Search Results (352)

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

Real-world speech restoration must handle coupled distortions, including acoustic noise and reverberation, codec artifacts, clipping, and artifacts left by upstream enhancement systems. Token-based generative systems offer a flexible route for such universal restoration, but discrete audio tokens can discard fine acoustic detail, and aggressive generative decoding may over-process inputs that are already close to clean speech. We propose AudioVAE-MASR, a continuous-latent masked autoregressive framework for multi-distortion speech restoration. A frozen AudioVAE maps clean and degraded speech into paired continuous latent sequences; a Conformer-based branch extracts the degraded-condition sequence $C_y$ from degraded latents; a two-stream masked autoregressive encoder-decoder conditions masked clean-latent recovery on both degraded context and visible clean tokens; and a lightweight diffusion head models the masked clean tokens in the continuous latent space. On the released CCF AATC 2025 blind test set, the main inference setting ($K=16$, temperature $0.5$) achieved WAcc 0.793, SIG 3.401, BAK 3.987, OVRL 3.111, PESQ 1.780, and ESTOI 0.798. Relative to the degraded input, these results improved WAcc and DNSMOS but did not improve PESQ; relative to the organizer baseline, they improved WAcc, SIG, OVRL, and PESQ but remained lower in BAK. A local subjective MOS evaluation with five listeners gave an overall mean score of 4.08 for AudioVAE-MASR, compared with 3.70 for the degraded input and 4.59 for the clean reference. Distortion-type, ablation, and parameter-sensitivity analyses further show that codec inputs remain vulnerable to over-restoration and that longer iterative decoding does not provide a consistent gain. The study therefore presents AudioVAE-MASR as a transparent continuous-latent restoration framework and identifies the fidelity-control problems that must be solved before such generative restoration can match the strongest lightweight discriminative systems. Full article

Distinguishing authentic gold products from high-quality counterfeits is a challenging fine-grained computer vision problem; counterfeit items are engineered to replicate surface texture, hallmark engravings, color, and geometry with remarkable fidelity, making visual discrimination unreliable even for trained professionals. In this paper, we address the problem of visual gold authentication from unconstrained smartphone imagery in three main contributions. First, we introduce GoldNet, a public benchmark dataset designed for this task, comprising 2127 real-world images of authentic and counterfeit gold items collected under diverse real-world conditions. Second, we evaluate fourteen classification architectures spanning classical handcrafted texture descriptors, convolutional neural networks (CNNs), and vision transformers under a rigorous transfer learning protocol, establishing the first comprehensive baseline for this problem. Third, we propose GoldFormer, a hybrid dual-stream algorithm that combines the local texture representations of ResNet-50 with the global contextual modeling capability of the Swin Transformer (Swin-T) through a newly designed Texture-Aware Attention Gate (TAAG) module. The TAAG dynamically modulates Swin feature dimensions using CNN-derived texture energy, providing improved discriminability and per-prediction interpretability without requiring post hoc attribution. Experimental results show that, under matched-resolution 5-fold cross-validation, the proposed GoldFormer attains the highest overall accuracy (95.02%, F1-score 0.9502) at roughly half the FLOPs of its higher-resolution setting, statistically tied with the strongest individual backbone (ViT-B/16, 94.31%; McNemar $p=0.23$) and on par with a training-free soft-voting ensemble (94.92%), while significantly improving on its own Swin-T backbone (93.65%) and adding built-in, attribution-free texture-gate interpretability. GoldFormer surpasses trained human-expert performance (89.80%) by approximately 5 percentage points. Full article

