X-ray computed tomography (CT) systems are inherently susceptible to ring artifacts, primarily caused by hardware imperfections such as inconsistent detector responses, uncertainties in X-ray source intensity, and variations in the energy thresholds of detector pixels, etc. These artifacts degrade CT images in more complex ways than imagined, thereby having negative impacts on subsequent analysis. In this paper, we propose P-LIWLR, an effective method for ring artifact removal that compensates for system deficiencies. The proposed method demonstrates strong robustness against diverse ring artifact patterns. Our approach is motivated by the discontinuity of stripe artifacts in the sinogram domain and the geometric properties of the ideal mean projection (MP). By leveraging these characteristics, unstable ring artifacts are decomposed into stable subtypes and separately eliminated via local MP restoration using a novel regression model, LIWLR. Within LIWLR, an intrinsic weighting mechanism is introduced to automatically distinguish and prioritize reliable data over corrupted samples, thereby ensuring the effectiveness of the regression results. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art ring artifact removal techniques, both in terms of artifact suppression and accurate restoration of critical image structures.
Single-photon LiDAR (SP-LiDAR), with its extremely high sensitivity, is well suited for imaging in photon-sparse conditions and under strong background noise. However, most existing approaches rely on histogram accumulation; in dynamic scenes, histogram accumulation mixes time-of-arrival (ToA) events across motion, jointly biasing/blurring depth and reflectivity estimates. To address these issues, we propose a Single-Photon Neural Assumed-Density Filter (SP-NADF) that operates directly on ToA events. At the system level, SP-NADF integrates photon-statistics physics with the spatio-temporal representation power of neural networks, enabling low-latency recovery of depth and reflectivity while providing actionable uncertainty estimates to assess reconstruction risk and guide robust inference. We further introduce an explicit physics-guided coupling mechanism for depth and reflectivity within this neural framework, so that the two estimates reinforce each other and improve both predictions. In addition, uncertainty-guided spatio-temporal propagation and updates help improve reconstruction stability and reliability under noise and motion. Experiments show that SP-NADF delivers higher reflectivity quality and lower depth error across noise levels and motion conditions, with clearer structures and more stable geometry, indicating strong robustness and generalization.
Three-dimensional (3D) inverse synthetic aperture radar (ISAR) imaging significantly improves the target discriminability. However, in practical scenarios such as high sea states, the complex target maneuvers will introduce severe projective distortions, which significantly degrade the quality of 3D reconstruction. First, this paper analyzes the projective transformation for 3D ISAR imaging in the single-antenna system, and the trans formation matrix that maps the 3D space to the parametric imaging space can be derived. Then, this paper further investigates the distortion compensation method for top view dominant scenarios to achieve the robust 3D structural recovery. The proposed method estimates the target's bottom plane structure and then compensates for the shear distortion by solving an optimization problem formulated to minimize the projected area. Finally, the effectiveness of the algorithm is demonstrated through the results of simulation experiments and processing of measured data.
Electrical Impedance Tomography is a non-invasive imaging modality that is radiation-free and cost-effective; yet, accurately estimating internal conductivity remains challenging due to its inherent ill-posedness and strong nonlinearity. Implicit Neural Representations have recently emerged as powerful unsupervised deep priors, enabling continuous parameterization of conductivity. However, the intuitively appealing strategy of solving forward problems on coarse discretizations for efficiency while querying on fine meshes for geometric details is affected by training-time discretization coupling. In existing methods, networks are queried only at coarse integration nodes during training, thereby coupling learning to the coarse discretization. This can lead to mesh-dependent distortions for continuous backbones and out-of-distribution degradation for topology-dependent backbones when transferred to finer meshes. To mitigate this issue, we propose Cross-Scale Neural Representation (CSNR), an unsupervised framework that explicitly decouples conductivity representation from forward-solver discretization within the training loop. CSNR integrates a Cross-Scale Regularization Network (CSR-Net) with L2 projection operators, enforcing fine-mesh representation under coarse-mesh physics while promoting cross-scale consistency and suppressing mesh-dependent distortions. Extensive simulated and experimental results demonstrate that CSNR achieves high reconstruction fidelity and mitigates backbone-specific failures across resolutions.
