Computationally Efficient Short-Range 3D SAR Imaging via Modified Augmented Lagrangian-Based FA-ADMM and Deep Learning Models

Article information

J. Electromagn. Eng. Sci. 2026;26(3):280-295
Publication date (electronic) : 2026 May 31
doi : https://doi.org/10.26866/jees.2026.3.r.364
Department of Smart Information and Technology Engineering, Kongju National University, Cheonan, South Korea
*Corresponding Author: Ic-Pyo Hong (e-mail: iphong@kongju.ac.kr)
Received 2025 February 20; Revised 2025 July 31; Accepted 2025 November 4.

Abstract

Short-range three-dimensional (3D) synthetic aperture radar (SAR) imaging has drawn significant attention across various domains, including security surveillance, non-destructive testing, and medical diagnostics. This paper introduces a fast adaptive alternative direction method of multipliers (FA-ADMM) framework designed to enhance both efficiency and accuracy in SAR image reconstruction. Our approach addresses two key challenges in the single holographic frequency ADMM (SFH-ADMM) model: image degradation from fast Fourier transform operations and slow convergence due to fixed ADMM penalty parameters. To overcome these issues, we refine the augmented Lagrangian formulation to ensure stable convergence and introduce an adaptive tuning mechanism that dynamically adjusts penalty parameters based on the connection between relaxed ADMM and relaxed Douglas–Rachford splitting. Additionally, we seamlessly integrate denoising convolutional neural network and autoencoder architectures into the iterative process to enhance noise suppression and image fidelity, respectively. The synergy of these innovations within a unified framework significantly accelerates convergence and improves reconstruction quality, making it well suited for real-world short-range 3D SAR applications.

I. Introduction

Millimeter-wave (MMW) synthetic aperture radar (SAR) imaging has prompted significant developments within the scientific community, especially with regard to short-range applications [1, 2]. This interest stems from the inherent advantages of MMW radar, which remains unaffected by optical constraints and enables high-resolution imaging in both the range and azimuth directions. Specifically, MMW SAR exhibits strong penetration capabilities through non-metallic materials and maintains operational effectiveness under diverse environmental conditions, including adverse weather, fire, fog, and dust, regardless of the time of day.

In recent years, considerable research efforts have been dedicated to advancing three-dimensional (3D) SAR holography, with a focus on developing novel imaging methodologies, image- focusing techniques, and super-resolution algorithms. In the context of short-range SAR imaging, fast Fourier transform (FFT)-based approaches, such as matched filter (MF) and range migration algorithm (RMA), have been widely adopted as primary image reconstruction techniques. The studies in [3] and [4] introduced phase compensation methods to enhance image focusing by mitigating the frequency-domain inconsistencies between transceiver antennas and their corresponding virtual elements that arise from FFT-based processing. Meanwhile, in [5], a multiplanar multi-static framework was constructed using RMA within a multi-input multi-output (MIMO) SAR configuration to address the irregular scanning geometries encountered in flexible imaging platforms, such as handheld scanners, and unmanned aerial vehicle (UAV)-based imaging.

Although FFT-based techniques help achieve high-quality imaging, they present two major limitations. First, their performance deteriorates under low signal-to-noise ratio (SNR) conditions, where strong clutter may obscure target detection. Second, they are constrained by Nyquist sampling criteria, which must be satisfied to prevent aliasing and preserve image texture [6]. To address the latter, several studies have explored the use of nonuniform FFT (NUFFT) algorithms [7, 8]. However, this method requires extensive interpolation, which significantly increases computational complexity while reducing image fidelity.

To address the limitations of FFT-based image reconstruction methods, recent research has increasingly focused on compressed sensing (CS) and sparse reconstruction algorithms [9, 10]. These approaches enable a substantial reduction in the number of required array elements, thereby eliminating the need for uniform element distribution in MIMO systems while preserving both image quality and accuracy. CS-based techniques generally fall into two categories: greedy algorithms and basis pursuit algorithms. The latter directly obtains the sparsest solution by solving a convex optimization problem that minimizes the ℓ1-regularization term [11]. Notable methods in this domain include the fast iterative shrinkage-thresholding algorithm (FISTA) [12] and the alternating direction method of multipliers (ADMM) [13].

ADMM has been widely adopted as an effective image reconstruction technique in SAR imaging. For instance, the study in [14] employed ADMM to jointly enforce constant false alarm rate (CFAR) and total variation (TV) regularizations, thereby enhancing sparse target detection in highly cluttered and noisy environments. Additionally, the authors of [15] introduced the equalized ADMM as an effective solver to improve autofocusing performance in scenes with moving targets and platforms. However, these approaches involve extensive matrix operations and inversions, which demand considerable hardware resources and processing time. To mitigate these issues, the work in [16] leveraged the single-frequency holography (SFH) method [17] to construct FFT-based forward and backward operators for processing frequency-domain data. By integrating the operators into the ADMM framework, the method circumvents vectorization and large-scale matrix inversions, significantly reducing both computational complexity and processing time. However, despite its advantages, the SFH-ADMM algorithm has two notable limitations. First, its ADMM parameters are manually set and remain fixed throughout the iterations, potentially leading to slow convergence. Second, SFH operators depend on FFT, which causes image degradation, particularly when applied in the context of CS-based sparse reconstruction.

Recent advancements in deep learning (DL) have led to the development of learned reconstruction methods [18, 19], which have emerged as effective alternatives to the traditional regularization- based approaches employed in plug-and-play (PnP) ADMM frameworks [20]. By combining data fidelity with prior information encoded by DL-based denoising models, these methods not only suppress noise in the reconstructed image but also recover target details that may have been distorted due to the sparsity constraints in reconstruction process [21, 22].

In this paper, we propose a fast adaptive ADMM-based image reconstruction framework tailored to short-range 3D SAR imaging. Our approach builds on the modified augmented Lagrangian (AL) formulation of the conventional ADMM algorithm by incorporating SFH operators [17] and exploiting the theoretical connection between relaxed ADMM [23] and relaxed Douglas–Rachford splitting (DRS) [24]. The proposed approach enhances convergence performance through a combination of simplified update steps, an adaptive ADMM parameter adjustment mechanism, and the avoidance of large-scale matrix inversions. Furthermore, complex-valued DL-based denoising models, specifically convolutional neural network (CNN) and autoencoder (AE), are seamlessly incorporated into the iterative ADMM process. This integration not only accelerates ADMM iteration under conditions of low sparse sampling and high phase errors but also leverages learned spatial patterns derived from complex-valued feature priors, further improving imaging performance.

II. Related Work

Fig. 1 illustrates an MMW frequency-modulated continuous wave (FMCW) radar operating across an aperture plane (x, y, 0), which transmits signals and receives echoes reflected from a target located at distance z on the target plane (x′, y′, z). The signals scattered from the target and collected at all measurement points on the aperture plane form a 3D frequency-domain data cube, which can be expressed as follows:

Fig. 1

The measurement system and coordinate definitions.

(1) s(x,y,k)=∭σ(x′,y′,z)e-j2kRdx′dy′dz

where s ∈ ℂNx×Ny×Nz refers to the 3D echo data cube, Nk denotes the number of fast-time sampling points, while Nx and Ny correspond to the number of slow-time sampling points in the x- and y-directions on the aperture plane, respectively. Furthermore, σ ∈ ℂNw×Nh×Nl represents the 3D complex-valued reflectivity coefficient of the specimen under test (SUT), while Nw, Nh and Nl are the number of pixels in the x′-, y′- and z-directions, respectively. In addition, j=-1, and k=2πfc is the wavenumber of the corresponding fast-time frequency f with respect to the speed of light c. In the context of 3D imaging, the instantaneous distance R between the transceiver antenna and a range cell r is given by:

(2) R=(x′-x)2+(y′-y)2+zr2

where zr refers to the location of the range cell r.

