A target cognition guiding-based super-sparse array millimeter wave three-dimensional imaging method

Abstract

The application discloses a kind of target cognitive guidance-based super-sparse array millimeter wave three-dimensional imaging method (Tcoge), it is by super-sparse array millimeter wave imaging system to obtain target echo data, distance focusing echo cube data is obtained after signal pre-processing, then by baseline imaging method to obtain rough initial imaging result, utilize lightweight perception network to extract target geometry and intensity perception attribute, finally, the perception attribute is fused with physical constraint, and the final super-sparse array millimeter wave three-dimensional imaging result is obtained by alternating direction multiplier method (ADMM) iterative optimization.Compared with prior art super-sparse array millimeter wave imaging method, the present application can effectively suppress grating lobe and clutter interference, better preserve target geometric structure and scattering energy, and has high imaging robustness under different sparsity, noise conditions and target scenarios, with significantly improved imaging quality.

Description

Technical Field

[0001] This invention belongs to the field of millimeter-wave imaging technology, and specifically relates to ultra-sparse array millimeter-wave three-dimensional imaging technology, which is applicable to millimeter-wave imaging application scenarios such as security screening, industrial inspection, target radar cross section (RCS) diagnosis, and intelligent sensing. Background Technology

[0002] Millimeter-wave three-dimensional imaging technology relies on the azimuth and elevation two-dimensional aperture expansion characteristics of two-dimensional planar arrays, combined with the short wavelength advantage of millimeter waves, to obtain the fine geometric scattering characteristics and electromagnetic intensity information of targets, realize high-resolution three-dimensional target characterization, and has all-weather and all-day working capability. It has become one of the important technologies in the fields of security, industry, and radar detection. For details, see the literature “Sheen DM, McMakin DL, Hall T E. Three-dimensional millimeter-wave imaging for concealed weapon detection. IEEE Transactions on Microwave Theory and Techniques, 2001, 49(9):1581-1592” and “Zhuge X, Yarovoy A G. Asparse aperture MIMO-SAR based UWB imaging system for concealed weapon detection. IEEE Transactions on Geoscience and Remote Sensing, 2011, 49(1):509-518”.

[0003] Ideal millimeter-wave 3D imaging needs to meet the Nyquist sampling criterion and rely on high-density arrays to complete spatial sampling. However, its hardware cost is high and the system complexity is large, making it difficult to meet the requirements of efficient engineering deployment. In practical applications, ultra-sparse array configurations are often used, which will cause serious spatial undersampling problems and form an inherently ill-posed inverse scattering problem. This leads to three core bottlenecks in imaging: first, the reduction of sampling points reduces the coherence accumulation gain and attenuates the effective scattering response; second, the sparse aperture introduces strong grating lobe interference, causing the target geometry to break and artifacts to increase; and third, clutter interference is aggravated, weak scatterers are submerged, and it is difficult to distinguish the target from the background. For details on sparse aperture array millimeter-wave / ultra-wideband imaging systems and their undersampling imaging problems, please refer to the literature "Zhuge X, Yarovoy A G. A sparse aperture MIMO-SAR based UWB imagingsystem for concealed weapon detection. IEEE Transactions on Geoscience and Remote Sensing, 2011, 49(1):509-518".

[0004] To address the problem of image quality degradation in ultra-sparse arrays, existing methods are mainly divided into two categories. One type is model-driven methods based on sparse regularization, such as introducing regularization terms like L1 norm and weighted L1 norm, which suppress interference by encoding general sparse priors. See the reference "Candès EJ, Wakin MB, Boyd S P. Enhancing sparsity by reweighted L1 minimization. Journal of Fourier Analysis and Applications, 2008, 14(5-6):877-905". However, these methods lack high-level visual cognition and are prone to raster lobe artifacts or loss of target structure information due to over-smoothing in ultra-sparse scenes, especially with poor detail preservation for weakly scattering targets. Another type is deep learning-driven imaging methods, such as deep unfolded networks and learnable priors, which utilize data-driven representation capabilities to extract target features. Related unfolded sparse coding learning methods can be found in the reference "Gregor K, LeCun Y. Learning fast approximations of sparse coding. Proceedings of the 27th International Conference on Machine Learning". 2010:399-406”; however, there are problems such as unreliable implicit inference of joint strength-geometric features, insufficient physical constraints, and no mutual correction of the accumulation of perception and cognition errors. It is easy to generate artifacts that violate the physical laws of millimeter wave propagation, and the generalization ability drops significantly under extremely sparse configurations.

[0005] Existing methods cannot simultaneously meet the three core requirements of interference suppression, structure fidelity, and physical effectiveness in ultra-sparse arrays. Therefore, there is an urgent need for a method that can integrate target perception features and physical imaging laws to achieve reliable millimeter-wave 3D imaging. To meet this need, this invention proposes an ultra-sparse array millimeter-wave 3D imaging method based on target cognition guidance. By decoupling target geometry and intensity perception and combining physical cognition constraints, high-quality millimeter-wave 3D imaging under ultra-sparse configuration is achieved. Summary of the Invention

[0006] This invention proposes a target cognition-guided ultra-sparse array millimeter-wave 3D imaging method (Tcoge). The core of this method is to construct an imaging framework that integrates target perception and physical cognition. The hierarchical perception mechanism of human vision is abstracted into a computational model. First, the geometric mask and scattering weight map of the target are extracted as structured priors through the target perception module. Then, three physical constraints, namely measurement consistency, single scatterer sparsity, and coherence of adjacent scatterers, are introduced through the physical cognition module. The perception prior is embedded as a spatially variable prior into a unified ADMM to carry out an iterative solution process, thereby realizing closed-loop imaging that combines perception-guided optimization and physical verification of perception.