Coded aperture imaging presents an efficient solution for compressively acquiring 4D light fields (LFs) by encoding high-dimensional data into compact 2D measurements. However, the primary bottleneck lies in the reconstruction process, where existing methods often struggle to effectively leverage the inherent contextual information within the coded measurements, leading to limited reconstruction fidelity. To address this challenge, we propose a novel framework based on state space model (SSM) for high-quality LF reconstruction from 2D measurements. To mitigate the inherent scanning order dependency of SSM, we introduce a multi-contextual SSM module designed to efficiently capture contextual information from different perspectives. Extensive experiments on both simulated and real-world coded measurements demonstrate that our proposed method achieves state-of-the-art reconstruction quality with reasonable inference speed.
Ultrafast Ultrasound (US) imaging is essential for applications requiring high temporal-resolution acquisition, such as fetal, cardiac, and musculoskeletal imaging. Its safety, portability, and real-time performance make US ideal for capturing rapid physiological motion. However, achieving high-quality image reconstruction from a single Plane Wave (PW) transmission remains a significant challenge due to limited spatial coherence, reduced signal-to-noise ratio, and spatially varying acoustic propagation speeds. Conventional Delay-and-Sum (DAS) beamforming from a single PW suffers from low contrast and strong artifacts, while multi-PW Coherent Compounding (CC) improves quality at the cost of reduced temporal resolution and greater acquisition complexity. Recent advances in computational imaging and machine learning can overcome these limitations and enable improved ultrafast US reconstruction. In this work, we propose a Reinforcement Learning (RL) framework that learns adaptive Speed-of-Sound (SoS) fields to optimize single-PW beamforming in a fully self-supervised manner. Unlike prior supervised deep learning methods, which require large, paired US data-to-image datasets or function as post-processing modules, our method directly integrates into the beamforming process, adapting reconstruction delays based on learned spatially varying SoS distributions. The RL agent optimizes SoS-dependent delays and channel apodization weights through deterministic policy gradients and perceptual image quality or signal coherence-based rewards, requiring neither ground-truth SoS maps nor multi-PW reference acquisitions. Experiments on publicly available PICMUS and CUBDL datasets demonstrated that the proposed RL-based adaptive beamformer consistently outperformed conventional DAS and supervised deep learning baselines. Notably, using a single PW, it achieved an image quality comparable to 25-PW CC while reducing the computational cost by approximately $7\times$. These results demonstrate that RL-driven adaptive SoS learning enables fast, label-free, high-quality US image reconstruction without requiring ground-truth training data.
With the rapid growth of aerospace activities, space situational awareness (SSA) has become increasingly important for space security. Compared with conventional two-dimensional (2D) inverse synthetic aperture radar (ISAR) image sequences, three-dimensional (3D) representations provide richer structural information. In addition, novel view synthesis (NVS) supports continuous visual interpretation and helps compensate for observation gaps. However, the limited angular coverage of single-pass observations makes stable 3D reconstruction and high-quality NVS difficult without accurate geometric calibration. To address these challenges, Radar Neural Splatting (RNSplat) is proposed for 3D reconstruction and NVS from ISAR image sequences. Specifically, a Pose Head Adaptation via Reprojection (PHARE) module is introduced to refine viewpoint parameters under a cross-view reprojection consistency constraint, thereby improving the estimation of a 3D point map. Together with the associated geometric attributes, the estimated point map is then used to construct the 3D Gaussian splatting (3DGS) representation. To better reflect ISAR image formation characteristics, a Rendering Stabilization Unit (RSU) is further introduced, using a Gaussian-sinc kernel as a PSF-based scattering response approximation to improve cross-view synthesis quality. Experimental results demonstrate that the proposed framework improves NVS performance and enhances visible structural consistency. Ablation studies further show that PHARE improves geometric consistency and 3D point map quality, while RSU enhances the consistency and quality of synthesized views.