This study presents a novel technological framework that integrates web-based 3D modeling and digital prototyping into interdisciplinary design education. Addressing the gap between traditional theoretical assessment and modern industry demands, the research investigates the implementation of interactive micro-websites and high-fidelity 3D product models as standard deliverables. Using a quasi-experimental design, the proposed digital workflow was tested on 53 final-year graphic design students at the University of Zagreb, divided into three groups based on the end users of their digital prototypes: real industry clients, peers, or academic mentors. The systemic reliability of the technological framework was measured through the technical quality of the final output (grades) analyzed via ANOVA, while user engagement with the digital process was tracked longitudinally. Results indicate that the implemented technological pipeline produced consistently high-quality outputs across all cohorts, with the client-facing group achieving the highest technical scores (M = 4.37; SD = 0.57). The lack of statistically significant variance between groups highlights a “ceiling effect,” demonstrating that the structured digital workflow itself is operationally stable and ensuring top-tier technical performance and prepress accuracy regardless of the evaluator. The study concludes that integrating advanced 3D web technologies and interactive public deliverables into the curriculum provides a scalable, industry-aligned technological model that successfully prepares design engineers for complex professional environments. Full article

Early detection of diabetes mellitus (DM) is essential for preventing disease progression and improving clinical outcomes. However, developing robust machine learning (ML) models for diabetes diagnosis is often constrained by limited data availability, privacy regulations, and challenges with data sharing. This study investigates a privacy-preserving synthetic data augmentation framework for classifying diabetic and non-diabetic blood serum samples using Fourier Transform Infrared (FTIR) spectroscopy. Two deep generative approaches, Autoencoders (AEs) and Generative Adversarial Networks (GANs), were evaluated for their ability to generate realistic synthetic FTIR spectra while preserving the statistical and biochemical characteristics of the original dataset. Synthetic datasets generated by the AE and GAN models were assessed using six ML classifiers: Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), Gradient Boosting (GB), Logistic Regression (LoR), and Decision Tree (DT). Model performance was evaluated using accuracy, precision, recall, F1-score, Receiver Operating Characteristic (ROC) curves, and Area Under the Curve (AUC). Results showed that AE-generated spectra retained stronger discriminative characteristics and were more easily distinguished from the original spectra, whereas GAN-generated spectra exhibited lower classifier separability, suggesting closer alignment with the original data distribution and greater realism for privacy-oriented data augmentation. Correlation analysis demonstrated high spectral fidelity for both approaches. Compared with the original spectra, AE-generated spectra achieved r = 0.9990 and R² = 0.9999, whereas GAN-generated spectra achieved r = 0.9982 and R² = 0.9965. The most prominent diabetes related spectral variations were observed in the carbohydrate (1000–1200 cm⁻¹), Amide I (~1650 cm⁻¹), and lipid-associated (3000–3500 cm⁻¹) regions. To explore the transferability of the proposed framework, a preliminary experimental feasibility study was conducted using independently acquired whole blood FTIR spectra. The generated spectra showed strong agreement with the measured whole blood spectra, demonstrating the potential applicability of the framework under alternative sampling conditions. Because the experimental cohort included only one diabetic volunteer, this analysis was intended solely as a proof-of-concept assessment of spectral feasibility and methodological transferability, rather than as a validation of diabetes classification performance. Overall, the findings demonstrate that synthetic data generation can effectively augment limited FTIR datasets while preserving privacy and key spectral characteristics. The proposed framework provides a promising foundation for privacy-aware biomedical data augmentation and future development of robust FTIR diabetes screening systems. The results should be interpreted as methodological evidence of feasibility and synthetic data utility rather than as evidence of clinical diagnostic readiness, as the serum dataset remains modest in size and the independent whole-blood experiment was intentionally exploring. Full article

Deep learning architectures generate highly effective but difficult-to-audit latent representations, creating a practical gap between predictive performance and verifiable explanations. Existing post hoc techniques often produce fragmented local attributions rather than dataset-level rulebooks. In this work, we propose Global SEMantic TRAnsition (SEMTRA), a post hoc framework that maps frozen representation features into semantic attributes, discretizes those attributes, and induces rough-set production rules with explicit coverage, conflict, fidelity, and abstention reporting. Evaluated on the Animals with Attributes 2 (AwA2) Protocol A, the semantic transition achieved a Mean Absolute Error (MAE) of $0.1029\pm 0.0005$. The extracted rulebook covered 84.80% of test instances, yielding a covered accuracy of 39.73% and a covered fidelity to the base predictor of 40.48%. Under the Protocol B split, continuous semantic-prototype transfer reached an unseen-object accuracy of $44.02\%\pm 1.22\%$ as a semantic-transfer validation. Cross-domain validations using SUN and Derm7pt demonstrated that the audit protocol is portable yet strongly dataset-dependent. In the controlled synthetic benchmark, SEMTRA achieved a macro-F1 score of 0.879 at zero semantic noise and degraded to 0.838 at the highest evaluated noise level. Ultimately, SEMTRA serves as a transparent audit layer to expose the verifiable logical subset of a model, rather than replacing the underlying predictor. Full article

