, during backpropagation gradient calculation, a smooth and everywhere differentiable Tanh function gradient smoothing is used instead of the originally non-differentiable step gradient. This mechanism successfully overcomes the gradient blocking bottleneck caused by hard quantization, allowing the reconstruction error of the subsequent imaging network to be smoothly backpropagated to the front-end MLP prediction network. After the real and imaginary parts are processed independently as described above, they are recombined to output the final complex single-bit quantized echo data. .

[0077] Step 3: Construct a deep unfolding network by combining Gaussian spatial energy prior and adaptive soft thresholding. Input the complex single-bit quantized echo data into the deep unfolding network and output the reconstructed image. The specific implementation process is as follows:

[0078] Receive the complex single-bit quantized echo data output from step 2. Along the fast time dimension Perform a Fast Fourier Transform to compress the range pulse, and adjust the range pulse based on the reference depth of the target. Extract the two-dimensional observation slice corresponding to the distance gate (bin), denoted as Subsequently, an imaging operator based on a propagation model was used. An initial inversion mapping is performed on the two-dimensional observation slice to obtain an initial scattering coefficient estimate for the target region. :

[0079]

[0080] In the formula, This indicates a normalization operation based on the maximum amplitude. Based on the near-field spherical wave to plane wave decomposition theory, an imaging operator based on a propagation model... Specifically defined as:

[0081]

[0082] In the formula, and They represent in Two-dimensional discrete Fourier transform and its inverse transform on a plane denoted as the spatial wavenumber in the depth direction.

[0083] Iterative updates for physical data consistency; expanding the traditional Iterative Shrink Thresholding (ISTA) algorithm into one that includes... A deep neural network with cascaded layers, in its deepest unfolded... layer( In this process, the data consistency gradient is first calculated based on the physical observation model, and then a learnable iterative step size parameter is used. Update intermediate state image :

[0084]

[0085] In the formula, This is the output image of the previous layer. The forward observation operator based on the propagation model. It is used to remap the spatial domain scattering coefficients back to the observation domain, and its specific definition is:

[0086]

[0087] This cascaded process ensures a high degree of consistency in the reconstructed image at the physical measurement propagation level.

[0088] Spatial energy prior extraction is based on learnable Gaussian kernels; conventional sparse reconstruction uses a global single penalty threshold, which easily leads to the filtering out of weak targets or blurring of edges. To achieve spatial perception, this embodiment... The layer introduces a learnable variance parameter. Two-dimensional Gaussian kernel :

[0089]

[0090] In the formula, This is a normalization constant used to ensure that the energy integral within the nucleus is 1.

[0091] Extract the complex magnitude matrix of the intermediate state image By using the aforementioned dynamic Gaussian kernel for two-dimensional spatial convolution smoothing, a priori map of local spatial energy density reflecting the true distribution of the target is calculated. :

[0092]

[0093] In the formula, This represents a two-dimensional convolution operation. To prevent division by zero by extremely small constants, this prior map precisely quantifies the energy concentration in various regions within the imaging field of view.

[0094] Inversely proportional driven adaptive complex soft thresholding activation; introduction of learnable baseline soft thresholding parameters. And combined with the prior map of local spatial energy density Construct a pixel-level dynamic inverse proportional threshold mapping map :

[0095]

[0096] The physical significance of this mechanism lies in: in regions where the target energy is dense ( Larger areas automatically reduce the penalty threshold to fully preserve target edges and details; in areas with pure background noise or artifacts ( Approaching 0), the penalty threshold is automatically amplified to forcefully suppress structured clutter.

[0097] Finally, this dynamic inverse proportional threshold mapping graph is used to... Perform complex-domain nonlinear soft-thresholding truncation to achieve image shrinkage, denoising, and phase preservation, and output the reconstructed image of this layer. :

[0098]

[0099] go through After cascading iterations, the network ultimately outputs a high-resolution single-bit millimeter-wave 3D target slice image. .