Computed tomography (CT) reconstruction under sparse-view acquisition is fundamentally ill-posed. Recently, 3D Gaussian Splatting (3DGS) has emerged as an efficient alternative to implicit neural fields for per-case tomographic reconstruction, offering explicit geometric primitives and fast differentiable rendering. However, under sparse projection supervision, existing 3DGS-based CT methods can become unstable because the Gaussian primitives are optimized largely independently, often leading to overfitting and needle-like artifacts. In this work, we propose GR-Gaussian, a graph-regularized radiative 3DGS framework for sparse-view CT reconstruction. Rather than introducing a new volumetric representation, our method augments radiative Gaussian optimization with local graph-guided structural cues. Specifically, we introduce two components: (1) a denoised point-cloud initialization strategy (De-Init), which filters artifact-contaminated FDK priors to provide a more reliable initialization for Gaussian placement and neighborhood construction; and (2) a Pixel-Graph-Aware (PGA) densification criterion, which supplements the baseline pixel-aware densification signal with local density contrast measured on the Gaussian neighborhood graph. In addition, we incorporate graph Laplacian and volumetric total variation regularization to improve structural consistency during optimization. Experiments on the X-3D and real-world CT datasets show that GR-Gaussian consistently improves reconstruction quality over the evaluated baselines, while providing cleaner structures and stronger suppression of sparse-view artifacts. These results indicate that initialization and graph-guided densification are effective practical extensions to radiative 3DGS for sparse-view CT reconstruction.
Synthetic aperture radar (SAR) imaging algorithms are designed to image stationary scatterers. A moving target is mis-positioned and results in artifacts in reconstructed images. In this paper, we present a novel approach to ground moving target imaging (GMTI) using a single channel SAR data based on atomic norm minimization (ANM) that is capable of localizing targets with sub-resolution cell precision. We use ANM to decompose the received SAR data into the echoes received from each scatterer in the scene, which facilitates processing echoes individually and isolating velocity related phase variations by cross-correlations across the slow-time samples. We use these estimates for motion compensation, and form the SAR image using the dual polynomial of the atomic decomposition. Motion parameters are then numerically estimated without a pre-specified grid. We derive the theoretical requirements on SAR imaging parameters, and detectable velocity limit; and demonstrate the off-grid, sub-resolution cell localization capability of our ANM-based method in SAR-GMTI setting by numerical simulations.
In situ inspection in turbid liquids and aerosols often requires identifying an embedded object's material, while recovering three-dimensional shape or volumetric optical parameters from steady-state intensity becomes ill-conditioned under multiple scattering. We propose a computational imaging pipeline for spectral-angular material identification in scattering media under a fixed acquisition geometry. Each measurement is a co-registered tensor from seven fixed yaw viewpoints $\lbrace 0^{\circ }, \pm 15^{\circ }, \pm 30^{\circ }, \pm 45^{\circ }\rbrace$ and three narrow spectral bands 405 nm, 540 nm and 640 nm; the task is material discrimination without three-dimensional reconstruction or volumetric inversion. Condition-rich supervision is generated by Monte Carlo light transport in participating media with wavelength-dependent Henyey-Greenstein scattering and a controlled cumulative volumetric-event cutoff. To reduce simulation-to-measurement mismatch while preserving wavelength- and view-dependent cues, we apply two-stage unpaired refinement: global domain correction followed by cue-aware patchwise refinement. A contrastive spectral-angular encoder aligns matched instances across simulated, refined, and measured tensors, enabling prototype-based one-shot or few-shot recognition and measurement-to-synthetic retrieval without retraining. We additionally evaluate sensitivity to moderate pose deviations and a reduced-view trade-off. On measured metallic objects across held-out simple convex canonical shapes (cube, plate, cylinder, pillar), the method achieves 84.7 percent top-1 accuracy and 82.8 percent retrieval mean average precision (measurement-to-refined-synthetic, $M{\to }S^{\prime }$). On a complementary ten-pattern primitive-stacked composite (non-convex) stress test restricted to copper, iron, and aluminum, evaluated without retraining, it retains 78.1 percent top-1 accuracy and 76.0 percent retrieval mean average precision ($M{\to }S^{\prime }$), supporting moderate same-layout generalization beyond simple convex shapes.