Now, considering a specific target slice located at range cell ri with a corresponding distance of zi from the radar, the wavenumber remains constant as k=2πfbc, where fb=2RKc is the beat frequency, representing the fast-time frequency difference caused by the time delay between signal transmission and reflection from the target at distance R with frequency slope K. Consequently, the 2D sub-echo slice at this range cell can be formulated as:

(3) s(x,y)=∬σ(x′,y′)e-j2k(x′-x)2+(y′-y)2+zi2dx′dy′

The exponential term in Eq. (3), representing a spherical wave, can be approximated as a superposition of the plane wave using the stationary phase method [25], yielding the following expression:

(4) e-j2k(x′-x)2+(y′-y)2+zi2≈∬e-j(kx(x′-x)+ky(y′-y)-kzzi)dkxdky

where kx, ky and kz are the components of k in the Cartesian coordinate system, which adhere to the following relationship:

(5) kz=4k2-kx2-ky2kx2+ky2≤4k2

By substituting Eq. (4) to Eq. (3), the following expression is attained:

(6) s(x,y)=∬σ(x′,y′)·∬e-j(kx(x′-x)+ky(y′-y)-kzzi)dkxdkydx′dy′

Furthermore, by rearranging the phase term in Eq. (6), we obtain:

(7) s(x,y)=∬[∬σ(x′,y′)e-j(kxx′+kyy′)dx′dy′]·ej(kxx+kyy)dkxdky·ejkzzi=FT2D†[FT2D[σ(x′,y′)]·ejkzzi]

where FT2D[·] and FT2D†[·] represent the 2D FFT and 2D inverse FFT (IFFT) operators, respectively. Consequently, the 2D image of the considered slice at range cell ri can be reconstructed from the complex reflectivity function using the SFH method, as follows:

(8) σ(x′,y′)=FT2D†[FT2D[s(x,y)]·e-jkzzi]

Assuming that the matrix representation of the sub-echo slice s(x, y) is S ∈ ℂNx×Ny and the corresponding SUT reflectivity σ(x′, y′) is ∑ ∈ ℂNw×Nh , the relationship between S and ∑, as established by Eqs. (7) and (8), can be expressed as:

(9) S=F{Σ}≜FT2D†[FT2D[Σ]⊙Φ]
(10) Σ=F†{P′{S}}≜FT2D[FT2D[P{S}]⊙Φ¯]

where and are the measurement and imaging operators corresponding to the SFH backward and forward operators, respectively. Furthermore, the phase matrix Φ is defined with element ejkzzi paired with its conjugate Φ̄ . Furthermore, ⊙ denotes the Hadamard product, and the zero-padding operator P{·} is defined as follows:

(11) P{S}mn={Smnif {Nw-Nx2≤m≤Nw+Nx2Nh-Ny2≤n≤Nh+Ny20other.

III. Proposed Method

In the proposed method, we employed SFH operators in the proposed framework with the aim of alleviating the computational burden caused by the large sensing matrix and repeated matrix inversions in conventional ADMM-based SAR imaging, as observed in the functional measurement model discussed in [16]. Building upon this foundation, this section presents a FA-ADMM framework that integrates a modified AL formulation with adaptive parameter tuning to improve convergence speed and reconstruction stability. In addition, DL-based denoisers were incorporated to further enhance noise suppression and preserve structural fidelity in the reconstructed SAR images.

1. Fast adaptive ADMM (FA-ADMM)

Considering M ∈ ℂNxNy×NwNh as the matrix representation of the sensing, with its elements given by e-j2k(x′-x)2+(y′-y)2+zi2, Eq. (3) was reformulated into a sparse matrix representation, expressed as follows:

(12) S=unvec(Mvec(Σ))

where unvec(·) denotes the matrix operator while vec(·) represents the vectorization operator. The imaging process can be interpreted as a reconstruction of the reflectivity function ∑ the acquired from the sub-echo S. By utilizing the sparsity inherent in S, Eq. (12) was refined into a linear-square problem using an augmented ℓ1-regularization term, as follows:

(13) Σ^=arg minΣ12‖S-unvec(Mvec(Σ))‖22+λ‖Σ‖1

where ||·||1 and ||·||2 are the ℓ1and ℓ2-norms capturing the sparsity and data fidelity properties of the target signal, respectively, while λ is the regularization coefficient. To solve the optimization problem using the ADMM approach, an auxiliary variable matrix Z ∈ ℂNw×Nh was introduced to decouple it into two independent penalty terms:

(14) Σ=arg minΣ12‖S-unvec(Mvec(Σ))‖22+λ‖Z‖1s.t. Σ=Z

Accordingly, the corresponding AL function was reformulated, diverging from the standard formulation proposed in [13], as follows:

(15) ℒρ(Σ,Z,U)=12‖S-unvec(Mvec(Σ))‖22+λ‖Z‖1+ρ2‖Σ-Z-U‖22

where U ∈ ℂNw×Nh represents the matrix form of the Lagrange multiplier (or dual variable) and ρ signifies the penalty parameter.

At this point, it must be emphasized that the relationship between the SFH operators in Eqs. (9) and (10) and the matrix operators remained consistent, meaning that:

(16) {S=F{Σ}⇔S=unvec(Mvec(Σ))Σ=F†{P{S}}⇔Σ=unvec(MHvec(S))

where (·)H is the conjugate transpose operator. Consequently, the AL function in Eq. (15) was reformulated as follows:

(17) ℒρ(Σ,Z,U)=12‖S-F{Σ}‖22+λ‖Z‖1+ρ2‖Σ-Z-U‖22

Notably, the variables of the AL function are separable, which means that they can be solved via the modeling of three iterative sub-problems:

(18) {Σ(t)=arg minΣℒρ(Σ,Z(t-1),U(t-1))Z(t)=arg minZℒρ(Σ(t),Z,U(t-1))U(t)=U(t-1)-(Σ(t)-Z(t))

where t denotes the iteration index. Drawing on the first derivatives of the AL function with respect to ∑ and Z, we managed to solve the first two sub-problems in Eq. (18). Furthermore, upon applying the relaxed ADMM method [23], the iterative solutions became:

(19) {Σ(t)=11+ρ(t)[F2{P{S}}+ρ(t)(Z(t-1)+U(t-1))]Σ˜(t)=α(t)Σ(t)+(1-α(t))Z(t-1)Z(t)=arg minZλ‖Z‖1+ρ(t)2‖(Σ˜(t)-U(t-1))-Z‖22 =J(Σ˜(t)-U(t-1);λρ(t))U(t)=U(t-1)-(Σ˜(t)-Z(t))

where α ∈ (0,2) is the relaxation parameter and is the soft-threshold function [16]. Notably, the convergence efficiency of a relaxed ADMM is significantly influenced by the choice of α(t) and ρ(t). Therefore, a rapid adaptive penalty method exploiting the intimate connection between relaxed ADMM and relaxed DRS [24] was employed to iteratively tune these parameters, thereby tailoring them to FA-ADMM.