Lead isotope ratios are frequently used in archaeometric provenance analysis; however, the overlap of isotopic fields within the Andean metallogenic belt complicates reliable provenance determination. This study presents a reproducible fidelity-based kernel method for the unsupervised clustering of Andean lead-isotope data and investigates whether a quantum-mechanical similarity space can reveal geologically significant structures beyond the classical Euclidean partition. A dataset of 1522 measurements of ²⁰⁶Pb/²⁰⁴Pb, ²⁰⁷Pb/²⁰⁴Pb, and ²⁰⁸Pb/²⁰⁴Pb was analyzed using a fidelity-based quantum kernel based on a three-qubit Pauli feature map and compared with classical K-means clustering, Gaussian mixture models, and Ward’s agglomerative clustering under various preprocessing strategies and cluster counts. The optimal quantum kernel setup achieved the highest silhouette score at k = 2. However, because analytical uncertainties were not consistently reported across all the compiled sources, an uncertainty-weighted similarity could not be applied. Geological insights indicate that this binary division separates less radiogenic, arc-related compositions from more radiogenic and thorogenic crustal signatures, a contrast that broadly follows the west-to-east crustal-contamination gradient across the Andes. Conversely, the traditional four-cluster approach provides more detailed subdivisions that align with the previously identified isotopic provinces. The reported separation reflects the geometry of the quantum feature space rather than any hardware-level speed-up, as this work represents only a simulation approach. Overall, these findings support a hierarchical and complementary approach to analyzing Pb isotope origins, in which quantum kernel clustering provides robust large-scale separation and classical clustering enhances regional understanding. Full article

Ceramic additive manufacturing offers strong potential for fabricating geometrically complex and application-specific components, yet achieving reliable dimensional fidelity remains challenging because dimensional deviation is governed by highly coupled material, process, thermal, and environmental factors. To address this problem, this study proposes an uncertainty-aware computational framework for dimensional error prediction in ceramic 3D printing under variable material and process conditions. The contribution is positioned as a system-level integration of established learning, uncertainty estimation, calibration, and reliability-interpretation components within a ceramic additive manufacturing dimensional-error prediction workflow, rather than as a fundamental methodological breakthrough. The validation is conducted using the publicly available Ceramic 3D Printing Process Control Dataset, a 1000-sample tabular dataset, and the resulting findings are therefore interpreted as dataset-specific computational evidence rather than direct proof of industrial deployment readiness. The methodology begins with a structured data-driven preprocessing pipeline that transforms the Ceramic 3D Printing Process Control Dataset into a multi-condition feature space through data cleaning, one-hot material encoding, min–max normalization, and engineered descriptors capturing extrusion–speed balance, thermal gradients, cooling intensity, deposition density, and material-conditioned interactions. A multi-branch deep computational architecture is then developed to encode material, process, thermal-environmental, and engineered-feature streams separately, followed by adaptive cross-condition fusion to learn nonlinear dependencies across ceramic printing regimes. To improve reliability beyond deterministic regression, the framework jointly models aleatoric and epistemic uncertainty and incorporates calibration refinement to align predictive confidence with observed error behavior, thereby enabling preliminary reliability-oriented interpretation of stable and high-risk operating conditions. Experimental results demonstrate that the full model achieves the best overall within-dataset performance, with a test MAE of 0.0118, RMSE of 0.0172, $R^2=0.999$, MAPE of 1.74%, calibration error of 0.003, PICP of 0.996, reliability score of 0.992, and a stable prediction rate of 98.7%. Although these values indicate strong predictive behavior under the current structured dataset, the exceptionally high $R^2$ should be interpreted cautiously because external experimental validation, larger measured datasets, and cross-machine ceramic printing trials are still required. These findings show that the proposed framework provides an effective system-level computational strategy for dataset-specific reliability-aware dimensional quality prediction in ceramic additive manufacturing and offers a preliminary data-driven foundation for uncertainty-aware intelligent process optimization. Full article