[0007] This method utilizes an ultra-sparse array millimeter-wave imaging system to acquire target echo data. After signal preprocessing, range-focused echo cube data is obtained. A coarse initial imaging result is then obtained using a baseline imaging method. A lightweight sensing network is used to extract the target's geometric and intensity sensing attributes. Finally, the sensing attributes are fused with physical constraints, and the final ultra-sparse array millimeter-wave 3D imaging result is obtained through iterative optimization using the Alternating Direction Multiplier Method (ADMM). Compared with existing ultra-sparse array millimeter-wave imaging methods, this invention effectively suppresses grating lobes and clutter interference, better preserves the target geometry and scattering energy, and exhibits high imaging robustness under different sparsity, noise conditions, and target scenes, resulting in a significant improvement in imaging quality.

[0008] To facilitate the description of this invention, the following terms are defined first: Ultra-sparse array millimeter-wave imaging system receiving target echoes

[0009] Definition 1. Method for receiving target echoes in an ultra-sparse array millimeter-wave (mmWave) imaging system.

[0010] An ultra-sparse array millimeter-wave imaging system refers to a physical imaging system that acquires millimeter-wave echo data of a target using a two-dimensional planar ultra-sparse array, under conditions where the array antenna spatial sampling rate is much lower than the Nyquist sampling criterion, and then inverts the three-dimensional spatial scattering distribution of the target. This system uses azimuth and elevation array elements to form a two-dimensional array aperture, transmitting a broadband linear frequency modulated millimeter-wave signal to the target to be imaged. The echo scattered by the target is determined by the array sampling coordinates. The finite array elements at the location receive the data and sample it along the range direction to obtain the raw echo data. ,in , and These represent the azimuth, elevation, and range sampling coordinates, respectively. The method for receiving target echoes in an ultra-sparse array millimeter-wave (mmWave) imaging system uses the target's spatial scattering intensity as the variable to be reconstructed. The forward imaging operator is constructed using the ultra-sparse array sampling position, millimeter-wave propagation phase, range delay, and array aperture response to describe the mapping relationship between the target scattering distribution and the array's received echoes. See the reference "Sheen DM, McMakin DL, Hall T E. Three-dimensional millimeter-wave imaging for concealed weapon detection. IEEE Transactions on Microwave Theory and Techniques, 2001, 49(9):1581-1592".

[0011] Definition 2. Wideband Linear Frequency Modulation (LFM) Signal

[0012] Wideband linear frequency modulation (LFM) signals refer to wideband millimeter-wave pulse signals whose instantaneous frequency changes linearly with time during the pulse duration. The frequency is continuously swept from the initial frequency to the termination frequency according to a fixed frequency modulation slope, and the signal bandwidth is determined by both the frequency modulation slope and the pulse duration. See the reference "Bao Zheng, Xing Mengdao, Wang Tong. Radar Imaging Technology [M]. Electronic Industry Press, 2005".

[0013] Definition 3. Azimuth Array Elements of Ultra-Sparse Arrays

[0014] Ultra-sparse array azimuth elements refer to the effective array elements in a two-dimensional planar millimeter-wave array arranged along the azimuth direction for transmitting or receiving target echo signals; their number is denoted as . The position of its array element corresponds to the array sampling coordinates Azimuth coordinates In this invention, azimuth array elements are used to acquire spatial sampling information of the target echo signal in the azimuth aperture. Since the number of azimuth array elements in an ultra-sparse array is significantly less than that in a full-sampling array, azimuth spatial sampling is insufficient, which easily leads to grating lobe interference, sidelobe enhancement, and target structure artifacts during direct imaging.

[0015] Definition 4. Ultrasparse array elevation vector element

[0016] Ultra-sparse array elevation-direction array elements refer to the effective array elements deployed along the elevation direction in a two-dimensional planar millimeter-wave array that participate in echo acquisition; their number is denoted as... Corresponding array sampling coordinates elevation coordinates in In this invention, elevation asterism elements are used to acquire spatial sampling information of the target echo in the elevation aperture, and together with azimuth asterism elements, they constitute a two-dimensional ultra-sparse array aperture. Since elevation sampling is also lower than that required for dense array sampling, the imaging results are prone to elevation artifacts, loss of weak scattering, and geometric discontinuities.

[0017] Definition 5, Scattering Intensity

[0018] Scattering intensity refers to the spatial position of the target The strength of the backscattered echo generated by the scattering body for the incident millimeter-wave signal is denoted as . In this invention, scattering intensity is used to characterize the reflection amplitude or scattering energy of each scattering unit in the three-dimensional space of the target, and is the physical meaning of the imaging result to be reconstructed. The greater the scattering intensity, the stronger the contribution of that location to the array's received echo; the smaller the scattering intensity, the more likely that the location is a weak scattering region, a background region, or a clutter region.

[0019] Definition 6. Traditional target-aware network extraction methods

[0020] Traditional target-aware network extraction methods are based on lightweight U-Net encoder-decoder networks, which use coarse initial imaging results. As input, output the target confidence map. The geometric mask and scattering weight map are obtained from the target confidence map, achieving decoupled extraction of target geometric attributes and scattering intensity attributes. Specifically, the target confidence map... Thresholding is performed to obtain a geometric mask that characterizes the spatial layout and connectivity of the target. Target confidence map Perform monotonic mapping The scattering weight map used to modulate the sparsity penalty is obtained. The parameters of the target perception network can be pre-trained using millimeter-wave coarse imaging samples of labeled target areas, or obtained through joint training with simulated imaging data and measured imaging data. For details on the U-Net network structure, please refer to the literature "Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical ImageComputing and Computer-Assisted Intervention, 2015:234-241".