First, h(Σ)=12‖S-F{Σ}‖22 and g(Z) = λ||Z||1 were taken as the subfunctions in Eq. (14), with θ(t) > 0 and ϕ(t) > 0 being their local curvature estimates, respectively. These curvature estimates were inferred from the outputs of the current iteration t and the preceding iteration t0 < t, as follows:

(20) θ(t)={θMG(t)if 2θMG(t)>θSD(t)θSD(t)-12θMG(t)otherwise

with

(21) {θMG(t)=〈Σ(t)-Σ(t0),U^(t)-U(t0)〉〈Σ(t)-Σ(t0),Σ(t)-Σ(t0)〉θSD(t)=〈U^(t)-U(t0),U^(t)-U(t0)〉〈Σ(t)-Σ(t0),U^(t)-U(t0)〉

where MG and SD stand for the minimum gradient and steepest decent, respectively. Furthermore, 〈M, N〉 refers to the inner product of M and N , and Û(t) = U(t−1) − (∑(t) − Z(t)) is the intermediate dual variable. Similarly, ϕ(t) for subfunction g(Z) was estimated by replacing ∑(t) − ∑(to) and Û(t) − U(to) with Z(t) − Z(to)and U(t) − U(to), respectively, as follows:

(22) φ(t)={φMG(t)if 2φMG(t)>φSD(t)φSD(t)-12φMG(t)otherwise

with

(23) {φMG(t)=〈Z(t)-Z(t0),U(t)-U(t0)〉〈Z(t)-Z(t0),Z(t)-Z(t0)〉φSD(t)=〈U(t)-U(t0),U(t)-U(t0)〉〈Z(t)-Z(t0),U(t)-U(t0)〉

The curvature estimates were then used to update the penalty and relaxation parameters ρ(t) and α(t), as follows:

(24) ρ(t)={θ(t-1)φ(t-1)if θcorr(t-1)>ɛcorr and φcorr(t-1)>ɛcorrθ(t-1)if θcorr(t-1)>ɛcorr and φcorr(t-1)≤ɛcorrφ(t-1)if θcorr(t-1)≤ɛcorr and φcorr(t-1)>ɛcorrρ(t-1)otherwise

and

(25) ρ(t)={θ(t-1)φ(t-1)if θcorr(t-1)>ɛcorr and φcorr(t-1)>ɛcorrθ(t-1)if θcorr(t-1)>ɛcorr and φcorr(t-1)≤ɛcorrφ(t-1)if θcorr(t-1)≤ɛcorr and φcorr(t-1)>ɛcorrρ(t-1)otherwise

where εcorr is the safeguard threshold for the local curvature estimates, while θcorr(t-1) and φcorr(t-1) indicate the correlation between the relaxed ADMM and relaxed DRS, expressed as:

(26) θcorr(t-1)=〈Σ(t-1)-Σ(t0),Ul(t-1)-U(t0)〉‖Σ(t-1)-Σ(t0)‖2‖u˙(t-1)-U(t0)‖2
(27) φcorr(t-1)=〈Z(t-1)-Z(t0),U(t-1)-U(t0)〉‖Z(t-1)-Z(t0)‖2‖U(t-1)-U(t0)‖2

Notably, updates to ρ(t) are selectively influenced by estimations pertaining to the model’s components. Meanwhile, in the case of updating α(t), when the model demonstrates strong performance in approximating h(∑) but not g(Z), a larger relaxation parameter α(t) = 1.9 should be adopted to enforce a more conservative update of Z relative to ∑. In contrast, when the model performs well for g(Z) but inadequately for h(∑), a smaller value α(t) = 1.1 must be employed to enable a more aggressive update of Z. In this context, to enhance stability and reduce computational overhead, both ρ(t) and α(t) should be updated at every other iterations rather than continuously, as follows:

(28) ρ(t)={Updated using (24)if tmod2==0ρ(t-1)if tmod2==1
(29) α(t)={Updated using (25)if tmod2==0α(t-1)if tmod2==1

This staggered update scheme helps avoid unnecessary recalculations while enhancing the convergence behavior of the algorithm [24]. Furthermore, the stopping criteria of the iteration process can typically be defined as follows:

(30) {‖Σ(t)-Z(t)‖2≤ɛtolmax(‖Σ(t)‖2,‖Z(t)‖2)‖ρ(t)(Z(t)-Z(t-1))‖2≤ɛtol‖U(t)‖2

where εtol is the stopping tolerance.

Therefore, the FA-ADMM framework is characterized by a parameter-tuning strategy, as illustrated in Fig. 2, with the maximum number of iterations being T. Each iteration t incorporates four cascaded modules—a reconstruction module ∑(t), a relaxation module ∑̃(t), an auxiliary module Z(t), and a dual variable updating module U(t)—corresponding to each of the subproblems in Eq. (19).

Fig. 2

Structure of the FA-ADMM method.

In this context, it is worth noting that the third subproblem in Equation (19) was designed to minimize the residual between (∑(t) − U(t−1)) and Z, which is analogous to a denoising process under prior constraint ||Z||1. In other words, this formulation presents an opportunity to establish a PnP approach by incorporating a robust denoiser, thereby enhancing imaging performance, including convergence precision and image quality, especially under sparse sampling and low SNR scenarios. Hence, the Z(t) module can be considered a denoising module that operates on the data originating from the relaxation module ∑̃(t), as follows:

(31) Z(t)=D{Σ˜(t)}

where represents a denoiser.

This FA-ADMM approach aims to effectively reduce the number of iterations required and enhance the robustness of the proposed framework with respect to parameter sensitivity. Notably, the adaptive adjustment of algorithmic parameters plays a vital role in ensuring that the ADMM updates remain responsive to changes in the reconstruction process, thereby promoting stability and convergence.

Nevertheless, owing to the FFT and IFFT in the SFH operators, the FA-ADMM approach is constrained by noise and degraded reconstruction quality of the target image stemming from the undersampling ratio and low SNR conditions. To address these issues, we enhanced denoising performance and improved the preservation of SUT information by incorporating unrolling techniques and integrating pre-trained DL-based denoising models into the iterative FA-ADMM framework, respectively, resulting in the development of a tailored FA-DL-ADMM methodology. The details of this integration are elaborated on in Section III-2.

2. Deep Learning-based Denoiser Integration

Fig. 3(a) provides an overview of the proposed FA-DL-ADMM framework for 3D SAR image reconstruction. The reconstruction process by subjecting a noisy image slice in range cell ri to the SFH forward operator , as expressed in Eq. (10), and then processed using the AL formulation. This formulation decomposes the reconstruction objective into three interrelated subproblems, each addressed by a dedicated update module:

Fig. 3

Diagram of the proposed framework: (a) the FA-DL-ADMM iterative procedure and (b) detailed structures of the three layers for solving the AL function.

  • (i) ∑-update (reconstruction layer): Depicted in Fig. 3(b) as the first block, this layer implements a physics-informed update that enforces data fidelity using the SFH operator. This step involves adaptive updates of the penalty parameter ρ(t) and relaxation parameter α(t), which are dynamically tuned based on local spectral curvatures to improve convergence efficiency.

  • (ii) Z-update (denoising layer): This is the second block in Fig. 3(b), which functions as a learned prior owing to a DL-based denoiser . This layer leverages both magnitude and phase information (|∑̃(t)| and ∠∑̃(t)) as dual-input features to preserve fine structural details and improve noise suppression. In this study, we employed two distinct denoising architectures— namely, CNN and AE, as depicted in Fig. 4, which are described later in this section.

