Search Results (287)

Sampling from high-dimensional posterior distributions is a central challenge in Bayesian inference and noisy inverse problems. Standard first-order Langevin-based methods often suffer from slow convergence and sensitivity to step-size hyperparameters, particularly in annealed score-based inverse imaging pipelines. We propose Adaptive Momentum Langevin Dynamics (AMLD), a practical stochastic correction kernel that introduces a momentum variable into the annealed posterior sampling framework and equips it with an annealing-aware momentum retention schedule. The method is fully compatible with the SNIPS framework and retains its coordinate-wise adaptive step structure, acting as a lightweight drop-in replacement for the conventional first-order Langevin correction step. Extensive experiments on three representative image inverse problems—Gaussian deblurring, inpainting, and $4\times$ super-resolution—demonstrate that AMLD consistently achieves strong PSNR and LPIPS performance, with competitive FID in most settings, compared to three state-of-the-art baselines (DDRM, DPS, SNIPS) under both nearly noiseless and noisy measurement conditions, while reaching target reconstruction quality using fewer sampling iterations. The proposed momentum-based sampler provides empirically improved exploration and robustness across evolving posterior landscapes, offering a practical and computationally efficient alternative to first-order annealed Langevin samplers in high-dimensional Bayesian inverse problems. Full article

Object removal is a widely used AI-eraser operation in photo editing and privacy protection, yet conventional workflows make the user paint the removal region by hand—a burden that is most acute when an image contains several objects of the same class and only one is to be erased. We present two pipelines for maskless selective removal that delete a designated object among many, using only a target ID, box, or point. Both are built around a selection mechanism: the detected candidates are indexed, and only the mask of the designated target—combined by a logical-OR when several are chosen—is constructed and inpainted, unlike the standard use of an inpainting model, which merely fills a mask that is already given. On the multi-object GQA-Inpaint benchmark, we compare the generative YOLO–SAM–SDXL (YSS) and the lightweight YOLO/MobileSAM–LaMa (YML) against PowerPaint and Inpaint-Anything. With no mask provided, the proposed pipelines select and remove the target far more accurately and reliably than the baseline (for YML, Selection-IoU 0.626 vs. 0.327 and residual object score 0.228 vs. 0.646), while their full-image quality stays within the baseline range. YML is the most reliable remover and far faster (1.31 s per image on GPU vs. 29.60 s for YSS, and 32× faster than a diffusion pipeline on CPU), whereas YSS reaches the most natural perceptual quality at the cost of a higher residual. The two are thus complementary options, chosen according to the desired speed and restoration character. Full article

Symmetry and asymmetry play an important role in image processing and computer vision, particularly when visual structures are corrupted by irregular and spatially heterogeneous degradation. In cultural heritage restoration, ancient murals often contain locally symmetric patterns, repeated decorative motifs, balanced compositions, and style-sensitive contours, while long-term aging introduces asymmetric damage such as cracks, pigment fading, flaking, and partial content loss. Restoring such images therefore requires models that can recover structural regularity from asymmetric degradation while preserving culturally meaningful visual details. In this paper, we propose a symmetry-aware progressive generative framework for non-invasive digital restoration of Dunhuang murals. The proposed model is implemented as a Cauchy–Schwarz-regularized cascading variational autoencoder, which decomposes restoration into three coarse-to-fine stages: global structural recovery, semantic and chromatic refinement, and fine-detail enhancement. To support this progressive process, the latent dimensionality is gradually expanded across stages, enabling the model to move from compact structural abstraction to detail-aware representation learning. Moreover, a Cauchy–Schwarz-divergence-based regularization strategy is introduced to align the aggregated posterior with a mixture-of-Gaussians prior, providing a tractable mechanism for modeling the multi-modal latent structure of mural images. Experiments on the MuralDH benchmark under irregular-mask, crack-like, and mixed degradation settings show that the proposed method achieves competitive restoration quality compared with representative inpainting and diffusion-based baselines, while requiring substantially lower inference cost. Qualitative results further demonstrate improved contour continuity, chromatic coherence, and texture preservation. These results suggest that symmetry-aware progressive generative modeling is a promising tool for sustainable, non-invasive cultural heritage restoration. Full article