[0021] Definition 7, Distance-Focused Echo Cube

[0022] Range-focused echo cube refers to millimeter-wave raw echo data that has undergone range compression, range-axis phase compensation, and range migration correction, and is then analyzed according to the coordinates of the azimuth array elements. Elevation matrix element coordinates and distance sampling coordinates The three-dimensional complex-valued data cube formed by the arrangement is denoted as ,in The range-focused echo cube is composed of multiple two-dimensional range slices arranged sequentially along the range direction, with fixed range sampling coordinates. Subsequently, the first [item] obtained from the distance-focused echo cube Each range slice is a two-dimensional complex-valued echo in the azimuth-elevation direction corresponding to the range cell, which will be abbreviated as in subsequent single range slice processing. It is used for matched filtering coarse imaging and measurement consistency constraint calculation.

[0023] Definition 8. Geometric Mask

[0024] A binary matrix (with elements of 0 or 1) characterizing the spatial layout and connectivity of the target, where 0 represents the background region and 1 represents the target scattering region, provides target spatial anchor points for imaging optimization, and is extracted from the coarse initial image by the target perception network.

[0025] Definition 9, Scattering Weight Map

[0026] The scattering weight map is a target confidence map output by the target perception network. The non-negative real-valued matrix representing the distribution of target scattering energy relative to background interference, obtained by the monotonically decreasing mapping, is denoted as... In this invention, the scattering weight map is used to spatially adaptively modulate the sparsity penalty of the imaging results, and its specific expression is as follows:

[0027]

[0028] in, Represents pixels Confidence level that it belongs to the target region It is a small, numerically stable constant. For locations with high target confidence, [this is important]. The smaller value results in a weaker sparsity penalty at that location, thus preserving the effective scattering response; for locations with lower target confidence, The larger value results in a stronger sparsity penalty at that location, thereby suppressing background clutter and raster lobe artifacts.

[0029] Definition 10. Measurement Consistency Constraints

[0030] The physical constraint term based on the millimeter-wave forward propagation model is represented by the Frobenius norm deviation between the imaging results and the measured echo data. This penalizes the deviation between the imaging results and the physical measurements, ensuring that the imaging results conform to the millimeter-wave wavefront propagation law.

[0031] Definition 11. Sparsity constraint of a single scattering volume

[0032] Based on prior constraints of millimeter-wave scattering physics, and imaging results of Norm representation utilizes the characteristic that "only a few scatterers of the target produce significant echoes" to suppress background clutter and preserve strong target scattering responses. For details on the ideas of sparse signal recovery and weighted sparse regularization, please refer to the literature "Candès EJ, Wakin MB, Boyd S P. Enhancing sparsity by reweighted L1 minimization. Journal of Fourier Analysis and Applications, 2008, 14(5-6):877-905".

[0033] Definition 12. Coherence Constraints of Adjacent Scatterers

[0034] Based on the prior constraints of the extended target geometry, the spatial dependence of the scatterer is modeled by Markov random field (MRF), penalizing the state abrupt changes of adjacent pixels, ensuring the spatial connectivity of the target scattering region and avoiding geometric structure fragmentation. For details, see the literature "Geman S, Geman D. Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1984, PAMI-6(6):721-741".

[0035] Definition 13. Traditional Alternating Direction Multiplier Method (ADMM)

[0036] This method decomposes multivariate joint optimization problems into multiple univariate convex optimization subproblems. By introducing auxiliary and dual variables, the constrained optimization problem is transformed into an unconstrained problem, achieving efficient iterative solution. For details, see the literature "Boyd S, Parikh N, Chu E, Peleato B, Eckstein J. Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers. Foundations and Trends in Machine Learning, 2011, 3(1):1-122".

[0037] Definition 14. Traditional Matched Filtering Method

[0038] Traditional matched filtering methods are baseline imaging techniques used in millimeter-wave imaging to generate coarse initial imaging results. For single range slice echoes... This method constructs a propagation path imaging matching phase factor based on the propagation distance from candidate imaging grid points to each array element position, performs phase compensation and coherent accumulation on the array received echo, thereby enhancing the scattering response of the corresponding grid points and obtaining the coarse initial imaging result of that range slice. Matched filtering is applied sequentially to all range slices, and the imaging results of each slice are stacked along the range direction to obtain a coarse initial 3D imaging result of the target scene. For details, please refer to the literature "Bao Zheng, Xing Mengdao, Wang Tong. Radar Imaging Technology [M]. Electronic Industry Press, 2005".

[0039] Definition 15. Traditional Distance Migration Correction (RCMC) Method

[0040] The traditional range migration correction (RCMC) method refers to correcting the range-compressed echo to eliminate the migration error of the scatterer at different range gates, so that the echo can be accurately returned to the corresponding range cell. It is a key preprocessing step for forming range-focused echo cubes. For details, see the literature "Bao Zheng, Xing Mengdao, Wang Tong. Radar Imaging Technology [M]. Electronic Industry Press, 2005".

[0041] Definition 16. Standard Linear Frequency Modulation Signal Pulse Compression Method

[0042] The standard linear frequency modulation (LFM) signal pulse compression method refers to using a filter at the receiving end that matches the transmitted broadband LFM signal to compress the original echo. A signal processing method for range compression is presented. This method focuses scattered echoes with the same range delay onto the corresponding range cell, compressing long pulse echoes into narrow pulse responses, thereby improving range resolution. In this invention, standard linear frequency modulated (LFM) signal pulse compression is used to obtain the range-compressed echo signal, providing a basis for subsequent phase compensation, range migration correction, and range-focused echo cube generation. For details on LFM pulse compression, please refer to "Skolnik MI. Radar Handbook [M]. 3rd ed. New York: McGraw-Hill, 2008".