  • (iii) U-update (dual variable layer): This layer is responsible for updating the Lagrange multiplier to ensure consistency between ∑̃(t) and Z(t), thereby enforcing the coupling constraint of the AL formulation.

These components operate within the FA-ADMM optimization loop, depicted in Fig. 3(a) as the area within the orange dashed boundary. Through iterative refinement across multiple stages, the algorithm produces a final reconstructed SAR image slice ∑(t) corresponding to each range cell ri .

  • (iv) Denoising network architectures: In conventional DL-based image denoising approaches, the magnitude component is usually prioritized due to its direct correlation with image intensity. However, such magnitude-only approaches may lead to loss of information, particularly in complex-valued data representations, where the phase component plays a critical role in defining the target’s structural integrity. Therefore, in this study, a dual-input strategy was leveraged to effectively capture not only intensity variations but also phase-related spatial features, leading to improved noise suppression and enhanced target shape reconstruction. Specifically, during training, these networks were configured to accept input sizes of 128×128×2 (Nw × Nh × 2), consisting of two 2D matrices corresponding to the magnitude and phase of complex-valued ∑̃(t).

As shown in Fig. 4(a), the CNN model consists of seven convolutional 2D (Conv2D) layers with convolutional channel sizes of 8, 8, 16, 32, 64, 128, and 2, respectively, maintaining a consistent 3 × 3 filter size across all layers. To enhance training stability and convergence, batch normalization (BN) layers are incorporated after each of the second through fifth Conv2D layers. Additionally, to mitigate overfitting, dropout layers at a rate of 0.25 are introduced after the BN layers following the third and sixth Conv2D layers.

Meanwhile, the AE architecture features an encoder–decoder framework, as depicted in Fig. 4(b). The encoder is composed of three Conv2D layers with a filter size of 3 × 3 and convolutional channel sizes of 128, 64, and 32, respectively, ending with a dropout layer with a rate of 0.25. By down-sampling the layers, the encoder effectively compresses the noisy input, eliminating redundant noise, while preserving the essential structural features of the target object. Meanwhile, the decoder, designed to reconstruct the denoised output, consists of two Conv2D transpose layers with the same filter size of 3 × 3 and convolutional channel sizes of 64 and 128 that are responsible for progressively up-sampling the latent space. This is followed by another Conv2D layer with a channel size of 128, and a dropout layer with a rate of 0.25. Finally, a Conv2D layer with a filter size of 2 and a sigmoid activation function are added to complete the reconstruction process produce the final denoised out

For both models, we employed a combined loss function, as formulated below:

(31) Lcombined=λ1Lmag+λ2Lphase+λ3LSSIM

where λ1 , λ2 and λ3 denote the weights of the magnitude loss (Lmag), phase loss (Lphase), and structural similarity index measure (SSIM) loss (LSSIM), respectively. These functions can be defined as follows:

(33) Lmag=‖Y^mag-Ymag‖1

where Ŷmag and Ymag denote the magnitude component of ground truth Ŷ and the corresponding reconstructed output Y obtained from the models, respectively. Furthermore,

(34) Lphase=1N∑iN(∣sin (ϕdiff,i)∣+∣cos (ϕdiff,i)∣)

where N refers to the number of elements of Ŷ and Y, while φdiff is their phase difference.

(35) LSSIM=1-SSIM(Y^,Y)

with

(36) SSIM(Y^,Y)=2μY^μY+C1μY^2+μY2+C1·2ςY^Y+C2ςY^2+ςY2+C2

where μŶ, μY, ςŶ and ςY denote the mean values and standard deviations of the elements in Ŷ and Y, respectively. Furthermore, ςŶY is the covariance of the elements in these two matrices, while C1 and C2 are the constants adopted to stabilize the division with a weak denominator.

During training, the loss weights λ1, λ2 and λ3 were kept constant at 1, 1, and 0.01, respectively. These chosen loss weights reflect balanced emphasis, indicating that equal importance was placed on magnitude and phase fidelity. The SSIM loss was assigned a smaller weight (0.01) to promote perceptual similarity without overwhelming pixel-wise losses. Notably, these values were determined empirically based on preliminary experiments, wherein they demonstrated a favorable tradeoff between quantitative metrics (e.g., PSNR) and qualitative visual performance.

The CNN and AE architectures used in this study were also selected through empirical evaluation. Various architectural configurations were investigated, including different network depths (3 to 8 layers), convolutional kernel sizes, and skip connections. We also tested different bottleneck sizes for the AE and compared input formats (magnitude-only vs. combined magnitude and phase). The selected configurations offered a favorable trade-off between reconstruction accuracy and model complexity. In particular, increasing the depth beyond six layers provided negligible performance gain while increasing overfitting risk and computational cost. Although we did not conduct a formal ablation study due to space limitations, these design choices were rigorously validated on a held-out synthetic dataset. Nonetheless, while the architectures selected in this study demonstrated satisfactory performance, we do not state them to be optimal solutions. Further benchmarking of network structures capable of generalizing across diverse scenarios remains a compelling area for further investigation.

IV. Simulation and Experiment Results

This section presents a comprehensive evaluation of the proposed methodologies through both simulation and real-world experiments. The performance of the FA-ADMM and FADL-ADMM frameworks was systematically compared with conventional techniques, including MF-based imaging and SFH-ADMM, which serve as benchmark methods. As outlined in Section III-2, the FA-DL-ADMM framework incorporates two deep learning models—CNN and AE—giving rise to the FA-CNN-ADMM and FA-AE-ADMM approaches, respectively. Numerical simulations were conducted under varying SNR level conditions and data sampling rates (SRs) to rigorously validate the effectiveness of the proposed methods. Additionally, experimental validation was performed using a commercial MMW 3D SAR system, further demonstrating the superior performance of the proposed framework.

1. Image Scheme Configuration and Data Augmentation

Table 1 presents the configuration of the imaging scheme employed in the simulations and experiments. We employed an FMCW radar operating in the 77 to 81 GHz frequency range with a frequency slope of 63.343 MHz per microsecond (μs) and a sampling frequency of 9.121 mega samples per second. For all simulations and experiments, the target was positioned at distance z, varying from 200 to 500 millimeters (mm) relative to the aperture plane of size 200 × 200 mm (Dx × Dy). The radar transmitted signals and captured echoes for each sampling interval of 2 mm in both horizontal and vertical directions. These radar and imaging configurations enabled us to obtain a 3D data cube size of 100 × 100 × 256 (Nx × Ny × Nk) for each SUT. Furthermore, the approximated azimuth and range resolutions of δa ∈ [2.92, 4.86] mm and δr ≈ 3.75 centimeters (cm), respectively, were calculated using the following equations:

Configuration of the imaging scheme

(37) δa≈λwz2D,         δr=c2B

where λw is the wavelength, D refers to the aperture length in the horizontal or vertical direction, and B indicates the bandwidth.

The initial parameters for the proposed FA-ADMM framework were empirically determined with the aim of balancing convergence speed and stability. A penalty parameter of ρ(0) = 1 provided a balanced starting point, avoiding excessive penalty strength and ensuring stable early iterations. Meanwhile, a relaxation factor of α(0) = 1.9, which is close to the relaxed ADMM–DRS upper limit, was selected to accelerate convergence without compromising stability. Meanwhile, λ was fixed at 0.01 across all experiments since it provided a balance between data fidelity and sparsity regularization during preliminary trials. Furthermore, a safeguard threshold of εcorr = 0.2 and a stopping tolerance of εtol = 0.01 were empirically set to ensure consistent and smooth convergence.