Driven by the high-end furniture industry’s demand for skin-tactile decorative boards, UV inkjet printing shows potential for wood-based surface finishing. Using primer-treated inkjet-printed melamine-faced panels, this study compared traditional UV varnish coatings with different thicknesses and UV curing intensities and 254 nm UV excimer-cured coatings with different radiant energies. Varnish thickness significantly affected surface roughness, 20° gloss, 85° gloss, and color difference, indicating a trade-off between matte tactile appearance and color fidelity. Thinner varnish coatings exhibited higher roughness and lower gloss but larger color differences, whereas thicker coatings better preserved color fidelity but resulted in higher gloss. For the UV excimer-cured system, one-way ANOVA showed significant treatment effects on acrylate conversion, water contact angle, 85° gloss, surface roughness, and abrasion mass loss. The coating prepared at an excimer radiant energy of 827.9 mJ/cm² showed the lowest 85° gloss of 5.28 GU and a pencil hardness of 3H, but also exhibited the highest abrasion mass loss in the short-cycle abrasion screening test. For both coating systems, three independently prepared specimens were tested for each processing condition. The UV varnish system was analyzed using two-way ANOVA, whereas the UV excimer-cured system was analyzed using one-way ANOVA. Friedman tests of sensory evaluation data showed significant differences among the eight selected samples for fineness, smoothness, and elasticity, with the excimer-cured coatings generally receiving higher fineness and smoothness scores than the UV varnish coatings. These results indicate that 254 nm UV excimer curing is a promising route for producing low-gloss, micro-wrinkle-induced skin-tactile surfaces on inkjet-printed melamine-faced panels, although optimization should balance tactile quality, gloss reduction, and abrasion resistance. Full article

We develop SAFIRE (Self-Attention Fuzzy-Inspired Rule Estimator), a differentiable fuzzy-inspired rule-scoring surrogate for binary medical tabular classification coupling multi-head self-attention, Gaussian membership functions, and Hard Concrete gates for continuous rule scoring. We position SAFIRE as a smooth surrogate of the discrete $L_0$-regularised rule-selection problem and establish five mathematical results and one complexity remark: (1) the relaxed objective is differentiable almost everywhere under positive Gaussian widths (enforced by a Softplus reparameterisation) and fixed batch-normalisation statistics; (2) the deterministic-inference active threshold is strictly stricter than the expected-nonzero training threshold, identifying Hard Concrete gates as continuous rule-scoring devices rather than automatic pruning mechanisms; (3) per-sample forward complexity identifies attention and rule layers as the dominant terms; (4) the Softplus–BatchNorm–linear rule operator violates all four triangular-norm axioms—with necessary and sufficient conditions per axiom and a no-finite-parameterisation impossibility result—while a Softplus reparameterisation restores coordinate-wise monotonicity; (5) a margin-based upper bound characterises disagreement between the full classifier and a top-k rule-only surrogate; and (6) the Softplus-reparameterised constrained variant is provably coordinate-wise monotone with explicit asymptotic regimes. Evaluated on four University of California, Irvine (UCI), medical binary tabular benchmarks under repeated stratified cross-validation, SAFIRE-Prog is statistically competitive with strong interpretable, modern, and gradient-boosting baselines, with one Bonferroni-significant gain over RuleFit on the Diabetic Retinopathy Debrecen corpus. The 48-configuration Hard Concrete sweep, constrained-variant comparison, and a top-k fidelity analysis (per-fold range 0.73–0.95) provide quantitative companion measurements for the mathematical framework. A supplementary large-scale hospital electronic health record (EHR) benchmark (Diabetes 130-US Hospitals, $n=\mathrm{101,766}$) shows the rule-scoring mechanism scales to ∼${10}^5$ records and, under severe class imbalance, statistically matches gradient boosting on accuracy while significantly exceeding it on macro-F1. The results offer a mathematically auditable pathway towards interpretable, auditable rule scoring for medical tabular classification, with rule signatures defined in a projected latent space rather than over raw clinical variables. Full article