Fluorescence molecular tomography (FMT) enables noninvasive 3D imaging of fluorescent probes, facilitating early tumor detection and drug metabolism studies. However, in highly scattering biological tissues, it is constrained by limited imaging depth and suboptimal spatial resolution. Although existing deep learning methods have enhanced reconstruction quality, they frequently induce physical modeling biases by overlooking the radial aggregation characteristics of fluorescent probes, and introduce reconstruction errors due to the inadequate characterization of photon transport dynamics in tissues. These issues ultimately manifest as target shape distortions, edge blurring, and noise sensitivity. To address these issues, we propose FDMamba-FMT, a feature-decoupled state-space model-based framework for FMT reconstruction. We design a physics-prior-guided spiral-centered scanning path that explicitly captures the radial gradient distribution of fluorescence intensity, effectively mitigating the physical biases arising from overlooked radiative attenuation in modeling. Furthermore, we introduce a hierarchical feature decoupling mechanism, utilizing pixel-level and region-level Mamba modules to suppress local noise and enhance representations in globally attenuated regions, respectively, thereby substantially mitigating reconstruction biases arising from signal attenuation-noise coupling. Experimental results demonstrate that, across simulation, phantom, and in vivo scenarios, the proposed method achieves strong overall reconstruction performance compared with state-of-the-art approaches, yielding reconstructions that are more physically consistent, structurally intact, and noise-reduced. This work holds significant promise for advancing precise in vivo tumor localization and 3D dynamic imaging.
Near-field synthetic aperture radar (SAR) imaging with sparsely sampled data attracts wide attention due to its ability to reduce data acquisition time. However, traditional low-rank or compressed sensing reconstruction methods are limited to random sampling scenarios, lacking adaptability to diverse sampling conditions and robustness against background clutter. To address these limitations, this paper introduces a 3-D radar imaging framework termed sampling adaptive matrix completion network (SAMCNet). The proposed framework employs progressive rank-one factorization and robust loss modeling, leveraging stable rank estimation to enhance adaptability under diverse sampling conditions. In addition, the optimization solver is designed with a dual-layer self-supervised iterative structure, which allows the network to reconstruct echo data under extreme sampling conditions without requiring external training data. Finally, extensive simulations and real-world experiments demonstrate that SAMCNet consistently achieves superior performance in sample recovery across various sampling conditions.
In this paper, a novel deep learning (DL) approach has been proposed to realize high-resolution electromagnetic (EM) inversion imaging. The newly proposed approach is based on the deep convolutional double-module structure (DCDMS), consisting of the pixel-interpolating module and the corresponding quality-improving module. While the pixel-interpolating module roughly increases the ‘resolution’ of the initial input, the following quality-improving module realizes quantitative EM imaging in high resolution. The input of the proposed DCDMS adopts the mixed input scheme, consisting of the received EM scattered field and the initial reconstruction in much low resolution computed from Gauss-Newton method. The output of the proposed model is the high-resolution contrast (permittivity) ‘image’ of the target domain. In such manner, the proposed DL approach can make use of much less measurement to realize high-resolution EM inversion imaging accurately and efficiently even for high-contrast scatterers, which can hardly be realized by conventional methods. The training of DCDMS is based on the simple synthetic dataset. Numerical benchmarks are offered to illustrate the excellent performance of DCDMS, which provides a novel thinking for conducting the real-time quantitative EM inversion imaging.