Based on the system configurations outlined above, we generated simulated 3D echo data cubes for 12 distinct categories of commonly observed metallic objects, as summarized in Table 2. Each category comprised multiple target instances, each possessing unique characteristics. To enhance data augmentation, each target instance was subjected to rotations at four discrete angles, ranging from 0° to 360° in increments of 90°, across 10 distinct range positions along the z-axis. Consequently, 40 simulated 3D data cubes were generated for each SUT. To generate ground truth data for training the DL-based denoising models, we reconstructed noise-free 2D SAR images for specific range-cell slices by implementing the methodology discussed in Section II, using the full data cube (100%) for each SUT. The corresponding noisy inputs were the 2D SAR images generated from the noisy data cubes snoisy, defined as:

Simulated object quantities

(38) snoisy=B⊙(s+n)

where n ∈ ℂNx×Ny×Nz is the additive Gaussian white noise (AGWN), and B ∈ ℤNx×Ny refers to the random binary mask matrix representing the sparse sampling pattern with an SR of κ: 0 < κ < 1, which is given by:

(39) κ≜‖B‖1Nx×Ny

In this paper, κ was randomly selected as 0.1, 0.2, 0.25, and 0.5. This simulation strategy enabled us to assemble a comprehensive dataset comprising 7,200 pairs of noisy inputs and their associated ground truths, of which 80% were used as training data and 20% served as testing data. Since the training dataset was specifically generated based on the SUT categories relevant to this study, each model was fine-tuned and optimized to identify and enhance the features specific to these classes rather than functioning as general-purpose models for all potential target categories.

Nonetheless, while the training data used in this study were used to effectively capture a range of metallic objects exhibiting varying shapes, orientations, and range positions, it is limited to only synthetic simulations of specific categories. This may constrain the generalization capability of the denoising models for real-world SAR data or unseen target types that exhibit different structural or scattering properties. However, unlike purely datadriven models, the PnP formulation separates the physics-based data fidelity term—derived from the SFH forward model—from the learned prior encoded by the denoising network. This modular design enables better adaptability to new scenes, even when the denoiser is trained on a limited dataset. Moreover, the iterative ADMM process integrated the denoiser into the flexible optimization loop, which helped maintain reconstruction stability and fidelity, even when the denoising prior did not perfectly match the test data. These properties contribute to the framework’s robustness in scenarios involving moderate domain shifts.

2. Simulation Results

To quantitively evaluate the performance of the considered methods in simulation, we took the SSIM (as in Eq. (36)), peak SNR (PSNR), and the normalized mean square error (NMSE) between the reconstructed image Y and its associated ground truth Ŷ into consideration, defined as:

(40) PSNR(Y^,Y)=10log10Ymax2‖Y^-Y‖22

and

(41) NMSE(Y^,Y)=10log10‖Y^-Y‖22‖Y^‖22

The PSNR metric quantifies image quality, with a higher PSNR value indicating superior reconstruction, while NMSE evaluates discrepancies between the reconstructed image and its ground truth, where a lower NMSE signifies a closer match to the ground truth, and vice versa. Meanwhile, SSIM reflects the similarity between Ŷ and Y, with the best possible score being 1.

Furthermore, we conducted Monte Carlo simulations under multiple conditions by varying both SNR and SR. For each condition, 25 independent trials were conducted using randomized noise realizations to evaluate the statistical consistency of the reconstruction results. Table 3 presents a quantitative comparison of image quality metrics obtained by all evaluated approaches for 100 distinct simulated target data, with the best scores highlighted in bold. First, we examined the reconstruction performance for different SRs, ranging from 10% to 90% in steps of 20%, while maintaining a constant SNR of 30 dB. As shown on the left side of Table 3, all ADMM-based methods demonstrate superior performance across all SRs compared to the conventional MF-based approach. Furthermore, the integration of DL-based denoisers enhanced the performance of the FA-DL-ADMM methods, with FA-AE-ADMM achieving the highest scores. Specifically, FA-AE-ADMM surpasses the FA-ADMM method by 11.03 dB in PSNR, 11.6 dB in NMSE, and 0.112 in SSIM, on average. This outcome underscores the advantages of implementing the unrolling technique in terms of improving image reconstruction efficiency.

Numerical evaluation of simulations under varied SRs and SNRs

Second, the image reconstruction performance was evaluated under varying SNR levels ranging from 0 to 20 dB at increments of 5 dB, with the SR fixed at 50%. A similar trend to the previous analysis is observed on the right side of Table 3, showing that the proposed methods consistently deliver superior performance across all SNR levels. The FA-DL-ADMM approaches, especially FA-AE-ADMM, demonstrate enhanced robustness in image reconstruction, yielding better evaluation metrics compared to the other approaches. Additionally, the last column in Table 3 presents the average computational duration of all methods, confirming that all the ADMM-based algorithms require significantly less processing time than the MF-based method, with the proposed fast adaptive schemes offering substantial computational efficiency improvements. Specifically, it is observed that the image reconstruction time for a single 3D SAR image using the FA-ADMM method is approximately one-third of that of the conventional SFH-ADMM. Moreover, the FA-DL-ADMM method achieved a further reduction in computational time, demonstrating significant efficiency in processing.

To assess computational efficiency, we analyzed the iterative performance of all the ADMM-based methods by measuring the average number of iterations and the average computation time required per iteration to reconstruct a 3D SAR image, the results of which are summarized in Table 4. Although the FA-ADMM method has a slightly higher per-iteration computation time compared to SFH-ADMM (0.26 vs. 0.21 seconds), its adaptive parameter updating mechanism significantly reduced the average number of required iterations (6.79 vs. 22.99), owing to the additional tuning parameter calculations. This resulted in a more than two-thirds reduction in the total runtime, highlighting the effectiveness of the adaptive mechanism in accelerating convergence while preserving reconstruction quality. Furthermore, the FA-DL-ADMM framework benefits from the integration of pretrained DL models and GPU acceleration, achieving not only the lowest average number of required iterations but also the lowest average per-iteration computation time and overall runtime. More specifically, the FA-AE-ADMM method took only 3.86 iterations to reconstruct a 3D SAR image, while the FA-CNN-ADMM method needed just 0.14 seconds to compute one iteration. These results further emphasize their efficiency in SAR image reconstruction.

Computational duration of ADMM-based methods

Fig. 5 presents boxplots of the PSNR and SSIM metrics obtained from 25 independent Monte Carlo trials under each SR. These results reflect the statistical stability and reconstruction consistency of the proposed FA-ADMM framework when integrated with two different DL-based denoising priors. The boxplots illustrate the distribution and variability of performance across trials. It is observed that as the sampling rate increases, both PSNR and SSIM improve, consistent with expectations. Notably, FA-AE-ADMM displays a higher median PSNR and SSIM across nearly all sampling rates, indicating superior reconstruction accuracy. Additionally, FA-AE-ADMM shows greater consistency in SSIM with narrower interquartile ranges, especially in the mid- to high-sampling regimes (30%–90%). While FA-CNN-ADMM demonstrates competitive PSNR performance at low sampling rates, it also exhibits higher variability in both metrics, suggesting reduced robustness under sparse sampling. These results confirm the adaptability of the FA-ADMM framework and highlight the performance advantage offered by the AE-based denoiser under a wide range of measurement conditions.