Significant progress in legal natural language processing (NLP) has enabled advancements in tasks such as legal judgment prediction, case retrieval, and question answering. However, the development of analogous technologies for Arabic legal texts remains severely constrained by the scarcity of large-scale, publicly available benchmarks for summarisation and classification. This paper addresses this gap by introducing a novel, comprehensive dataset of 9699 Arabic legal cases sourced from the Saudi Board of Grievances. This corpus is unique in pairing full-length court decisions with expertly human-crafted abstractive summaries and multi-class category labels (Administrative, Commercial, and Criminal), establishing a dedicated benchmark for Arabic legal NLP. The dataset was constructed via a robust, reproducible pipeline that ensures high textual fidelity, incorporating specialised optical character recognition (OCR) via Google Document AI and precise structural segmentation into facts, reasons, and summaries. To establish robust baselines, we conduct an extensive empirical evaluation of seven summarisation models—encompassing four extractive algorithms (TextRank, LexRank, Latent Semantic Analysis, and Luhn) and three transformer-based abstractive architectures (AraT5v2, AraBART, and mBART)—each evaluated in both base and fine-tuned configurations. Results across ROUGE, BERTScore, BLEU metrics and human evaluation demonstrate substantial performance gains achieved through domain-specific fine-tuning, with the fine-tuned AraBART model achieving the strongest performance among all evaluated models. Furthermore, we present a novel analysis of the downstream utility of generated summaries by evaluating their performance on legal category classification using five machine learning models. This investigation reveals a strong positive correlation between summarisation quality and classification accuracy, empirically demonstrating that domain-adapted abstractive summarisation not only enhances intrinsic evaluation scores but also significantly boosts extrinsic task performance. By providing this essential dataset and comprehensive benchmarking, our work contributes a much-needed resource to the field, facilitating future research and innovations in Arabic legal text analysis. Full article

Group Relative Policy Optimization (GRPO) streamlines the alignment of Large Language Models (LLMs) and Vision–Language Models (VLMs) by eliminating the Critic model. However, its efficiency heavily depends on the group size, G. While a larger G improves reward estimation and stabilizes the Advantage, $A_i$, it drastically increases VRAM usage and reduces throughput. Standard heuristics like a fixed G of 64 create significant bottlenecks in resource-constrained settings. This paper introduces an Efficiency-Aware optimization framework utilizing Multi-fidelity Bayesian Optimization and Hyperband (BOHB) to dynamically identify the optimal group size, $G^{\ast }$. The method uses a multi-objective function that balances reward accuracy, $A_i$ variance, and hardware utilization, applying z-score normalization. By employing Successive Halving to quickly evaluate candidates at low fidelity, the framework reduces search costs by up to 74% compared with random search. Tested across text-only LLMs (Qwen2.5-7B/1.5B) and multimodal VLMs (Qwen2.5-VL-3B), the framework demonstrates that the discovered $G^{\ast }$ saves up to 72.5% in VRAM compared with the baseline of 64, while maintaining reward accuracy within 5.8%. Sensitivity analyses on hyperparameters like $\lambda$, $\alpha$, and $\beta$ confirm the framework’s robustness. Rather than treating group size as a mere engineering heuristic, this study establishes a principled methodological advance by formalizing the trade-off between statistical estimation stability and hardware constraints into a unified optimization framework for resource-efficient RLHF. Full article