[0043] Definition 17, Fast Time

[0044] In radar imaging, fast time is used to characterize the rapid sampling time dimension of the echo range direction, in order to distinguish target scattered echoes of different range units.

[0045] Definition 18. Traditional Preprocessing Methods

[0046] Traditional preprocessing methods refer to performing preprocessing such as range compression and range migration correction on the raw radar echoes to normalize the data into a three-dimensional range-focused echo cube.

[0047] Definition 19. Traditional Stacking Operation Method

[0048] The traditional stacking operation method refers to arranging the two-dimensional azimuth-elevation imaging results corresponding to each distance slice along the distance direction according to the distance slice index to form a three-dimensional imaging result of the target scene.

[0049] Definition 20, Frobenius Norm

[0050] The Frobenius norm is a matrix norm used to measure the global numerical deviation of a complex-valued matrix. It is defined as the square root of the sum of the squares of the moduli of all elements of the matrix. In millimeter-wave ultra-sparse array imaging optimization models, its square form is often used as a measurement consistency constraint term to quantify the fitting error between the measured data and the output data of the imaging forward propagation model. It is a core metric for evaluating the physical fidelity of the imaging results.

[0051] Definition 21, Augmented Lagrangian function

[0052] The augmented Lagrangian function is the core mathematical model for solving the constrained optimization problem of millimeter-wave ultra-sparse array imaging. It is constructed by introducing a quadratic penalty term and dual variables based on the standard Lagrangian function. It is used to transform multivariate imaging optimization problems with equality constraints into unconstrained optimization problems, including auxiliary variables. , and binary state field The constrained optimization problem is transformed into an unconstrained optimization problem that can be solved iteratively and alternately, which is the key solution carrier for the target cognition-guided imaging algorithm. For details, please refer to the literature "Boyd S, Parikh N, Chu E, Peleato B, Eckstein J. Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers. Foundations and Trends in Machine Learning, 2011, 3(1):1-122".

[0053] Definition 22, Classical Range-Directional Phase Compensation Factor

[0054] Range-phase compensation factor refers to the factor applied to compensate for the phase compression of the original echo after linear frequency modulation (LFM) pulse compression. The complex exponential phase factor, constructed from the round-trip propagation phase corresponding to each distance slice, is denoted as... ,in, This represents the natural exponential function. The imaginary unit, Pi The carrier frequency for transmitting broadband linear frequency modulated signals, The speed at which light travels through the air. For wave number, For distance-oriented sampling coordinates, For the first The center distance corresponding to each distance slice; this factor is used to compensate for the distance in the echo after range compression. Corresponding round-trip propagation phase term This unifies the phase reference for different range slices, providing a data basis for phase alignment for subsequent range migration correction and range focusing echo cube generation. For details, please refer to the literature "Gao Jingkun, Deng Bin, Qin Yuliang, et al. Near-field three-dimensional imaging technology of scanning MIMO array [J]. Journal of Radar, 2018, 7(6): 676-684".

[0055] Definition 23, Forward Operator for Millimeter-Wave Imaging

[0056] In single-range slice imaging, the millimeter-wave imaging forward operator is used to describe the propagation mapping relationship between the imaging result to be optimized and the echo data received by the two-dimensional ultra-sparse array. Let the... The imaging results corresponding to each distance slice are The effective element coordinates in a two-dimensional ultrasparse array are:

[0057]

[0058] in , Let the first The coordinates of the imaging grid points in each distance slice are:

[0059]

[0060] in The propagation distance between the array element and the imaging grid points is:

[0061]

[0062] Then the first Forward imaging operator corresponding to each distance slice Defined as:

[0063]

[0064] in, Indicates the first Imaging grid points in a distance slice Scattering intensity at that location, Let the wave number be denoted by . For a binary state field . Modulated imaging results The forward propagation result is as follows:

[0065]

[0066] To simplify the description, in the subsequent single-distance slice optimization process, , , , Abbreviated as , , , .

[0067] This invention provides a target-aware guided ultra-sparse array millimeter-wave three-dimensional imaging method, comprising the following steps:

[0068] Step 1: Initialize relevant parameters and acquire echo data:

[0069] For 3D scenes requiring imaging and modeling, the ultra-sparse array millimeter-wave imaging system described in Definition 1 is used to receive target echoes. A broadband linear frequency modulated signal as described in Definition 2 is transmitted, and the following core parameters are initialized: the speed of light in air... Pi carrier frequency , wave number Definition 3: Number of azimuth elements in a supersparse array Definition 4: Number of elements in the elevation vector array of ultra-sparse arrays Array sampling point coordinates ,in These represent the sampling coordinates of the azimuth aster and the elevation aster, respectively, in the imaging space. Distance sampling number Distance sampling coordinates Regularization weights (Single scattering volume sparsity constraint parameters in Definition 11). (Neighboring scatterer coherence constraint parameters in Definition 12), ADMM method penalty parameters in Definition 13 Numerical stability small constant 4-neighbor operator Iteration termination threshold .

[0070] For a 3D scene target, the method for receiving target echoes using the ultra-sparse array millimeter-wave imaging system described in Definition 1 is employed to obtain the raw echo data. ,in For fast time as defined in 17.

[0071] Step 2: Echo signal preprocessing to obtain range-focused echo cubes:

[0072] The raw echo data obtained in step 1 is processed using the traditional preprocessing method described in Definition 18 to obtain the range-focused echo cube in Definition 7. The specific steps are as follows:

[0073] Step 2.1: Process the raw echo obtained in Step 1. The distance compression is performed using the standard linear frequency modulation signal pulse compression method in Definition 16 to obtain the distance-compressed echo signal.