Fig. 5

Boxplots of (a) PSNR and (b) SSIM obtained by conducting over 25 Monte Carlo runs for different sampling rates using the two FA-DL-ADMM methods

To evaluate the structural reconstruction quality of the proposed methods, we present the 3D imaging results for four distinct object types—a gun, a spanner, a knife, and a pair of pliers— in Fig. 6. These objects were placed at varying distances from the synthetic aperture plane: 411 mm (gun), 403 mm (spanner), 323 mm (knife), and 332 mm (pliers). Furthermore, to simulate real-world undersampling conditions, each object was reconstructed using a different SR and SNR: 10% and 30 dB (gun), 30% and 20 dB (spanner), 50% and 10 dB (knife), 70% and 0 dB (pliers). These settings allowed us to assess reconstruction robustness across a wide range of signal sparsity levels. In Fig. 6, each row corresponds to a different object at a specific SR, while the columns present the results pertaining to ground truth, the MF-based method, SFH-ADMM, and the proposed FA-ADMM, FA-CNN-ADMM, and FA-AE-ADMM. All outputs were normalized and rendered using consistent isosurface thresholds.

Fig. 6

Reconstructed 3-D SAR images obtained using different methods, with SRs–SNRs being 10%–30 dB (gun), 30%–20 dB (spanner), 50%–10 dB (knife), and 70%–0 dB (pliers), respectively.

Fig. 6 shows that MF-based reconstructions suffer from severe noise, resulting in distorted structures, particularly under low SR. Meanwhile, the SFH-ADMM and FA-ADMM approaches yield comparable visual quality, both restoring the overall shape and structure of the targets. However, as mentioned before, FA-ADMM recorded a significantly faster reconstruction time due to its simplified update strategy and adaptive penalty tuning.

Further improvements are observed in the case of the DL-augmented variants. FA-CNN-ADMM provides the most structurally faithful reconstructions with crisp object contours and reduced background clutter, while FA-AE-ADMM generates smoother and more natural representations, particularly effective in suppressing undesired noise. Crucially, for sparse inputs, such as the gun at 10% SR and the spanner at 30% SR, the FA-DL-ADMM model shows strong performance, successfully restoring object shapes and features that are heavily degraded in other approaches. These results underscore the robustness and generalizability of the proposed framework in handling extreme subsampling and varied imaging geometries.

3. Experimental Results

Unlike simulation scenarios, where ground truth images are available, real-world experiments often lack reference data for direct comparison. Therefore, quantitative evaluation of all the methods was conducted using image entropy (ENT) and the target-to-background ratio (TBR). Notably, ENT quantifies the randomness or uncertainty in pixel intensity values within an image, where a lower ENT value corresponds to enhanced image-focusing performance. It is mathematically expressed as follows:

(42) ENT(Y)=-1EY∑iNw∑jNh(∣Yi,j∣2ln(∣Yi,j∣2))+ln (EY)

where EY=ΣiNwΣjNh∣Yi,j∣2. Meanwhile, TBR assesses the relative intensity of contrast between a region of interest (i.e., the target) and its surrounding background. A higher TBR value indicates greater contrast, which is typically associated with improved target-detection capability. The TBR metric is defined as follows:

(43) TBR(Y)=10log10‖Y∈ΩT‖22‖Y∈Ωc‖22

where ΩT and Ωc denote the target area and clutter region, respectively.

For all the experiments conducted in this study, the radar and imaging configurations were kept consistent with those used in the simulation setup. Specifically, we employed Texas Instruments’ AWR1443BOOST, which is a commercial FMCW radar operating in the 77–81 GHz frequency range. The radar parameters were configured to match the simulated values listed in Table 1. The radar module was connected to a DCA1000MVE data capture card transmitting raw IF signal data to a host PC via an Ethernet interface, as illustrated in Fig. 7(a).

Fig. 7

Measurement system: (a) radar system mounted on (b) 2D motion platform.

The radar system and data acquisition hardware were mounted on a 2D motorized motion platform to perform mechanical scanning across the aperture plane. The scanning area was 200 mm × 200 mm, with a spatial sampling interval of 2 mm, resulting in a total of 100 × 100 spatial samples. With 256 range bins per location, this configuration yielded a 3D data cube of dimensions 100 × 100 × 256 for each SUT. The target objects were positioned in front of the radar scanning plane at distances ranging from 200 mm to 500 mm, as depicted in Fig. 7(b).

Measurements were conducted for a knife with dimensions of 24.3 cm × 3.7 cm × 2 cm (length × width × thickness) that was placed at a distance of 290 mm, and on a pair of pliers, whose size was 14.7 cm × 94 cm × 1.4 cm, that was positioned at a distance of 225 mm. As depicted in Fig. 8, the MF-based method produced images heavily contaminated with clutter noise at all SR levels, resulting in significantly degraded target recognition as the SR decreased. Meanwhile, Fig. 8(a) shows that the ADMM-based methods provide satisfactory 3D reconstructions with clear target shapes at 100% SR. However, their performance deteriorated considerably at 50% SR, as evident from Fig. 8(b). In particular, SFH-ADMM and FA-ADMM suffer from severe distortions and information loss under reduced sampling conditions, highlighting the limitations of traditional iterative approaches in handling sparse data. In contrast, the proposed FA-DL-ADMM methods equipped with DL-based denoising models accomplish superior target recovery and noise suppression across all SR levels. Even at 50% SR, both the FA-CNN-ADMM and FA-AE-ADMM methods effectively restore target details while maintaining high image quality with minimal noise. This notable improvement underscores the effectiveness of deep learning integration in addressing undersampling challenges and highlights the contribution of FA-DL-ADMM methods in enhancing 3D SAR image reconstruction under sparse data conditions.

Fig. 8

Experimental 3D image results at SR of (a) 100% and (b) 50%.

Additionally, the ENT and TBR metrics summarized in Table 5 allow for a quantitative comparison of image quality across the employed methods at various SR levels. The findings reveal the limitations of the conventional MF-based method, which suffers from high ENT and low TBR, resulting in poor noise suppression and degraded target details. In contrast, the non-DL-based ADMM methods exhibit significantly improved performance as a result of the iterative optimization process, which effectively suppresses noise while retaining essential structural information. The FA-DL-ADMM models achieved the most favorable results, outperforming all other approaches across all SRs. This performance advantage is attributable to the synergistic combination of ADMM optimization with DL-based denoisers. Moreover, between the two FA-DL-ADMM models, FA-CNN-ADMM is observed to be more effective under low-SR conditions, whereas FA-AE-ADMM excels in higher-SR scenarios, delivering superior image quality with well-defined target details. Overall, these results emphasize the practical effectiveness and innovative contribution of the proposed FA-DL-ADMM framework.

Comparison of experimental results

While this study primarily focuses on metallic targets, such as hand tools, we acknowledge the need to evaluate model robustness across diverse material types encountered in practical environments, such as plastic, fabric, and wood. Compared to metal, these materials typically exhibit lower reflectivity and different backscatter characteristics at 77–81 GHz MMW frequencies, which makes their detection and reconstruction more challenging. Future research must incorporate dielectric material modeling into simulations and account for multi-material test objects in experiments to assess the performance of the proposed FA-DL-ADMM framework in mixed-material scenarios.