Immunohistochemistry (IHC) is a widely used method for detecting specific proteins in tissue samples, helping diagnose diseases such as cancer. Traditional analysis methods rely heavily on human interpretation, which can lead to inconsistencies. In this study, we propose DeepFlare, a weakly supervised deep learning framework for cross-modality translation and segmentation of immunofluorescence and immunohistochemistry images. The proposed method utilizes multiplex immunofluorescence (mpIF) and co-registered IHC images, combined with preprocessing techniques such as affine transformation, stain normalization, noise reduction, and artifact removal. Multiple imaging channels, including hematoxylin, DAPI, Lap2, and nuclear envelope signals, are leveraged to generate segmentation masks using a U-Net++ architecture. The final segmentation mask is obtained through weighted fusion of modality-specific outputs. A generative adversarial network (GAN) is employed to measure translation fidelity between generated and real images. Weakly supervised learning techniques, including image-level supervision and consistency constraints, are applied to enhance performance under limited annotation scenarios. Pretrained pathology foundation encoders such as UNI and Virchow are integrated to extract multi-scale morphological and contextual features. Explainable AI techniques are incorporated to highlight critical regions and refine model attention. Experimental results demonstrate strong performance, achieving an SSIM of 0.7077 for image translation and a Dice score of 0.7424 for segmentation. The integration of the UNI encoder provides marginal improvement over the baseline (0.72 Dice score), indicating limited domain adaptation without fine-tuning on the dataset of 1264 training samples. Full article

This paper describes design, implementation and initial evaluation of Digital Twin Metaverse Classroom for higher education. Digital Twin Metaverse Classroom refers to highly realistic digital replicas or virtual replicas or prototypes of university classrooms or learning spaces. This paper focuses on creating high-fidelity digital replica of typical university lecture room. The main purpose of the Digital Twin Metaverse Classroom is to support teaching and learning in addition to traditional videoconferencing. The pilot involved thirty-two undergraduate students. A single-group pre-test/post-test quiz measured short-term learning, while the Technology Acceptance Model (TAM) measured acceptance through perceived usefulness, perceived ease of use, attitude toward use, and behavioral intention. A single session raised the mean quiz score from 6.41 to 9.19, a within-session gain that reached statistical significance, while all four TAM constructs scored highly. Because the sample was small and confined to one institution, with neither a control group nor a follow-up, these findings are best read as early evidence of feasibility, short-term improvement, and favorable acceptance rather than as proof of comparative effectiveness. Full article

Background: Nuclear medicine healthcare professionals (NMHP) sustain chronic occupational exposure to iodine-131 (I-131), conferring an elevated risk of radiation-induced solid thyroid cancer. Established radiobiological prediction tools derive risk coefficients from atomic bomb survivor data but are not configured for rapid individualized risk assessment in occupational exposure settings. This study examined whether machine learning algorithms can serve as high-precision computational surrogates for excess relative risk estimation in NMHP. Aim: The study aimed to (i) develop and validate three machine learning algorithms for predicting the excess relative risk per unit absorbed dose for radiation-induced solid thyroid cancer (ERR/Gy.RST), (ii) characterize relationships between dosimetric and demographic features and predicted risk, and (iii) identify the optimal algorithm for deployment in occupational health surveillance. Methods: A dataset of 4657 observations was constructed from Life Span Study-derived ERR/Gy parameters, adapted to occupational low-dose conditions, using a dose-and-dose-rate effectiveness factor of 2.0, per ICRP Publication 103. Five features (gender, age at exposure, current age, distance from the I-131 source, and cumulative absorbed dose in the thyroid) were used to train a decision tree regressor (dtcr), a random forest regressor (rfr), and a multilayer perceptron (MLP) neural network algorithm. Results: Cumulative absorbed dose in the thyroid correlated positively with ERR/Gy.RST (r = 0.63, p < 0.01), while radiation source distance demonstrated a strong inverse association (r = −0.38, p < 0.01). The MLP algorithm achieved R² score = 0.999, MSE = 0.002, and MAE = 0.010, substantially outperforming the rfr (R² score = 0.700, MSE = 0.410, MAE = 0.295) and the dtcr (R² score = 0.510, MSE = 0.654, MAE = 0.289). Conclusions: The MLP algorithm provides a high-fidelity surrogate for established ERR/Gy.RST projection tools in the NMHP context, enabling computationally efficient, feature-integrated occupational radiation-induced thyroid cancer risk quantification. These findings suggest that machine learning-based surrogate modeling is a practical, scalable complement for occupational health practitioners and radiation protection officers to support individualized surveillance of radiation-induced thyroid cancer risk in nuclear medicine departments. Full article