Cross-view re-identification of Hanwoo cattle remains challenging: weak texture provides only limited identity cues, and large pose and viewpoint changes easily distort these cues, causing severe feature dispersion. As a result, samples from the same individual may be scattered in the feature space, whereas visually similar individuals can form ambiguous local neighborhoods. To address this problem, we propose a framework that explicitly improves identity compactness for weakly textured cattle ReID. Specifically, Pose- and Text-Conditioned Inpainting Diffusion (PTID) is trained to generate pose-diverse yet identity-consistent samples and, during inference, aggregates their features with the original input feature to approximate a compact identity center and reduce pose-induced feature dispersion. Dual-Adaptive Viewpoint-Aware Feature Centralization (DVFC) further refines retrieval features by adaptively exploiting viewpoint consistency and neighborhood reliability, thereby suppressing unreliable neighbor mixing under large cross-view variation. We also construct a Hanwoo ReID dataset containing 37 identities and 12,480 images collected from three farm environments with substantial variations in pose, illumination, occlusion, and background conditions. Extensive experiments on multiple strong baselines under both closed-set and open-set settings demonstrate consistent improvements, with gains of up to 40.8 points in mAP and 26.1 points in Rank-1 accuracy. These findings demonstrate that explicitly improving identity compactness through identity-consistent generation and viewpoint-aware feature centralization is a promising solution for the robust re-identification of weakly textured cattle. Full article

Background/Objectives: Melanoma is a life-threatening skin cancer characterized by aggressive progression and high metastatic potential, making early diagnosis essential for improving patient survival and treatment outcomes. However, accurate automated skin lesion classification remains challenging due to variations in lesion appearance, illumination, image quality, and the presence of artifacts. This study proposes a unified framework for robust multi-class skin cancer classification by integrating preprocessing, lesion segmentation, feature extraction, optimization, and classification within a single end-to-end architecture. Methods: The proposed framework employs an iterative hair artifact removal strategy based on the fusion of Frangi vesselness filtering, Gabor texture filtering, morphological refinement, and Telea inpainting to preserve lesion integrity. A novel dermoscopic lesion segmentation network (CutisNet) is introduced to accurately delineate lesion boundaries. Hybrid representation learning combines deep features extracted using MobileNetV2 with uniquely selected handcrafted descriptors to capture complementary texture, structural, and contextual information. Gray Wolf Optimization is utilized for feature fusion and refinement, while a hybrid GNN–CNN classifier performs robust multi-class skin lesion classification. Results: Extensive experiments conducted on multiple benchmark dermoscopic datasets demonstrate the effectiveness and generalization capability of the proposed framework. The proposed model consistently outperformed existing state-of-the-art methods, achieving a maximum classification accuracy of 96.7% on the PH² dataset while maintaining competitive performance across other benchmark datasets. Conclusions: The proposed framework effectively integrates novel preprocessing, segmentation, hybrid feature representation, feature optimization, and classification strategies to improve the robustness and accuracy of automated skin cancer classification. These results demonstrate its potential to support reliable computer-aided diagnosis and assist clinicians in the early detection of skin cancer. Full article

The Social Internet of Things (SIoT) connects smart devices into social networks in which they share, forward, and consume visual content on behalf of their owners. As generative models become more capable, the images that circulate among these connected devices are increasingly easy to fake or manipulate, which threatens the trust relationships that hold an SIoT network together. Most existing forgery detectors return only a real or fake label, which gives a connected device no basis on which to decide whether to trust a neighbor or relay a piece of content. We propose an explainable image authentication framework for SIoT that classifies an image as real or fake and also provides a human-readable explanation and localized visual evidence for its decision. Our architecture, the Domain-Tag Guided Explainable Forgery Detection Module (DTE-FDM), uses a domain tag generator to predict the manipulation type (Photoshop, DeepFake, or AI-generated inpainting) and feeds it as a prompt to a multimodal large language model, which improves cross-domain generalization and produces a textual rationale. A Multimodal Forgery Localization Module (MFLM) and then grounds the explanation in the image by highlighting manipulated regions using a Tamper Comprehension Module combined with the Segment-Anything Model. We train the two modules in two stages on a multimodal tampered image dataset (MMTD) with triplet annotations. On MMTD, the method reaches 87.17% accuracy and 0.8696 F1, outperforming recent baselines, generates more relevant explanations (0.8566 CSS, 0.4348 ROUGE-L), and localizes manipulated regions with a mean IoU of 0.3438. The results show that large multimodal models can support accurate, transparent, and trust-aware content authentication for SIoT. Full article

