Short-range SAR imaging with fast adaptive plug-and-play ADMM-based approach stationary phase [10], a spherical wave can be revealed as a superposition of a plane wave: ZZ T.H. Pham and I.P. Hong ej(kx (x−x ej2kR0 ≈ Fast Fourier transform (FFT) is typically applied to the majority of recent short-range synthetic aperture radar (SAR) imaging algorithms to reconstruct the target image, whose quality is particularly sensitive to the high displayed dynamic range. To tackle this issue, we proposed a SAR imaging scheme that employs a fast adaptive plug-and-play alternating direction method of multiples (FA-PnP-ADMM) framework adopting the state-of-the-art ADMM-based image-solving model. This approach retains good forcefulness denoising performance at a low signal-to-noise ratio (SNR) and low sampling ratio (SR) conditions while simultaneously significantly reducing the computing load. The 2-dimensional (2-D) imaging simulation results illustrate the superiority of the proposed scheme compared to those of the existing FFT-based algorithm. ′ )+ky (y−y ′ )+kz z0 ) dkx dky , where kx , ky and kz are the components of k in the Cartesian coordinate system which defer to the following relation: kz = q 4k2 − kx2 − ky2 , kx2 + ky2 ≤ 4k2 . (3) By submitting (2) to (1), the echo signal can be presented as: ZZ ZZ s(x, y, k) = σ(x′ , y ′ )e−j(kx x ′ +ky y ′ ) dx′ dy ′ × ejkz z0 ej(kx x+ky y) dkx dky . Introduction: As an important remote sensing method, synthetic aperture radar (SAR) imaging makes use of electromagnetic (EM) waves’ capacity to create images free from the interference of opaque substances such as light, clouds, fog, walls, etc. The short-range millimeter-wave (MMW) technology has drawn a lot of interest lately since it offers better resolution in both azimuth and range directions than other lower-frequency bands. As a result, this technology has been used as a useful tool for several close-range civilian applications, such as medical diagnostics [1], nondestructive testing (NDT) [2], and security surveillance [3]. The image reconstruction algorithms need to take into account the important factors of image reconstruction, namely "speed" and "quality" for these types of applications. The back-projection algorithm (BPA), which can be applied to any shape or layout, is a spatial-domain processing method that is widely used for reconstructing SAR images [4]. However, the computational load of this technique is substantial since it needs to scan over every point in the reconstruction area to form an image. The range migration algorithm (RMA), which is a representation of the fast Fourier transform (FFT), is another method for recreating SAR pictures. Nevertheless, uniform sampling during measurement is necessary for the FFT-based approaches [5]. Although the non-uniform FFT (NUFFT) approach can overcome the aforementioned problem [6], it is necessary to perform a large amount of interpolation, resulting in increasing the computational load. Furthermore, the FFT operation for inverting the target image has a considerable influence on the image quality in a highly displayed dynamic range. The alternating direction method of multipliers (ADMM), an algorithm based on compressed sensing (CS) that exhibits good convergence performance, has demonstrated its significant promise in SAR image reconstruction under sparse conditions [7]. However, this algorithm’s major drawback is that it requires massive amounts of matrix multiplication and inversion, which significantly increases the computational burden [8]. An FFT-based function was suggested as an efficient alternative to high-dimensional matrix operations in order to tackle this problem and drastically reduce the computation duration [9]. On the basis of this most advanced fast algorithm, we proposed a fast