[0074] Step 2.2: Multiply the range-compressed echo signal obtained in Step 2.1 by the classic range-direction phase compensation factor in Definition 22 to compensate for the round-trip propagation phase corresponding to each range slice, and obtain the phase-compensated echo signal.

[0075] Step 2.3: The echo signal obtained in Step 2.2 is corrected using the conventional range migration correction (RCMC) method in Definition 15 to obtain the corrected echo signal;

[0076] The corrected echoes are arranged according to the azimuth element coordinates, elevation element coordinates, and range sampling coordinates to obtain the range-focused echo cube as defined in Definition 7. And the two-dimensional range slices arranged along the range direction in the echo cube, i.e., the two-dimensional complex-valued echoes in the azimuth-elevation direction. .

[0077] Step 3: Generate coarse initial imaging results:

[0078] The two-dimensional complex-valued echo in azimuth and elevation directions obtained in step 2.3 The traditional matched filtering method in Definition 14 is used to process the image to obtain a coarse initial imaging result. .

[0079] Step 4: The target perception module extracts structured perception priors:

[0080] The coarse initial imaging results obtained in step 3 The geometric mask described in Definition 8 is obtained by processing the data using the traditional target perception network extraction method described in Definition 6. and the scattering weight map described in Definition 9 .

[0081] make ,in, express In position Elements at the location; geometric mask The element at the corresponding position in As The initial value, i.e. Written in matrix form as .

[0082] Step 5: Construct the objective function for target cognition-guided imaging optimization:

[0083] Based on the geometric mask obtained in step 4 and scattering weight map Construct a constrained imaging optimization objective function:

[0084]

[0085] in, Indicated by and For the variable to be optimized, minimize the sum of its subsequent terms; The imaging results are to be optimized. , express In position Elements at the location; The initial value is the geometric mask in definition 8 obtained in step 4. ,Right now ; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; For the millimeter-wave imaging forward operator in definition 23; This is the scattering weight map in Definition 9 obtained in step 4; This represents the Hadamard product, which is the multiplication of elements at corresponding positions. This represents the Frobenius norm in Definition 20, with the superscript 2 indicating that the norm is squared. This represents the sum of the moduli of all elements in the matrix; and All are non-negative parameters. To define the single-scatterer sparsity constraint parameters in definition 11, Define the coherence constraint parameters for adjacent scatterers in definition 12; This indicates the coherence constraint between adjacent scatterers in Definition 12, with the subscript ac serving as the symbol for this constraint.

[0086]

[0087] in, and for The position index of the element in the middle; Indicates position The four neighboring regions, including the upper neighbor position. , next to Left neighbor position and right neighbor position ; For the four neighboring areas The position index in the formula; the first and second summations in the formula represent summations over 𝑖 and 𝑗 respectively, and the third summation represents summations over the four neighboring positions (𝑝, 𝑗) of position (𝑖, 𝑗); Indicates position With position Interaction weights between them; and They represent In position and location Elements at the location; It represents the square of the difference between two elements.

[0088] Let auxiliary variables Then the objective function in step 5 is expressed as:

[0089]

[0090] Wherein, the minimization symbol at the beginning of the expression represents... , , and For the variable to be optimized, minimize the sum of its subsequent terms; This indicates that the following equation represents the constraints that each variable to be optimized should satisfy.

[0091] According to the definition of the augmented Lagrangian function in Definition 21, the augmented Lagrangian function corresponding to the above objective function is:

[0092]

[0093] in, The augmented Lagrangian function as defined in 21; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; The millimeter-wave imaging forward operator as defined in definition 23; and Auxiliary variables introduced for this step; , , , and The meaning is the same as that in the objective function mentioned above in step 5; and The penalty parameters initialized in step 1; and Constraints and The corresponding dual variables have the following initial values: , .

[0094] Step 6: Solve using ADMM-based expansion iteration:

[0095] Using the traditional alternating direction multiplier method described in Definition 13, the augmented Lagrangian function obtained in step 5 is... Perform iterative solutions and update sequentially. , , , , and The specific steps are as follows:

[0096] Step 6.1, fix the first The next iteration , and Update the auxiliary variable as follows , obtained the The next iteration :

[0097]

[0098] in, For the millimeter-wave imaging forward operator described in definition 23 The adjoint operator; To and Identity matrices of the same dimension; superscript Indicates matrix inversion; superscript and Indicates the number of iterations for the corresponding variable; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; The penalty parameters initialized in step 1.

[0099] Step 6.2, fix the first The next iteration and Update the auxiliary variable as follows , obtained the The next iteration :

[0100]

[0101] in, This represents the element-wise soft thresholding function, which applies to any input element. ,when Time output ,when Output 0 at time. Soft threshold, , and This has the same meaning as in step 5.

[0102] Step 6.3, Fix , , , and Update the imaging results to be optimized using the following formula. In position The element at that position, obtain :

[0103]

[0104] in, and The position index of the element. , , , , , and They represent , , , , , and In position The element at that location, and The penalty parameters initialized in step 1, The numerically stable small constants are initialized for step 1.

[0105] Step 6.4, Fix , and Update as follows In position The element at that position, obtain :

[0106]

[0107] in, , , and They represent In the upper adjacent position , next to Left neighbor position and right neighbor position The element at that location, , and Same meaning as in step 6.3, and This has the same meaning as in step 5.

[0108] Step 6.5: Based on the iteration results from Steps 6.1 to 6.4, update the dual variable using the following formula. and :

[0109]

[0110]

[0111] in, This represents the Hadamard product, which is the multiplication of elements at corresponding positions.

[0112] Step 6.6: Calculate the difference between the imaging results obtained from two adjacent iterations using the Frobenius norm as defined in Definition 20. Less than the iteration termination threshold initialized in step 1 When, output Otherwise Then return to step 6.1.