Intensity diffraction tomography enables three-dimensional (3D) imaging via refractive index (RI) reconstruction of objects without needing chemical contrast agents. However, the accuracy of RI reconstruction is typically limited because incorporating multiple scattering between different regions of the sample is difficult. While new approaches are emerging, multiple scattering is considered a problem rather than an asset. We present a proof-of-concept result of scattering-enhanced intensity-based tomography (SEIT) that uses multiple scattering as an opportunity to significantly improve the RI reconstruction accuracy. We use coherent laser light in side-illumination configuration, i.e., illumination perpendicular to the collection path of the microscope, where coherence and side illumination enhance multiple scattering and preserve its contrast in the final image. Then we use a full-wave multiple-scattering based 3D RI solver, contraction integral equation based inversion method, to reconstruct the sample image particularly designed for highly scattering samples. Our experimental results show that SEIT provides improved accuracy and resolution in RI tomography even with Delta n = 0.48, and opens opportunities for high contrast RI tomography.
Magnetic Particle Imaging (MPI), an emerging imaging modality characterized by its high sensitivity and spatiotemporal resolution, holds immense promise for advanced biomedical applications. However, its clinical translation is significantly hindered by the laborious and time-intensive physical calibration process requisite for acquiring the system matrix (SM). To address this limitation, this paper introduces the Adaptive Chebyshev Tensor Decomposition (ACTD) method. By incorporating an online-optimizable adaptive affine basis into a compact tensor-product model, the proposed method achieves an accurate representation of the intrinsic structure of the SM. We further establish a unified, SM-aware reconstruction framework wherein the iterative solving process is entirely transformed into the low-rank Chebyshev spectral domain, thus circumventing the need for explicit storage and manipulation of the complete SM. Experimental evaluations on the public OpenMPIData demonstrate that ACTD facilitates high-fidelity reconstruction from merely 1.5625% sparse sampling, significantly outperforming existing state-of-the-art methods with respect to key metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Moreover, the proposed framework reduces the memory footprint for reconstruction to only 6.7% of that required by conventional methods and inherently supports arbitrary-scale super-resolution imaging. Consequently, ACTD provides an efficient and robust approach for rapid and flexible MPI reconstruction, independent of supervised data-driven paradigms, thereby offering strong potential for practical applications. The code for this work has been open-sourced at: https://github.com/GALIFeng/ACTD/.
In this paper, we introduce an innovative deep learning (DL) methodology designed for real-time quantitative microwave imaging (MWI). Our approach is centered around the utilization of a deep convolutional asymmetric encoder-decoder structure (DCAEDS), which requires only a single-frequency far-field measurement of the electromagnetic (EM) scattered field as input and subsequently predicts the contrasts (permittivities) of the target materials. During the offline training process, we incorporate an EM forward solver specifically crafted to compute the EM scattered field generated by the predicted target contrasts (permittivities) produced by the DCAEDS. The DCAEDS is seamlessly integrated with this EM forward solver to optimize the loss function. This loss function comprises two fundamental components: (1) Data-induced loss, directly quantifying the dissimilarity between the predictions of our proposed DCAEDS and the actual labels for the target contrasts (permittivities); (2) Physics-induced loss, evaluating the distinctions between the measured EM scattered field and the computed EM scattered field derived from the predicted target contrasts (permittivities) generated by the DCAEDS. Our DL approach excels in delivering precise results, even for high-contrast targets, overcoming limitations associated with conventional methods, such as computational cost and ill-conditioning. Numerical benchmarks using dielectric targets underscore the practicality and effectiveness of our DL-based approach.