V. Conclusion

This paper proposes an FA-ADMM framework for short-range 3D SAR imaging, achieving substantial improvements in both computational efficiency and reconstruction fidelity. The framework leverages a modified augmented Lagrangian formulation that simplifies the iterative update process, together with an adaptive parameter-tuning mechanism that dynamically adjusts penalty and relaxation parameters to enable enhanced convergence stability. Furthermore, when integrated with DL–based denoisers, the proposed FA-DL-ADMM variants achieved improved noise suppression and structural preservation, resulting in clearer and more accurate reconstructions. The advantages of the proposed method lie in its computational efficiency, adaptive convergence control, and flexibility to incorporate data-driven priors. Nevertheless, its performance may be influenced by the representativeness of the training dataset used for the DL-based denoisers.

Future research should focus on improving generalization by diversifying the training data to include non-metallic and irregularly shaped objects, in addition to applying domain adaptation strategies and exploring transformer-based denoisers to better capture global features. Nonetheless, owing to its efficiency and adaptability, the proposed framework has strong potential for real-world short-range SAR applications, such as concealed object detection, industrial inspection, and autonomous navigation.

Notes

This work was supported in part by the Priority Research Center Program through the National Research Foundation under Grant RS-2019-NR040074, and in part by a Korea Basic Science Institute (National Research Facilities and Equipment Center) Grant funded by the Ministry of Education under Grant RS-2022-NF000835.

References

1. Batra A., Barowski J., Damyanov D., Wiemeler M., Rolfes I., Schultze T., Balzer J. C., Gohringer D., Kaiser T.. Short-range SAR imaging from GHz to THz waves. IEEE Journal of Microwaves 1(2):574–585. 2021;https://doi.org/10.1109/JMW.2021.3063343.
2. Kang Y. G., Jung D. H., Shin G. H., Kim C. K., Park S. O.. A study on the feasibility of stop-and-go approximation in FMCW SAR. Journal of Electromagnetic Engineering and Science 22(3):210–217. 2022;https://doi.org/10.26866/jees.2022.3.r.79.
3. Cho B., Hong S., Yoon K.. Omega-K algorithm using plane wave approximation for forward-looking imaging radar. Journal of Electromagnetic Engineering and Science 22(5):571–575. 2022;https://doi.org/10.26866/jees.2022.5.r.124.
4. Schorlemer J., Schulz C., Pohl N., Rolfes I., Barowski J.. Compensation of sensor movements in short-range FMCW synthetic aperture radar algorithms. IEEE Transactions on Microwave Theory and Techniques 69(11):5145–5159. 2021;https://doi.org/10.1109/TMTT.2021.3108399.
5. Smith J. W., Torlak M.. Efficient 3-D near-field MIMO-SAR imaging for irregular scanning geometries. IEEE Access 10:10283–10294. 2022;https://doi.org/10.1109/ACCESS.2022.3145370.
6. Tan K.. A fast omega-K algorithm for near-field 3-D imaging of MIMO synthetic aperture radar data. IEEE Geoscience and Remote Sensing Letters 18(8):1431–1435. 2021;https://doi.org/10.1109/LGRS.2020.3003504.
7. Wang J., Aubry P., Yarovoy A.. 3-D short-range imaging with irregular MIMO arrays using NUFFT-based range migration algorithm. IEEE Transactions on Geoscience and Remote Sensing 58(7):4730–4742. 2020;https://doi.org/10.1109/TGRS.2020.2966368.
8. Song J., Liu Q. H., Torrione P., Collins L.. Two-dimensional and three-dimensional NUFFT migration method for landmine detection using ground-penetrating radar. IEEE Transactions on Geoscience and Remote Sensing 44(6):1462–1469. 2006;https://doi.org/10.1109/TGRS.2006.870412.
9. Xu G., Zhang B., Yu H., Chen J., Xing M., Hong W.. Sparse synthetic aperture radar imaging from compressed sensing and machine learning: theories, applications, and trends. IEEE Geoscience and Remote Sensing Magazine 10(4):32–69. 2022;https://doi.org/10.1109/MGRS.2022.3218801.
10. Song S., Dai Y., Sun S., Jin T.. Efficient image reconstruction methods based on structured sparsity for shortrange radar. IEEE Transactions on Geoscience and Remote Sensing 62article no. 5212615. 2024;https://doi.org/10.1109/TGRS.2024.3404626.
11. Wahid A., Shah J. A., Khan A. U., Ahmed M., Razali H.. Multi-layer basis pursuit for compressed sensing MR image reconstruction. IEEE Access 8:186222–186232. 2020;https://doi.org/10.1109/ACCESS.2020.3028877.
12. Beck A., Teboulle M.. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences 2(1):183–202. 2009;https://doi.org/10.1137/080716542.
13. Neal P., Eric C., Borja P., Jonathan E.. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning 3(1):1–122. 2011;https://doi.org/10.1561/2200000016.
14. Li P., Ding Z., Zhang T., Wei Y., Gao Y.. Integrated detection and imaging algorithm for radar sparse targets via CFAR-ADMM. IEEE Transactions on Geoscience and Remote Sensing 61article no. 5204015. 2023;https://doi.org/10.1109/TGRS.2023.3251732.
15. Qin X., Deng B., Wang H., Zeng Y., Chen X., Xiao H.. An equalized ADMM-based high-resolution autofocusing imaging algorithm for THz-SAR ground moving targets. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17:8450–8460. 2024;https://doi.org/10.1109/JSTARS.2024.3386583.
16. Wang M., Wei S., Zhou Z., Shi J., Zhang X.. Efficient ADMM framework based on functional measurement model for mmW 3-D SAR imaging. IEEE Transactions on Geoscience and Remote Sensing 60article no. 5226417. 2022;https://doi.org/10.1109/TGRS.2022.3165541.
17. Amineh R. K., Nikolova N. K., Ravan M.. Real-Time Three-Dimensional Imaging of Dielectric Bodies Using Microwave/Millimeter Wave Holography New York, NY: John Wiley & Sons Inc; 2019. https://doi.org/10.1002/9781119538875.
18. Seo S. M., Choi Y., Lim H., Chae D. Y.. A measurement image translation-automatic target recognition technique based on CycleGAN with SAR simulation DB. Journal of Electromagnetic Engineering and Science 22(3):326–334. 2022;https://doi.org/10.26866/jees.2022.3.r.94.
19. Seo S. M., Choi Y., Lim H., Park J. H.. A deep learning-based compact weighted binary classification technique to discriminate between targets and clutter in SAR images. Journal of Electromagnetic Engineering and Science 22(4):412–418. 2022;https://doi.org/10.26866/jees.2022.4.r.103.
20. Oral O., Oktem F. S.. Plug-and-play regularization on magnitude with deep priors for 3D near-field MIMO imaging. IEEE Transactions on Computational Imaging 10:762–773. 2024;https://doi.org/10.1109/TCI.2024.3396388.
21. Alver M. B., Saleem A., Cetin M.. Plug-and-play synthetic aperture radar image formation using deep priors. IEEE Transactions on Computational Imaging 7:43–57. 2021;https://doi.org/10.1109/TCI.2020.3047473.
22. Wang Z., Song C., Wang B., Qiu X., Xiang M.. Langevin sampling plug-and-play synthetic aperture radar imaging algorithm. IEEE Transactions on Geoscience and Remote Sensing 62article no. 5222415. 2024;https://doi.org/10.1109/TGRS.2024.3457819.
23. Xu Z., Figueiredo M. A., Yuan X., Studer C., Goldstein T.. Adaptive relaxed ADMM: convergence theory and practical implementation. In : Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA; 2017; p. 7234–7243. https://doi.org/10.1109/CVPR.2017.765.
24. Eckstein J., Bertsekas D. P.. On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators. Mathematical Programming 55(1):293–318. 1992;https://doi.org/10.1007/BF01581204.
25. Weyl H.. Ausbreitung elektromagnetischer Wellen über einem ebenen Leiter. Annalen der Physik 365(21):481–500. 1919;https://doi.org/10.1002/andp.19193652104.