[0113] Step 7: Generate 3D millimeter-wave imaging results:

[0114] The optimized imaging results obtained from each distance slice in step 6 The target scene is processed using the traditional stacking operation method described in Definition 19 according to the distance direction to obtain the three-dimensional millimeter-wave sparse imaging result, thus completing the three-dimensional characterization of the target under the ultra-sparse array.

[0115] The innovations and advantages of this invention are reflected in multiple aspects: Firstly, it applies the hierarchical perception mechanism of human vision—from coarse geometric perception to fine intensity verification—to ultra-sparse array millimeter-wave imaging. By decoupling target geometry and intensity perception, it avoids mutual interference when traditional methods jointly extract features, solves the problem of unreliable feature inference under ultra-sparse conditions, and provides structured spatial anchors for imaging optimization. Secondly, it constructs a dual-pathway fusion imaging framework of target perception and physical cognition, embedding perceptual priors such as geometric masks and scattering weight maps as spatially variable priors into the physical model, forming a closed loop of perception-guided optimization and physical verification of perception, effectively solving the perception problems of existing methods. To address the issues of disconnect from physics and insufficient physical constraints, this method ensures that imaging results possess both structural fidelity and physical validity. An ADMM-based expansion iterative solution strategy is designed, decomposing the complex joint optimization problem into multiple sub-problems with closed-form solutions. This improves solution efficiency while maintaining interpretability, avoiding the insufficient interpretability and artifact generation issues of end-to-end deep learning methods, making it more suitable for engineering deployment. Furthermore, the perception module and physical cognition module are decoupled, allowing for flexible replacement of the perception network structure and adjustment of physical constraints. This enables adaptation to different ultra-sparse array topologies and various millimeter-wave imaging scenarios, demonstrating good compatibility and scalability. Attached Figure Description

[0116] Figure 1 A schematic diagram of the flowchart of the ultra-sparse array millimeter-wave three-dimensional imaging method based on target cognition guidance provided by the present invention.

[0117] Figure 2 The table below shows the imaging quality evaluation results of the wrench scene under different sampling rate conditions using the method of the present invention. Detailed Implementation

[0118] This invention is primarily verified using experimental data. All steps and conclusions have been verified as correct using the mathematical calculation software MATLAB 2023b and the deep learning framework PyTorch 2.0. The hardware platform consists of a GeForce RTX 2080 Ti (11GB), an Intel Core i7-8700K processor, and 64GB of RAM. The specific implementation steps are as follows:

[0119] Step 1: Initialize relevant parameters and acquire echo data:

[0120] For 3D scenes requiring imaging and modeling, the ultra-sparse array millimeter-wave imaging system described in Definition 1 is used to receive target echoes. A broadband linear frequency modulated signal as described in Definition 2 is transmitted, and the following core parameters are initialized: the speed of light in air... Pi carrier frequency , wave number Definition 3: Number of azimuth elements in a supersparse array Definition 4: Number of elements in the elevation vector array of ultra-sparse arrays Array sampling point coordinates Imaging space Distance sampling number Distance sampling coordinates Regularization weights (Single scattering volume sparsity constraint parameters in Definition 11). (Neighboring scatterer coherence constraint parameters in Definition 12), ADMM method penalty parameters in Definition 13 Numerical stability small constant 4-neighbor operator Iteration termination threshold .

[0121] For a 3D scene target, the method for receiving target echoes using the ultra-sparse array millimeter-wave imaging system described in Definition 1 is employed to obtain the raw echo data. ,in For fast time as defined in 17.

[0122] Step 2: Echo signal preprocessing to obtain range-focused echo cubes:

[0123] The raw echo data obtained in step 1 is processed using the traditional preprocessing method described in Definition 18 to obtain the range-focused echo cube in Definition 7. The specific steps are as follows:

[0124] Step 2.1: Process the raw echo obtained in Step 1. The distance compression is performed using the standard linear frequency modulation signal pulse compression method in Definition 16 to obtain the distance-compressed echo signal.

[0125] Step 2.2: Multiply the range-compressed echo signal obtained in Step 2.1 by the range-direction phase compensation factor in Definition 22 to compensate for the round-trip propagation phase corresponding to each range slice, and obtain the phase-compensated echo signal.

[0126] Step 2.3: Correct the echo signal obtained in Step 2.2 using the conventional range migration correction (RCMC) method in Definition 15 to obtain the corrected echo signal;

[0127] The corrected echoes are arranged according to the azimuth element coordinates, elevation element coordinates, and range sampling coordinates to obtain the range-focused echo cube as defined in Definition 7. And the two-dimensional range slices arranged along the range direction in the echo cube, i.e., the two-dimensional complex-valued echoes in the azimuth-elevation direction. .

[0128] Step 3: Generate coarse initial imaging results:

[0129] The two-dimensional complex-valued echo obtained in step 2.3 The traditional matched filtering method in Definition 14 is used to process the image to obtain a coarse initial imaging result. .

[0130] Step 4: The target perception module extracts structured perception priors:

[0131] The coarse initial imaging results obtained in step 3 Using the traditional target perception network extraction method described in Definition 6, the geometric mask described in Definition 8 is obtained. and the scattering weight map described in Definition 9 .make ,in, express In position Elements at the location; geometric mask The element at the corresponding position in As The initial value, i.e. Written in matrix form as .