adaptive plug-and-play ADMM (FA-PnP-ADMM)-based imaging scheme to accomplish two goals: Initially, determine adaptive regularization coefficients over each iteration to accelerate the image reconstruction process. Secondly, robustly forming high-quality images even under situations with low signal-to-noise ratio (SNR). Signal model and conventional FFT-based imaging: Assume that the target is situated at a distance z0 on a target plane (x′ , y ′ , z0 ) and being scanned by a frequency-modulated continuous wave (FMCW) radar on 2D aperture plane (x, y, 0). The reflected signal s ∈ CNx ×Ny ×Nk can be defined in the frequency domain as: ZZ s(x, y, k) = σ(x′ , y ′ )e(j2kR0 ) dx′ dy ′ , (1) where Nx and Ny are the sampling points over aperture plane in x- and y -directions, respectively; Nk is number of sampling points in range direction; σ ∈ CNw ×Nh is the target reflectivity with Nw and x′ - and y ′ -directions; R0 = p Nh are the number of pixel in 2π 2 ′ 2 ′ 2 (x − x) + (y − y) + z0 and k = f is the wave number of the beat b frequency fb at the corresponding distance R0 . By applying the method of ELECTRONICS LETTERS 22nd November 2023 (2) (4) It is interesting to note that the term inside the square brackets can be considered as the 2-D Fourier transform of the reflectivity σ which is denoted as F2D [σ(x′ , y ′ )]. Hence, s(x, y, k) can be rewritten as: ZZ s(x, y, k) = F2D [σ(x′ , y ′ )] × ejkz z0 ej(kx x+ky y) dkx dky −1 = F2D F2D σ(x′ , y ′ )ejkz z0 (5) , −1 where F2D [·] represents the inverse 2-D Fourier transform operator. Consequently, we can infer the target complex reflectivity as: −1 σ(x′ , y ′ ) = F2D F2D [s(x, y, k)] × e−jkz z0 . (6) Related work: By denoting the s = vec(s) ∈ CNx Ny ×1 and σ = vec(σ ) ∈ CNw Nh ×1 as the vectorial forms of the echo data and scattering coefficients, where the vec(·) is the vectorizing operation, (1) can be redefined as: s = Hσ, (7) where H ∈ CNx Ny ×Nw Nh is the sensing matrix with element being e−j2kz0 R0 . It can be seen that the imaging process is correspondingly the reconstruction of σ from the echo data s. The sparsity allows (7) to be modeled as linear least-square problem including ℓ1 regularization term as below: σ̂ = arg min σ 1 ∥s − Hσ∥22 + λ∥σ∥1 , 2 (8) where ∥·∥2 and ∥·∥1 are the ℓ2 (Euclidean) and ℓ1 norms indicating the data fidelity and the sparsity of targeting signal, respectively, λ is introduced as a regularization coefficient to control the balance between two terms. Using ADMM theory [11], an extra auxiliary variable (b) is added to split the optimization problem into two independent penalty terms: σ̂ = arg min σ 1 ∥s − Hσ∥22 + λ∥b∥1 2 s.t. σ = b. (9) The corresponding augmented Lagrangian (AL) function [11] is expressed as: 1 2 Lρ (σ, b, u) = ∥s − Hσ∥22 + λ∥b∥1 + uH (σ − b) + ρ ∥σ − b∥22 , 2 (10) where u stands for the Lagrange multiplier and its conjugate transpose uH ; ρ is the penalty parameter. As AL function offers all separable variables in Lρ (σ, b, u), (8) can be solved by decomposing (10) to three iterative subproblems [11]. However, according to this approach, the large-scale matrix inversion for calculating σ brings about a high computational burden, and thus, leads to slow convergence and consumes a large amount of hardware resources. Wang et al. [9], introduced the fast ADMM (FADMM) to solve the aforementioned problem by utilizing the relation between the matrix-representing echo s and reflection coefficient σ . Let’s Vol. 00 No. 00 (t) define: S = s(x, y, kz0 ) and X = σ(x′ , y ′ ) by replacing ∆h(t) and ∆Û with ∆g (t) = B(t) − B(t0 ) and ∆U(t) = U(t) − U(t0 ) , respectively. The parameters ρ and γ are then updated based on the following rules: p (t−1) (t−1) α(t−1) β (t−1) if αcor > ϵcor and βcor > ϵcor (t−1) (t−1) α(t−1) if αcor > ϵcor and βcor ≤ ϵcor