Biography

The-Hien Pham, https://orcid.org/0000-0003-3800-0703 received a B.E. degree in communication and networking from Ho Chi Minh City University of Transport, Vietnam, in 2012, and an M.E. degree in information and communication engineering from Kongju National University, South Korea, in 2022. He is currently pursuing a Ph.D. in information and communication engineering, at Kongju National University, Korea. His major research interests are non-destructive testing, communication and networking, wireless communication systems, synthetic aperture radar imaging, machine learning, and optimization methods.

Ic-Pyo Hong, https://orcid.org/0000-0003-1875-5420 received his B.S., M.S., and Ph.D. degrees in electronics engineering from Yonsei University, Seoul, South Korea, in 1994, 1996, and 2000, respectively. From 2000 to 2003, he worked with the Information and Communication Division, Samsung Electronics Company, Suwon, South Korea, where he was a senior engineer involved in CDMA mobile research. He was a visiting scholar at Texas A&M University, College Station, TX, USA, in 2006, and at Syracuse University, Syracuse, NY, USA, in 2012. Since 2003, he has been with the Department of Smart Information and Technology Engineering, Kongju National University, Cheonan, South Korea, where he is currently a professor. His research interests include numerical techniques in electromagnetics, and the application of periodic electromagnetic structures in wireless communication.

Article information Continued

Fig. 1

The measurement system and coordinate definitions.

Fig. 2

Structure of the FA-ADMM method.

Fig. 3

Diagram of the proposed framework: (a) the FA-DL-ADMM iterative procedure and (b) detailed structures of the three layers for solving the AL function.

Fig. 4

Architectures of denoising models: (a) CNN and (b) autoencoder.

Fig. 5

Boxplots of (a) PSNR and (b) SSIM obtained by conducting over 25 Monte Carlo runs for different sampling rates using the two FA-DL-ADMM methods

Fig. 6

Reconstructed 3-D SAR images obtained using different methods, with SRs–SNRs being 10%–30 dB (gun), 30%–20 dB (spanner), 50%–10 dB (knife), and 70%–0 dB (pliers), respectively.

Fig. 7

Measurement system: (a) radar system mounted on (b) 2D motion platform.

Fig. 8

Experimental 3D image results at SR of (a) 100% and (b) 50%.

Table 1

Configuration of the imaging scheme

Parameter Value
Radar
 Frequency 77–81 GHz
 Bandwidth, B 4 GHz
 Slope frequency 63.343 MHz/μs
 Sampling frequency, fs 9.121 × 106 samples/s
Imaging
 Target range, z 200–500 mm
 Synthetic aperture size (Dx × Dy) 200 × 200 mm
 Spatial sampling intervals (dx, dy) 2 mm

Table 2

Simulated object quantities

# Category Quantity Number of simulated pairs
1 fork 10 400
2 gun 16 640
3 hammer 20 800
4 key 10 400
5 knife 30 1,200
6 pliers 15 600
7 scissors 10 400
8 screwdriver 30 1,200
9 spanner 19 760
10 spatula 5 200
11 spoon 10 400
12 whisk 5 200
Total 180 7,200

Table 3

Numerical evaluation of simulations under varied SRs and SNRs

Method Sampling rate (%), SNR 3 0dB Average SNR in dB sampling rate = 50% Time (s)


10% 30% 50% 70% 90% 0 dB 5 dB 10 dB 15 dB 20 dB
MF-based 20.53
 PSNR 13.95 16.40 17.51 18.39 18.62 16.97 15.47 16.41 17.11 17.45 17.49
 NMSE 1.86 −0.65 −2.45 −2.59 −3.05 −1.38 0.35 −1.17 −1.89 −2.22 −2.30
 SSIM 0.05 0.07 0.09 0.11 0.19 0.10 0.05 0.07 0.07 0.08 0.09
SFH-ADMM 4.82
 PSNR 16.41 17.72 18.38 18.73 18.94 18.04 17.18 17.87 18.31 18.34 18.36
 NMSE −0.60 −1.96 −2.59 −2.94 −3.13 −2.24 −1.36 −2.11 −2.42 −2.54 −2.55
 SSIM 0.80 0.85 0.86 0.87 0.87 0.85 0.84 0.86 0.86 0.86 0.86
FA-ADMM 1.73
 PSNR 16.36 17.66 18.36 18.77 18.93 18.02 17.21 17.87 18.34 18.35 18.36
 NMSE −0.55 −1.91 −2.58 −2.97 −3.11 −2.22 −1.39 −2.11 −2.45 −2.57 −2.57
 SSIM 0.80 0.85 0.86 0.87 0.87 0.85 0.84 0.86 0.86 0.86 0.86
FA-CNN-ADMM 0.54
 PSNR 20.16 24.57 27.06 29.51 33.14 26.89 23.26 25.04 26.56 26.85 27.05
 NMSE −4.35 −8.81 −11.28 −13.71 −17.32 −11.10 −7.44 −9.28 −10.67 −11.05 −11.23
 SSIM 0.89 0.93 0.95 0.97 0.98 0.94 0.92 0.94 0.95 0.95 0.95
FA-AE-ADMM 0.68
 PSNR 21.09 25.76 29.05 31.69 37.65 29.05 23.77 26.57 27.79 28.61 28.89
 NMSE −5.84 −10.53 −13.82 −16.47 −22.42 −13.82 −8.54 −11.33 −12.57 −13.38 −13.66
 SSIM 0.92 0.96 0.97 0.98 0.99 0.96 0.94 0.96 0.97 0.97 0.97

The bold font indicates the best performance in each test.

Table 4

Computational duration of ADMM-based methods

Method Average number of iterations Average per iteration time (s) Average runtime (s)
SFH-ADMM 22.99 0.21 4.82
FA-ADMM 6.79 0.26 1.73
FA-CNN-ADMM 4.02 0.14 0.54
FA-AE-ADMM 3.86 0.18 0.68

The bold font indicates the best performance in each test.

Table 5

Comparison of experimental results

Method Sampling rate

25% 50% 75% 100%
MF-based
 ENT 8.67 8.43 8.10 7.44
 TBR 4.89 5.42 8.95 9.79
SFH-ADMM
 ENT 7.13 5.51 5.44 5.38
 TBR 14.26 15.98 18.43 22.20
FA-ADMM
 ENT 7.02 5.47 5.44 5.38
 TBR 17.44 17.58 18.45 22.33
FA-CNN-ADMM
 ENT 6.23 5.42 5.35 5.21
 TBR 19.71 23.47 27.60 27.67
FA-AE-ADMM
 ENT 6.49 5.40 5.30 5.18
 TBR 19.06 23.93 27.53 28.23

Bold font indicates the best result for each metric at each sampling rate.