[0132] Step 5: Construct the objective function for target cognition-guided imaging optimization:

[0133] Based on the geometric mask obtained in step 4 and scattering weight map Construct a constrained imaging optimization objective function:

[0134]

[0135] in, Indicated by and For the variable to be optimized, minimize the sum of its subsequent terms; The imaging results are to be optimized. , express In position Elements at the location; The initial value is the geometric mask in definition 8 obtained in step 4. ,Right now ; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; For the millimeter-wave imaging forward operator in definition 23; This is the scattering weight map in Definition 9 obtained in step 4; This represents the Hadamard product, which is the multiplication of elements at corresponding positions. This represents the Frobenius norm in Definition 20, with the superscript 2 indicating that the norm is squared. This represents the sum of the moduli of all elements in the matrix; This indicates the coherence constraint between adjacent scatterers in Definition 12, with the subscript ac serving as the symbol for this constraint.

[0136]

[0137] in, and for The position index of the element in the middle; Indicates position The four neighboring regions, including the upper neighbor position. , next to Left neighbor position and right neighbor position ; For the four neighboring areas The position index in the formula; the first and second summations in the formula represent summations over 𝑖 and 𝑗 respectively, and the third summation represents summations over the four neighboring positions (𝑝, 𝑗) of position (𝑖, 𝑗); Indicates position With position Interaction weights between them; and They represent In position and location Elements at the location; This represents the square of the difference between two elements. Let the auxiliary variable be... Then the objective function in step 5 is expressed as:

[0138]

[0139] Wherein, the minimization symbol at the beginning of the expression represents... , , and For the variable to be optimized, minimize the sum of its subsequent terms; This indicates that the following equation represents the constraints that each variable to be optimized must satisfy. According to Definition 21, the augmented Lagrangian function corresponding to the above objective function is:

[0140]

[0141] in, The augmented Lagrangian function as defined in 21; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; The millimeter-wave imaging forward operator as defined in definition 23; and Auxiliary variables introduced for this step; , , The meaning is the same as that in the objective function mentioned above in step 5; and Constraints and The corresponding dual variables have the following initial values: , .

[0142] Step 6: Solve using ADMM-based expansion iteration:

[0143] Using the alternating direction multiplier method described in Definition 13, the augmented Lagrangian function obtained in step 5 is... Perform iterative solutions and update sequentially. , , , , and The specific steps are as follows:

[0144] Step 6.1, fix the first The next iteration , and Update the auxiliary variable as follows , obtained the The next iteration :

[0145]

[0146] in, For the millimeter-wave imaging forward operator described in definition 23 The adjoint operator; To and Identity matrices of the same dimension; superscript Indicates matrix inversion; superscript and Indicates the number of iterations for the corresponding variable; This is the two-dimensional complex echo in the azimuth-elevation direction obtained in step 2.

[0147] Step 6.2, fix the first The next iteration and Update the auxiliary variable as follows , obtained the The next iteration :

[0148]

[0149] in, This represents the element-wise soft thresholding function, which applies to any input element. ,when Time output ,when Output 0 at time. This has the same meaning as in step 4.

[0150] Step 6.3, Fix , , , and Update the imaging results to be optimized using the following formula. In position The element at that position, obtain :

[0151]

[0152] in, and The position index of the element. , , , , , and They represent , , , , , and In position The element at that location.

[0153] Step 6.4, Fix , and Update as follows In position The element at that position, obtain :

[0154]

[0155] in, , , and They represent In the upper adjacent position , next to Left neighbor position and right neighbor position The element at that location, , and This has the same meaning as in step 6.3.

[0156] Step 6.5: Based on the iteration results from Steps 6.1 to 6.4, update the dual variable using the following formula. and :

[0157]

[0158]

[0159] in, This represents the Hadamard product, which is the multiplication of elements at corresponding positions.

[0160] Step 6.6: Calculate the difference between the imaging results obtained from two adjacent iterations using the Frobenius norm as defined in Definition 20. Less than When, output Otherwise Then return to step 6.1.

[0161] Step 7: Generate 3D millimeter-wave imaging results:

[0162] The optimized imaging results obtained from each distance slice in step 6 The target scene is processed using the traditional stacking operation method described in Definition 19 according to the distance direction to obtain the three-dimensional millimeter-wave sparse imaging result, thus completing the three-dimensional characterization of the target under the ultra-sparse array.

[0163] The method of this invention achieves effective suppression of grating lobes and clutter in the imaging results of the measured target at an ultra-sparse sampling rate of 10%, while preserving the target's geometric structure and scattering energy; compared with existing methods... Compared with regularization methods and deep learning DeepPrior methods, it improves PSNR by 3~5dB and SSIM by 8%~14%. It can still preserve the interpretable global layout of the target under 5% extreme ultra-sparse sampling, and its imaging robustness is significantly better than existing methods.