ρ(t) = (22) (t−1) (t−1) (t−1) β if αcor ≤ ϵcor and βcor > ϵcor (t−1) ρ otherwise, (11) (5) and (6) can be correspondingly reformulated as: −1 S = F {X} ≜ F2D [F2D [X] ⊗ Φ] , −1 X = F† {S} ≜ F2D F2D [P{S}] ⊗ Φ̄ , (12) where F{·} and F† {·} represent the measurement and imaging processing, respectively, Φ ∈ CNw ×Nh is the phase matrix and its conjugate Φ̄ with element being e−jkz z0 , ⊗ denotes the Hadamard product (element-wise multiplication), and P{·} ∈ CNw ×Nh (Nx < Nw , Ny < Nh ) defines the zero-padding operation as follow: ( Nw −Nx x ≤ i ≤ Nw +N S , if 2 2 ij Nh −Ny Nw +Ny [P{S}]ij = (13) ≤j≤ 2 2 0, others. γ (t) = √ α(t−1) β (t−1) 1 + 2α(t−1) +β (t−1) 1.9 1.1 1.5 (t−1) if αcor (t−1) > ϵcor (t−1) βcor (t−1) βcor ≤ ϵcor > ϵcor and βcor (t−1) αcor (t−1) αcor if > ϵcor and if ≤ ϵcor and otherwise, > ϵcor (23) (t) where ϵcor is a safeguarding threshold for the curvature estimates; αcor (t) and βcor indicate the correlation between relaxed ADMM and relaxed DRS [13] which are expressed as: E D D E (t) ∆h(t) , ∆Û ∆g (t) , ∆U(t) (t) (t) αcor = (24) and βcor = (t) ∥∆g (t) ∥2 ∥∆U(t) ∥2 ∥∆h(t) ∥2 ∥∆Û ∥2 The image reconstruction problem is now can be redefined based on (12) as: 1 2 X̂ = arg min ∥S − F {X}∥22 + λ∥B∥1 X s.t. X = B, (14) and the AL function is accordingly formulated as: 1 2 Lρ (X, B, U) = ∥S − F {X}∥22 + λ∥B∥1 ρ + Tr UH (X − B) + ∥X − B∥22 , 2 The stepsize updating process is performed every 2 iterations and the safeguarding threshold is fixed at ϵcor = 0.2 (15) Simulation results: In this section, we validate the effectiveness of the proposed scheme by simulating SAR images of objects located at a distance (z0 ) of 200 mm from the aperture plane. Two other algorithms, namely, RMA and FADMM-ST [9] are used to illustrate the performance of the proposed scheme employing FA-ADMM-BM3D and FA-ADMMST approaches. The simulated FMCW signal works at a start frequency of 77 GHz, with a bandwidth of 4 GHz and a slope frequency of 63.343 MHz per second. An echo data cube size of 100 × 100 × 512 is collected. Then, we randomly select a reduced set of elements in this data cube with a down-sampling ratio (SR) of 50% for reconstructing a SAR image. where B and U are the matrix forms of the auxiliary variable (b) and the dual variable (u), respectively. The Tr{·} stands for the matrix trace calculation. The problem (14) can be solved by using an iterative method to update variables in (15) over finite iterations [11]: 1 (t−1) † (t) (t−1) X(t) = F { S } + ρ B − U (16) (t) 1+ρ (t) = γ (t) X(t) + (1 − γ (t) )B(t−1) X̃ B(t) = arg min λ∥B∥1 + B ρ (t) ∥B − (X̃ + U(t−1) )∥22 2 (t) U(t) = U(t−1) + ρ(t) (X̃ − B(t) ) (17) (18) (19) Here, we apply the relaxed ADMM (as in (17)) to improve the convergence [12], where γ ∈ (0, 2) is the relaxation parameter. Relaxed ADMM coincides with the original non-relaxed version if γ = 1. It is noteworthy that the sub-problem (18) is equivalent to a denoiser, and the idea of PnP ADMM [16] is to plug in a powerful denoising algorithm in place of (18). In this letter, we employ two denoisers, namely: the shrinkage threshold (ST) and block-matching and three-dimensional filtering (BM3D). FA-PnP-ADMM-based approach: It is important to note that the convergence performance of the ADMM iteration (16)-(19) strongly depends on the choices of ρ and γ . Hence, in this letter, we studied a fast adaptive penalty method to automatically tune these parameters over each iteration, denoted as ρ(t) and γ (t) . By exploiting the relationship between relaxed ADMM and the relaxed Douglas-Rachford Splitting (DRS) [13, 14], the adaptive stepsizes of ρ(t) and γ (t) can be derived. First, two local curvature estimates were defined as α(t) > 0 and β (t) > 0 [15]: ( (t) (t) (t) αM G if 2αM G > αSD (t) α = (20) (t) (t) αSD − αM G /2 otherwise, Fig. 1 The convergence