Borehole electrical imaging logs provide important sensor-derived information for identifying fractures, bedding structures, and reservoir heterogeneity. However, strip-like missing regions commonly arise from uneven tool pad distribution, incomplete borehole coverage, and acquisition limitations, which can distort geological structures and reduce interpretation reliability. Because the true values inside real missing regions are unobservable, complete ground-truth labels are generally unavailable. This study therefore proposes a real missing-region-constrained self-supervised inpainting framework for borehole electrical imaging logs. The key idea is to use real missing masks to define safe intact regions and to generate artificial strip-like training masks only within those reliable regions, thereby avoiding supervision contamination from originally missing or adjacent unstable areas. A UNet-based inpainting model is developed, and two attention-enhanced variants, UNet-SE and UNet-CBAM, are evaluated together with Telea and Navier–Stokes baselines. In addition to the main comparison, mask-level ablation, model-level ablation, repeated-seed robustness analysis, and lightweight expert-assisted geological assessment are used to examine the reliability of the proposed strategy. The results show that learning-based methods consistently outperform conventional approaches for structurally complex strip-like gaps. UNet-CBAM achieves the best overall performance, with mask-MAE, mask-PSNR, and mask-SSIM values of 14.8, 21.2 dB, and 0.8, respectively. The safe-region-constrained strategy further reduces supervision contamination and improves reconstruction quality compared with random or less restrictive mask-generation strategies. These findings indicate that the proposed framework offers a practical self-supervised solution for improving the quality and interpretability of sensor-captured borehole imaging logs when complete labels are unavailable. Full article

Recent advances in NeRF-based inpainting have enabled the completion of masked regions across multi-view images. However, two major challenges remain: generating accurate masks efficiently in the presence of complex multi-object interference and maintaining view consistency without floating artifacts in large-scale, unbounded 360° environments. To address these challenges, we propose a 3D inpainting framework with two principal components. First, a depth-aware mask-generation pipeline produces high-quality, view-consistent masks from sparse annotations. Second, spherical parameterization is combined with a joint objective comprising smooth inter-layer and original-scene constraints to suppress floating artifacts and content drift in unbounded scenes. We also introduce IM2360, a multi-object 360° dataset for evaluating 3D inpainting methods. Experiments show that our approach outperforms existing NeRF-based inpainting methods in PSNR, LPIPS, and FID. On a single NVIDIA GeForce RTX 4090, our method requires an average of 25.21 min of training per scene, compared with 98.14 min for SPIn-NeRF, 38.71 min for OR-NeRF, and 11.29 min for InFusion. These results indicate that the proposed method provides a favorable balance between reconstruction quality and computational cost for high-fidelity restoration of complex immersive scenes. Full article