Learning-based multiview stereo (MVS) has achieved remarkable progress, largely driven by cascade-based architectures. However, existing methods inevitably suffer from error accumulation across stages because of simple upsampling. Furthermore, relying solely on RGB images often leads to photometric ambiguity in textureless or non-Lambertian regions and a lack of explicit 3D constraints, resulting in erroneous correspondences. To address these challenges, we propose NormalMVS, an efficient framework that explicitly incorporates surface normal priors to enhance reconstruction quality. Specifically, we design a geometry-aware fusion (GAF) module that synergistically combines RGB and normal features and extracts discriminative geometric cues to resolve photometric ambiguities. To mitigate error propagation, we design a normal-guided depth refinement (NGDR) module. Unlike conventional bilinear interpolation, NGDR refines upsampled depth maps on the basis of local coplanarity and is supervised by a novel neighbor-weighted consistency loss that enforces geometric coherence. Additionally, a four-stage architecture and a sparse sampling strategy are introduced to significantly reduce the computational and memory overhead of 3D CNNs. Extensive experiments on the DTU, Tanks & Temples, and ETH3D benchmarks demonstrate that NormalMVS outperforms most existing methods in terms of efficiency, effectively balancing reconstruction quality and resource consumption.
Future satellite LiDARs such as NASA's CASALS promise global monitoring but suffer from coarse footprints, sparse sampling, and severe photon/background/readout noise. We present a unified Bayesian framework that jointly performs compressed-sensing (CS) recovery, denoising, and super-resolution of HyperHeight Data Cubes (HHDCs). The inverse problem combines a physics-informed hierarchical likelihood that models PSF blur/decimation, photon statistics with background, and readout noise, with a data-driven prior parameterized by a diffusion model. Posterior-guided diffusion sampling couples the prior score with the likelihood gradient, yielding high-posterior reconstructions without hand-tuned regularizers. To our knowledge, this is the first evaluation on real CASALS engineering-flight data (Virginia, Nov. 2024) in addition to extensive emulation. On emulated datasets, the method improves PSNR by up to 5 dB over PnP-ADMM+BM3D and strong supervised deep-learning baselines (3D U-Net, SwinIR) while achieving better SSIM/LPIPS, especially at low SNRs and 25-50% illumination. On real data, it improves no-reference IQA (e.g., CHM NIQE 5.45 vs. 8.06) and achieves 0.085 bits/voxel better log-likelihood under the hierarchical model, with similar gains under a Gaussian approximation. The approach narrows the resolution-noise gap between satellite and airborne LiDAR, enabling more reliable canopy and terrain products for biomass, forest structure, and disaster applications.
Hyperspectral images (HSIs) capture rich spatial-spectral information and are widely used in various fields. However, the spectral ranges of hyperspectral imaging cameras are generally limited to a narrow range, like visible (VIS) or near-infrared (NIR), due to the constraints of optical elements and imaging sensors. Existing methods primarily focus on enhancing the visual effect in the spatial or spectral domain, neglecting the correlation between different spectral ranges. Therefore, this work aims to accurately correlate HSIs across different spectral ranges, enabling efficient spectral extension using algorithms. To achieve spectral extension, we explore the sparsity of real-world HSIs and propose a sparse representation framework based on the spectral convolutional dictionary. This approach reveals latent features shared across different spectral ranges, which are effectively captured by learned convolutional dictionaries and their associated sparse coefficients. Building upon this representation, we introduce the Unfolding Convolutional Sparse Representation Transformer (UCSRT) model, jointly optimizing the convolutional dictionaries and sparse coefficients. The Spectral Coding Transformer in the UCSRT achieves accurate and efficient sparse encoding through a sparse self-attention mechanism guided by spectral sparsity. Additionally, Spectral Convolutional Dictionary Learning Module is proposed to learn the convolutional dictionary for the VIS-NIR spectral range. This framework enables the extension of HSIs from a narrow VIS range to a wider VIS-NIR range. Extensive experiments on real-world datasets demonstrate the effectiveness and accuracy of our spectral extension method. Comparative analyses with state-of-the-art spectral reconstruction models, as well as practical applications for small target detection in disguised environments and enhancement in snapshot hyperspectral imaging, further highlight the superiority of our UCSRT method.