performance of the FADMM and the proposed method First, we evaluate the computational performance of the proposed scheme and the conventional FADMM-ST by deriving the dual residual ∥X − B∥2 over the iterative process of ADMM [11]. In all ADMM-based approaches, the tuning parameters of ρ, γ , and λ are initiated as 1, 1.7, and 0.01, respectively [9]. Due to the adaptation of parameters (ρ, γ ) of the proposed method, the problem can be quickly solved. Fig. 1 illustrates that our method needs considerably fewer iterations to reach the optimum. where D E D E (t) (t) (t ) ∆h(t) ,∆Û X(t) −X(t0 ) ,Û −Û 0 (t) αM G = ⟨∆h(t) ,∆h(t) ⟩ = X(t) −X(t0 ) ,X(t) −X(t0 ) D E D⟨ E⟩ (t) (t) (t) (t ) (t) (t ) ∆Û ,∆Û Û −Û 0 ,Û −Û 0 (t) E = D E αSD = D (t) (t) (t0 ) (t0 ) (t) (t) ∆h ,∆Û X −X ,Û Table 1: Numerical comparison of different imaging methods Methods (21) RMA −Û FADMM-ST (t) where Û = U(t−1) + ρ(X(t) − B(t) ) is the intermediate dual variable; MG and SD stand for minimum gradient and steepest descent, respectively; ⟨a, b⟩ is the inner product of a and b. Similarly, β (t) can be estimated FA-ADMM -BM3D FA-ADMM -ST 2 Metrics PSNR (dB) SSIM PSNR (dB) SSIM PSNR (dB) SSIM PSNR (dB) SSIM 0 17.180 0.856 19.726 0.861 19.725 0.861 19.726 0.861 SR = 0.5, SNR (dB) 10 20 19.424 19.709 0.938 0.947 29.302 36.661 0.973 0.986 29.301 36.660 0.972 0.986 29.302 36.661 0.973 0.986 30 20.474 0.951 39.714 0.987 39.713 0.987 39.714 0.987 Average time (s) 5.189 0.531 182.88 0.295 improved SSIM and a considerable increase in PSNR. Additionally, at all SNR levels, FA-ADMM-BM3D performance is slightly worse than the other ADMM-based methods while the FADMM-ST and FA-ADMMST algorithms archive similar good SSIM and PSNR. Furthermore, the FA-ADMM-BM3D consumes a large amount of time due to the blockmatching finding nature of BM3D which does not relate to the procedure of finding adaptive parameters. In contrast, the FA-ADMM-ST approach has less computational complexity than its counterparts, as a result, it appreciably consumes less necessary duration to form high-quality images. Conclusion: This letter proposed the FA-PnP-ADMM-based SAR imaging scheme to archive high-quality image reconstruction while considerably reducing computational burden. The performance of the proposed method is superior to the original fast ADMM as it consumes fewer iterations to find the optimal solution while retains good image quality at very low SNR conditions. Moreover, we also validated the performance of the PnP scheme by integrating different denoisers into the ADMM problem-solving process. In future work, a novel artificial intelligence (AI)-based denoising technique will be applied to our imaging system which is predicted to improve denoising capability. Acknowledgment: This work has been supported by The IET J. Smith and A. N. Other (The IET, Stevenage, UK) E-mail: jsmith@theiet.org References 1 C. Das, M. Z. Chowdhury and Y. M. Jang,: ‘A Novel Miniaturized mmWave Antenna Sensor for Breast Tumor Detection and 5G Communication’, IEEE Access, 2022, 10, pp. 114856-114868 Anderson, P.: ‘A poor man’s derivation of scaling laws for the Kondo problem’, J. Phys. C., 1960, 3, p. 2436 2 M. A. Abou-Khousa, M. S. U. Rahman, K. M. Donnell and M. T. A. Qaseer,: ‘Detection of Surface Cracks in Metals Using Microwave and Millimeter-Wave Nondestructive Testing Techniques—A Review’, IEEE Trans. Instrum. Meas., 2023, 72, pp. 1-18 3 B. Su and M. Yuan,: ‘Object Recognition for Millimeter Wave MIMO-SAR Images Based on High-Resolution Feature Recursive Alignment Fusion Network’, IEEE Sens. J., 2023, 23, (14), pp. 16413-16427 4 D. Feng, D. An and X. Huang,: ‘An Extended Fast Factorized Back Projection Algorithm for Missile-Borne Bistatic Forward-Looking SAR Imaging’, IEEE Trans. Aerosp. Electron. Syst., 2018, 54, (6), pp. 