[0100] Step 4: Based on the gradient isolation mechanism and joint loss function, perform end-to-end fine-tuning and joint training on the overall network. The specific process is as follows:

[0101] The overall architecture (AdaD-Net) comprising a feature-aware MLP, differentiable layers, and a deep unfolded subnetwork is trained using the paired training dataset constructed in step 1. In each training iteration, normalized complex echoes and one-dimensional statistical features are used. The input network outputs a reconstructed complex target image via forward propagation, denoted as the predicted image. Simultaneously, utilizing unquantized, high-precision raw echo data, a fast Fourier transform and a physical adjoint operator are employed. Mapping, dynamically generating the corresponding target ground truth distribution, denoted as the high-precision reference image. .

[0102] When calculating the network reconstruction error, if extreme value normalization is directly used to calculate the mean squared error, the network optimizer will tend to maliciously increase the extreme points to take shortcuts and reduce the overall loss, resulting in gradient hijacking and abnormally compressing the overall amplitude of the reconstructed image. To solve this problem, this embodiment extracts the complex amplitude matrices of the predicted image and the reference image respectively before calculating the loss function. and Furthermore, during global maximum division normalization, a gradient isolation (Detach) operation is forced on the denominator constant term. The normalized predicted image amplitude... Amplitude of reference image They are represented as follows:

[0103]

[0104]

[0105] In the formula, This indicates that the link that propagates gradients back to the node in the computation graph is cut off. To prevent the elimination of zero minima, this mechanism forces the network to reduce errors by substantially optimizing the global shape of the image, rather than by opportunistic scaling. To balance target shape fidelity with sparsity suppression of background clutter, a shape fidelity loss mechanism is constructed. and amplitude sparse penalty loss The joint objective function is as follows:

[0106]

[0107]

[0108]

[0109] In the formula, Let be the mean square error function. Calculate the structural differences in the target shape using the extreme value normalized image after isolating the gradient; The physical significance of calculating the sparsity penalty using the mean operation is that it eliminates the interference of imaging field of view size or discrete grid resolution on the absolute order of loss, ensuring that the penalty term objectively reflects the absolute energy sparsity of the entire reconstructed scene. Based on this, fixed weight hyperparameters... It can stably generalize to imaging scenes of different sizes, thereby stably driving the network to adaptively activate the soft thresholds of each layer, strongly suppressing widely distributed background artifacts and clutter energy, and making the target scattering point highly focused.

[0110] Step 5: Obtain the echo features of the real scene, input them into the trained and converged network model, and output a 3D millimeter-wave reconstructed image. The specific process is as follows:

[0111] First, we acquire the measured echoes from the current real-world scene. Unlike the training phase, which relies on high-precision data, real-world applications do not require complete, unquantized, high-precision echoes. Within a short pre-sampling time window, the hardware front-end sets a fixed comparison threshold. Threshold detection is performed on the real and imaginary parts of the simulated echo in the current scene to obtain a lightweight single-bit presampled sequence.

[0112] The threshold occupancy rate in the presampled sequence is calculated, and the scale mapping relationship established offline during the training phase is invoked to directly estimate the approximate real and imaginary amplitude scales of the current echo. and Normalized feature vectors are constructed using this coarse scale. .

[0113] Will The input is fed into the trained and weighted MLP prediction subnetwork, and the output of a single inference is the optimal jitter coefficient that matches the current scene. Radar hardware front-end based on A physical jitter signal is generated using the estimated coarse scale, injected into the analog echo channel, and then subjected to purely physical zero-threshold hard symbol truncation via a single-bit analog-to-digital converter to obtain formal single-bit imaging data. .

[0114] Finally, for Distance compression is performed and observation slices are extracted and input into a depth unrolling reconstruction subnetwork with solidified weights to quickly output a high-resolution 3D millimeter-wave reconstructed image.

[0115] Through the above process, the training phase can make full use of high-precision echo data to establish a stable jitter parameter prediction and reconstruction network model; in the real application phase, it only relies on the fixed threshold single-bit pre-sampling results to estimate the coarse scale of the echo, and accordingly completes adaptive jitter modulation, single-bit acquisition and depth unfolding imaging, avoiding dependence on complete unquantized high-precision echo statistics.