Image reconstruction from incomplete measurements is a fundamental problem in signal and image processing, with applications ranging from medical imaging to computational photography. In recent years, deep learning approaches have shown promising performance, particularly when combined with physics-inspired formulations such as deep unrolling. This paper presents a systematic comparative study of classical interpolation and variational reconstruction methods, direct convolutional neural networks (CNNs), and unrolled data-consistency CNN architectures for image reconstruction under different sampling patterns. We consider three representative mask types: structured block masks, nonuniform masks, and random sampling patterns, with sampling ratios ranging from 10% to 50%. Experiments are conducted on the public BSDS500 image dataset, using a fixed grayscale preprocessing pipeline and a reproducible train/validation/test split. Experimental results demonstrate that reconstruction performance strongly depends on the sampling pattern. For random masks, the full unrolled DC-CNN achieves the best quantitative and qualitative performance, reaching a PSNR of 30.98 dB and an SSIM of 0.921 at 50% sampling. In contrast, for structured block and nonuniform masks, TV-based inpainting provides the strongest overall performance, showing that classical model-based reconstruction remains highly competitive when the sampling pattern contains spatially coherent missing regions. A block-size sensitivity analysis further confirms that the difficulty of structured-mask reconstruction is governed by the geometric severity of the missing region. Statistical analysis using paired tests with Holm correction confirms that the main performance differences are significant across the evaluated configurations. Furthermore, we show that a lightweight unrolled model with shared weights and reduced depth achieves a substantially lower parameter count and lower computational cost than the full unrolled architecture, although with reduced accuracy in the most favorable random-sampling cases. These findings provide practical insights into the relationship between sampling strategies and reconstruction performance, offering guidance for the design of efficient and robust learning-based reconstruction systems. Full article

Four unvarnished oil paintings by the abstract expressionist artist Joan Mitchell were studied and conserved at the Carnegie Museum of Art (CMOA) and the Indianapolis Museum of Art at Newfields (IMA). These paintings, Sans Neige, Sans Neige II, Low Water, and Diabolo (neige et fleurs), all dating to 1969, share a similar color palette and degradation phenomenon in thickly applied cadmium yellow brushstrokes. Samples from three works were studied using Raman and FTIR spectroscopies, stereo microscopy, and scanning electron microscopy with energy dispersive spectrometry. A protocol was developed at CMOA for the consolidation of the crumbling yellow matte impasto through the careful application of funori with gentle manipulation of the paint. Acrylic fills were used to mimic Mitchell’s paint handling in areas of loss, and the fills were textured and inpainted using funori with ground original paint samples that had detached and could not be reattached to the painting’s surface. The treatment was successfully repeated at the IMA through online collaboration with CMOA during the COVID-19 pandemic. Samples were analyzed to identify, for the first time, the artist’s late 1960s palette. Waxy nodules observed in cross sections of the crumbling yellow paint were determined to be zinc stearate agglomerates, focused along break edges and incipient cracks, strongly suggesting a relationship between the instability of the paint and these zinc soaps. Full article

We study fixed-point-corrected numerical schemes for reverse-time diffusion sampling. Instead of treating the reverse sampler only as an explicit Euler–Maruyama discretization, we formulate each reverse-time step as a local implicit equation and approximate its solution by a finite number of fixed point corrections. After fixing the backward-time drift convention, we prove well-posedness of the local implicit step, contraction of the inner solver for sufficiently small step sizes, and a conditional global error decomposition separating terminal initialization, score approximation, assumed implicit time-discretization error, and fixed point truncation error. The estimate clarifies how additional inner corrections reduce numerical residuals at the cost of extra score evaluations. Multi-seed exact-score experiments, implicit-gap diagnostics, controlled score perturbations, a fitted Gaussian-mixture score example, and a neural-score Fashion-MNIST experiment with inpainting demonstrate the distinction between algebraic solver consistency and distributional sample quality. The paper is intended as a solver-level numerical analysis of reverse diffusion dynamics, not as a new large-scale generative architecture or a complete convergence theorem for diffusion models. Full article

Background/Objectives: Digital intraoral photography is widely used for clinical documentation, longitudinal monitoring, and AI-assisted dental image analysis. However, specular highlights caused by saliva and intense illumination can obscure tooth texture and compromise image fidelity. This study aimed to develop an automated method for removing specular highlights from intraoral images while preserving tooth surface texture. Methods: A three-stage pipeline consisting of adaptive threshold prediction, mask generation, and image inpainting was proposed. Initially, the Hue, Saturation, Value (HSV) statistical features were extracted from each image and used to train a regression model that predicts an image-specific threshold. Subsequently, the predicted threshold was applied in the CIE LAB color space, followed by a condition-based dynamic adjustment algorithm to refine the mask area and distribution. Finally, an Aggregated Contextual Transformation (AOT)-based generator network was used to restore the masked regions. Results: The proposed dynamic adjustment reduced over-masking compared with regression-only processing and better preserved tooth surface texture. Pixel distribution analysis demonstrated a lower distributional discrepancy, with the Wasserstein distance reduced from 2.9601 to 1.3505 and the Kullback–Leibler divergence reduced from 0.3451 to 0.0618. In the clinical expert evaluation, the proposed method was preferred in 69.5% of the 200 evaluation responses, and the preference difference was statistically significant (p < 0.001). Conclusions: As a proof-of-concept study conducted under controlled conditions using synthetic images, the proposed pipeline reduced specular highlights while better preserving tooth surface texture than the baseline approaches. These findings suggest that the pipeline may support standardized preprocessing of dental image datasets, although broader applications such as long-term monitoring and AI-assisted diagnostic workflows require validation on real clinical photographs. Full article