2724-2734 5 M. E. Yanik, D. Wang and M. Torlak,: ‘Development and Demonstration of MIMO-SAR mmWave Imaging Testbeds’, IEEE Access, 2020, 8, pp. 126019-126038 6 J. Wang, P. Aubry and A. Yarovoy,: ‘3-D short-range imaging with irregular MIMO arrays using NUFFT-based range migration algorithm’, IEEE Trans. Geosci. Remote Sens., 2020, 58, (7), pp. 4730-4742 7 Moradikia, M., Samadi, S., and Cetin, M.: ‘Joint SAR imaging and multifeature decomposition from 2-D under-sampled data via lowrankness plus sparsity priors", IEEE Trans. Comput. Imaging, 2018, 5, (1), pp. 1–16 8 B. Zhao, L. Huang and W. Sun,: ‘Target Reconstruction in Deceptively Jammed SAR via ADMM’, IEEE Sens. J., 2019, 19, (11), pp. 4331-4339 9 M. Wang, S. Wei, Z. Zhou, J. Shi and X. Zhang,: ‘Efficient ADMM Framework Based on Functional Measurement Model for mmW 3-D SAR Imaging’, IEEE Trans. Geosci. Remote Sens., 2022, 60, pp. 1-17 10 H. Weyl: ‘Ausbreitung elektromagnetischer Wellen über einem ebenen Leiter’, Ann. Phys, 1919, 365, (21), pp. 481–500 11 S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein,: ‘Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers’, 2011 12 J. Eckstein and M. C. Ferris,: ‘Operator-splitting methods for monotone affine variational inequalities, with a parallel application to optimal control’, INFORMS J. Comput., 1998, 10, pp. 218–235 13 Z. Xu, M. A. T. Figueiredo, X. Yuan, C. Studer and T. Goldstein,: ‘Adaptive Relaxed ADMM: Convergence Theory and Practical Implementation’, 2017 Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern. Recognit., 2017, pp. 7234-7243 14 J. Eckstein and D. Bertsekas,: ‘On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators’, Math. Program., 1992, 55, (1-3), pp. 293–318 15 Xu, Zheng, Mario Figueiredo, and Tom Goldstein. ‘Adaptive ADMM with spectral penalty parameter selection’, Artif. Intell. and Stat. 2017, PMLR, pp. 718-727 16 R. Hou, F. Li and G. Zhang,: ‘Truncated Residual Based Plug-andPlay ADMM Algorithm for MRI Reconstruction’, IEEE Trans. Comput. Imaging, 2022, 8, pp. 96-108 Fig. 2 SAR images of a target reconstructed at four SNR levels by using RMA (first row) and FA-ADMM-ST (second row) Fig. 2 demonstrates the SAR images of a scissor exploited from imaging approaches over four SNR levels, as correspondingly shown in four columns. The first and second rows in each column of Fig. 2 are obtained by using the RMA, and FA-ADMM-ST algorithms, respectively. It is important to mention that the ADMM-based methods have the same structure, and the SAR images formed by them are similar. The computational complexity is the only difference between them. Therefore, only the images archived by FA-ADMM-ST are depicted in this figure, and it demonstrates that the FA-ADMM-ST has good performance at all SNR levels with higher contrast and less noise, ghost areas, and distortion. Furthermore, to quantitatively compare the image quality and the computational load of the algorithms, two imaging metrics, namely, peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM); and computational time are deliberated. The PSNR is the metric to quantify the reconstructed image quality, whose higher value indicates a better-quality image, expressed as: xmax , (25) PSNR(x̂, x) = 20 log10 ∥x − x̂∥22 where x denotes the reconstructed image, and x̂ represents the ground truth. Whilst, SSIM with the highest value of 1, indicates the similarity of the x̂ and x which is defined as: SSIM(x̂, x) = (2µx̂ µx + C1 )(2σx̂x + C2 ) , (µ2x̂ + µ2x + C1 )(σx̂2 + σx2 + C2 ) (26) where µx̂ , µx , σx̂ and σx are the mean values and standard deviations of x̂ and x, respectively. C1 and C2 are the constants to stabilize the division with a weak denominator. Table 1 demonstrates how the ADMMbased algorithms perform better than the traditional RMA as they achieve 3