Claims

1. A method for target cognition-based guided super-sparse array millimeter wave three-dimensional imaging, characterized in that Including the following Several steps: Step 1: Initialize relevant parameters and acquire echo data: For 3D scenes requiring imaging and modeling, an ultra-sparse array millimeter-wave imaging system is used to receive target echoes. A broadband linear frequency modulated signal is emitted, and the following core parameters are initialized: the speed of light in air. Pi carrier frequency , wave number Ultra-sparse array azimuth array element number Ultra-sparse array elevation vector number of elements Array sampling point coordinates ,in These represent the sampling coordinates of the azimuth aster and the elevation aster, respectively, in the imaging space. Distance sampling number Distance sampling coordinates Regularization weights It is a single-scattering volume sparsity constraint parameter. These are the coherence constraint parameters for adjacent scatterers and the penalty parameters for the ADMM method. Numerical stability small constant 4-neighbor operator Iteration termination threshold ; For 3D scene targets, a method using an ultra-sparse array millimeter-wave imaging system to receive target echoes is employed to obtain raw echo data. ,in To save time; Step 2: Echo signal preprocessing to obtain range-focused echo cubes: The raw echo data obtained in step 1 is processed using traditional preprocessing methods to obtain range-focused echo cubes. The specific steps are as follows: Step 2.1: Process the raw echo obtained in Step 1. The standard linear frequency modulation signal pulse compression method is used to compress the distance, and the distance-compressed echo signal is obtained. Step 2.2: Multiply the range-compressed echo signal obtained in Step 2.1 by the classical range phase compensation factor to compensate for the round-trip propagation phase corresponding to each range slice, and obtain the phase-compensated echo signal. Step 2.3: The echo signal obtained in Step 2.2 is corrected using the traditional range migration correction (RCMC) method to obtain the corrected echo signal; The corrected echoes are arranged according to the azimuth element coordinates, elevation element coordinates, and range sampling coordinates to obtain the range-focused echo cube. And the two-dimensional range slices arranged along the range direction in the echo cube, i.e., the two-dimensional complex-valued echoes in the azimuth-elevation direction. ; Step 3: Generate coarse initial imaging results: The two-dimensional complex-valued echo in azimuth and elevation directions obtained in step 2.3 The initial imaging results were obtained by using a traditional matched filtering method. ; Step 4: The target perception module extracts structured perception priors: The coarse initial imaging results obtained in step 3 The geometric mask is obtained by processing the data using traditional target perception network extraction methods. and scattering weight map ; make ,in, express In position Elements at the location; geometric mask The element at the corresponding position in As The initial value, i.e. Written in matrix form as ; Step 5: Construct the objective function for target cognition-guided imaging optimization: Based on the geometric mask obtained in step 4 and scattering weight map Construct a constrained imaging optimization objective function: in, Indicated by and For the variable to be optimized, minimize the sum of its subsequent terms; The imaging results are to be optimized. , express In position Elements at the location; The initial value is the geometric mask obtained in step 4 (8). ,Right now ; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; For millimeter-wave imaging forward operator in 23; The scattering weight map in step 4 is shown in step 9. This represents the Hadamard product, which is the multiplication of elements at corresponding positions. denoted by 20, the Frobenius norm in 20 is indicated by the superscript 2, which indicates that the norm is squared. This represents the sum of the moduli of all elements in the matrix; and All are non-negative parameters. For the single-scattering volume sparsity constraint parameters in 11, The coherence constraint parameters for adjacent scatterers in 12; This indicates the coherence constraint between adjacent scatterers in 12, with the subscript ac serving as the symbol for that constraint; in, and for The position index of the element in the middle; Indicates position The four neighboring regions, including the upper neighbor position. , next to Left neighbor position and right neighbor position ; For the four neighboring areas The position index in the formula; the first and second summations in the formula represent summations over 𝑖 and 𝑗 respectively, and the third summation represents summations over the four neighboring positions (𝑝, 𝑗) of position (𝑖, 𝑗); Indicates position With position Interaction weights between them; and They represent In position and location Elements at the location; Represents the square of the difference between two elements; Let auxiliary variables Then the objective function in step 5 is expressed as: Wherein, the minimization symbol at the beginning of the expression represents... , , and For the variable to be optimized, minimize the sum of its subsequent terms; This indicates that the following equation represents the constraints that each variable to be optimized should satisfy. Based on the augmented Lagrangian function, the augmented Lagrangian function corresponding to the above objective function is: in, The augmented Lagrangian function described in 21; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; For millimeter-wave imaging forward operator; and Auxiliary variables introduced for this step; , , , and The meaning is the same as that in the objective function mentioned above in step 5; and The penalty parameters initialized in step 1; and Constraints and The corresponding dual variables have the following initial values: , ; Step 6: Solve using ADMM-based expansion iteration: The augmented Lagrangian function obtained in step 5 is processed using the traditional alternating direction multiplier method. Perform iterative solutions and update sequentially. , , , , and The specific steps are as follows: Step 6.1, fix the first The next iteration , and Update the auxiliary variable as follows , obtained the The next iteration : in, For millimeter-wave imaging forward operator as described in 23 The adjoint operator; To and Identity matrices of the same dimension; superscript Indicates matrix inversion; superscript and Indicates the number of iterations for the corresponding variable; This is the two-dimensional complex-valued echo in the azimuth-elevation direction obtained in step 2; The penalty parameters initialized in step 1; Step 6.2, fix the first The next iteration and Update the auxiliary variable as follows , obtained the The next iteration : in, This represents the element-wise soft thresholding function, which applies to any input element. ,when Time output ,when Output 0 at time. Soft threshold, , and Same meaning as in step 5; Step 6.3, Fix , , , and Update the imaging results to be optimized using the following formula. In position The element at that position, obtain : in, and The position index of the element. , , , , , and They represent , , , , , and In position The element at that location, and The penalty parameters initialized in step 1, Numerical stable small constants initialized in step 1; Step 6.4, Fix , and Update as follows In position The element at that position, obtain : in, , , and They represent In the upper adjacent position , next to Left neighbor position and right neighbor position The element at that location, , and Same meaning as in step 6.3, and Same meaning as in step 5; Step 6.5: Based on the iteration results from Steps 6.1 to 6.4, update the dual variable using the following formula. and : in, This represents the Hadamard product, which is the multiplication of elements at corresponding positions. Step 6.6: Calculate the difference between the imaging results obtained from two adjacent iterations using the Frobenius norm described in 20. Less than the iteration termination threshold initialized in step 1 When, output Otherwise And return to step 6.1; Step 7: Generate 3D millimeter-wave imaging results: The optimized imaging results obtained from each distance slice in step 6 The target scene is processed using a traditional stacking method based on distance, resulting in a three-dimensional millimeter-wave sparse imaging result, thus completing the three-dimensional characterization of the target under an ultra-sparse array.

Images