Dense simultaneous localization and mapping (SLAM) via 3D Gaussian splatting (3DGS) faces memory bottlenecks due to the explosive growth of primitives during long-sequence mapping. We propose SATA-SLAM, a framework featuring spatial-aware truncation and adaptive anti-hallucination. The online front-end maintains a constant memory footprint via a spatial-aware pruning module (SAPM), which employs a survival scoring function that couples primitive opacity with view-frustum projection coverage and a temporal protection window. Subsequently, an anti-hallucination generative refinement module (AGRM) utilizes texture priors from pretrained diffusion models for offline inpainting of residual regions. In addition, an adaptive gating mechanism to verify and suppress AIGC-induced hallucinations caused by pose drift, ensuring multiview consistency. Experiments on the public Replica dataset show that SATA-SLAM improves rendering quality from 12.5 dB to 37.44 dB (averaged over the Replica room0 and office0 scenes) while using only 26% of the original memory, outperforming the unconstrained baseline. This study provides a pathway toward low-power, high-fidelity environmental perception for mobile robots. Full article

With the wide application of generative models in the field of SAR image inpainting, inadequate reconstruction quality of scattering characteristics and insufficient global coherence of semantic logic remain the core challenges of such tasks. To address these issues, this paper proposes a Feature-Encoding Diffusion Model (FEDM). Guided by local valid regions, the proposed model accurately learns the microwave backscattering distribution law of ground features through a SAR-specific Variational Auto-Encoder (SAR-VAE), thus improving the reconstruction accuracy of backscattering statistics. Meanwhile, it integrates semantic embedding and cross-attention mechanism to strengthen the semantic constraints of SAR scenes, ensuring the logical rationality of the ground feature layout. With progressive diffusion generation and sliding window strategy, the model achieves high-quality reconstruction with coherent semantics and consistent global spatial structure for large-scale missing regions. Experiments on public datasets including OSdataset, SEN1-2, SRSDD-v1.0 and MRSSC show that the proposed method achieves excellent performance in terms of scattering characteristic reconstruction quality and globally coherent generation of semantic logic, and realizes high-quality SAR image inpainting. Full article

Implied volatility surfaces describe option-implied volatilities across strikes, and maturities and play a central role in derivative pricing and risk management. However, in practice, they are often incomplete due to illiquidity or sparse trading, requiring reliable reconstruction of missing regions. Existing approaches typically rely on parametric assumptions or latent space optimisation methods, which may be restrictive or computationally intensive. This study proposes a data-driven framework based on conditional generative adversarial networks (GANs) to map partially observed surfaces to completed ones in a single forward pass. The approach is evaluated in a controlled setting using synthetic data generated from the Heston stochastic volatility model, with varying levels of missingness (10–96%). The generator objective incorporates penalty terms enforcing the absence of call-spread, butterfly-spread, and calendar-spread arbitrage, together with a smoothness regulariser on the implied risk-neutral density. Compared with a conditional variational autoencoder (VAE), the Bates model, and the stochastic volatility-inspired (SVI) parameterisation, the proposed approach achieves lower reconstruction errors across all levels of missingness, including unseen cases, while preserving the no-arbitrage properties. An ablation study shows that the conditional GAN implicitly learns no-arbitrage behaviour, with density smoothness regularisation being the only constraint that meaningfully improves reconstruction quality. Full article