[0116] Example 2:

[0117] A depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method includes the following steps:

[0118] 1. Obtain the imaging experiment dataset, specifically:

[0119] Setting the core hardware parameters of a single-station grating scan FMCW radar system: Radar start frequency FM slope ADC sampling rate Number of fast-time sampling points for a single pulse The radar antenna is in Perform a 2D scan on a plane, with the grid size set to [value]. Horizontal scanning interval Vertical scanning interval Set the reference depth for the target area. .

[0120] Under the aforementioned physical parameters, a complex target scene consisting of 5 to 20 point scatterers is randomly generated, and a dynamic range of [value missing] is injected into the original complex echo. After random noise is normalized using root mean square, a total of 1000 pairs of "feature-echo-reference image" datasets are generated for subsequent processing.

[0121] 2: Network model training settings, specifically:

[0122] The end-to-end deep unfolding network proposed in this invention is constructed based on the PyTorch framework. The hidden layer dimension of the adaptive jitter prediction network (MLP) is set to 16; the number of cascaded layers of the deep unfolding reconstruction subnetwork is also set. The Gaussian prior kernel size is set to The Adam optimizer was used during network training, and the initial learning rate of the model was set to [value missing]. The joint loss function includes a mean squared error shape fidelity loss based on the extreme value-normalized image, and an amplitude sparsity penalty loss based on the global mean, where the weight hyperparameter of the sparsity penalty term is fixed. Using the aforementioned dataset of 1000 samples, with a batch size equivalent to 4, end-to-end joint training for 50 epochs was performed in a GPU environment. During training, to prevent extreme value hijacking, gradient isolation was enforced on the division denominator; simultaneously, a pass-through estimator (steepness parameter set to 3.0) was used to implement single-bit truncated backpropagation of gradients.

[0123] 3: Imaging effect comparison and verification, specifically:

[0124] To verify the effectiveness of the method of this invention in balancing harmonics and noise and improving imaging quality, a comparison was made with existing traditional imaging methods based on the publicly available 3-DRIED dataset for the same complex target scene: Range Migration Algorithm (RMA) using full-precision data, RMA using zero-threshold single-bit data, RMA using zero-threshold single-bit data with conventionally injected jitter parameters (RMA(DT)), a method using zero-threshold single-bit data with conventionally injected jitter parameters and based on soft-threshold sparse iterative reconstruction (RMA-ISTA(DT)), a method using zero-threshold single-bit data with conventionally injected jitter parameters and based on Gaussian kernel convolution denoising sparse iterative reconstruction (RMA-ISTA-GUA(DT)), and the method of this invention (AdaD-Net).

[0125] The quantitative evaluation results of each method are shown in Table 1:

[0126] Table 1 Quantitative Assessment Results

[0127]

[0128] In summary, when imaging the echoes of real, complex targets, the method presented in this paper can present clearer target features, improve image contrast, and optimize runtime.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the invention in any way. Any simple modifications, alterations, and equivalent changes made to the above embodiments based on the inventive essence shall still fall within the protection scope of the present invention.

Claims

1. A depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method, characterized in that, Includes the following steps: S1. Construct a pairwise training dataset based on root mean square normalization and extract one-dimensional statistical features; S2. The adaptive jitter parameter prediction subnetwork constructed by inputting one-dimensional statistical features generates adaptive jitter amplitude coefficients, and performs differentiable single-bit quantization with a pass-through estimator mechanism in the differentiable quantization layer to obtain complex single-bit quantized echo data. S3. A deep unfolding network is constructed by combining Gaussian spatial energy prior and adaptive soft thresholding. Complex single-bit quantized echo data is input into the deep unfolding network and the reconstructed image is output. S4. Based on the gradient isolation mechanism and joint loss function, the overall network, including the adaptive jitter parameter prediction sub-network, differentiable layer and deep unfolding network, is fine-tuned and jointly trained using paired training datasets. S5. Obtain the echo features of the real scene, input them into the trained and converged network model, and output the 3D millimeter wave reconstructed image.

2. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, S1 includes: The radar antenna is controlled to scan and sample on a planar rectangular grid. The antenna receives the original echo and obtains the three-dimensional complex echo tensor. The global root mean square value of the three-dimensional complex echo tensor is calculated and normalized to obtain the normalized complex echo. The standard deviations of the real and imaginary parts of the normalized complex echo in the global space are calculated respectively, and the obtained real and imaginary standard deviations are concatenated to form a one-dimensional statistical feature. Construct a pairwise training dataset containing one-dimensional statistical features, corresponding normalized complex echoes, and corresponding reference images.

3. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 2, characterized in that, S1 also includes: A single-bit presampled normalized complex echo was performed using a fixed threshold. The threshold occupancy rate was statistically analyzed, and an offline mapping relationship between the threshold occupancy rate and the echo amplitude scale was established.

4. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, In S2, the adaptive jitter parameter prediction subnetwork is composed of a multi-layer sensing mechanism, which includes a linear mapping layer, a nonlinear activation layer, and an output mapping layer in sequence; the adaptive jitter amplitude coefficient corresponds to the current echo amplitude distribution state.

5. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 4, characterized in that, S2 include: Multiply the uniform random matrix by the adaptive jitter amplitude coefficient and then multiply by the real part standard deviation and the imaginary part standard deviation respectively to construct the real part random jitter matrix and the imaginary part random jitter matrix respectively. The real part random jitter matrix and the imaginary part random jitter matrix are injected into the real part channel and the imaginary part channel of the normalized complex echo data, respectively, to obtain the complex signal after jitter injection; The real and imaginary parts of the complex signal after jitter are quantized separately using differentiable single-bit quantization. The differentiable single-bit quantization outputs of the real and imaginary parts are combined to obtain complex single-bit quantized echo data.

6. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, In S2, differentiable single-bit quantization processing includes: Forward propagation uses a hard sign function to truncate the output, while backward propagation uses a hyperbolic tangent smoothing function as a gradient substitute. The output of differentiable single-bit quantization is the sum of the difference between the hard quantization result and the soft quantization result after gradient truncation and the soft quantization result.

7. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, S3 includes: Range compression processing is performed on complex single-bit quantized echo data to extract two-dimensional observation slices at the reference depth; An imaging operator based on an electromagnetic wave propagation model performs an initial inversion on a two-dimensional observation slice to obtain an initial scattering coefficient estimate. The initial scattering coefficient estimate is input into a depth-unfolded network containing K-level cascade nodes for iterative reconstruction. After K-level node iteration, the reconstructed image is output.

8. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 7, characterized in that, In S3, each node at each level of the deep-unfolded network performs the following operations: Based on the forward observation operator and the imaging operator, data consistency is updated with a learnable step size to obtain the intermediate state image. The amplitude of the intermediate state image is spatially convolved with a Gaussian kernel with a learnable variance parameter to obtain a local spatial energy density prior map. A learnable baseline threshold is introduced, and a dynamic inverse proportional threshold mapping map is obtained based on the local spatial energy density prior map. The intermediate state image is then truncated using a complex-domain nonlinear soft thresholding method with the dynamic inverse proportional threshold mapping map to obtain the reconstructed image of the current node.

9. The depth-unfold-driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, S4 includes: The mean square error of the amplitudes of the reconstructed image and the reference image after normalization to the maximum value is used as the shape fidelity loss, where the normalization denominator is subjected to gradient isolation processing; the global mean of the amplitude of the reconstructed image is used as the amplitude sparsity penalty loss; a total loss function is constructed based on the shape fidelity loss and the amplitude sparsity penalty loss, and the overall network is jointly optimized.

10. The depth-unfolded driven adaptive jitter modulation single-bit millimeter-wave imaging method according to claim 1, characterized in that, S5 include: The echo of the target under test is presampled with a fixed threshold single bit, the threshold occupancy rate is calculated, and a rough amplitude scale estimate is obtained through offline mapping relationship; Normalized scale features are constructed using coarse amplitude scale estimation and input into a trained adaptive jitter parameter prediction subnetwork to output the optimal jitter coefficients. Physical jitter signals are generated based on the optimal jitter coefficients and injected into the echo channel. Formal single-bit imaging data is obtained through a single-bit analog-to-digital converter. After distance compression processing, the single-bit imaging data is input into a trained depth unfolding network to output a three-dimensional millimeter-wave reconstructed image.

Images