Search Results (4,908)

Few-shot font generation can be viewed as a challenging conditional image generation task, where the goal is to synthesize target glyphs from only a few reference samples while preserving structural fidelity and style consistency. This problem becomes particularly difficult for characters with complex spatial layouts and fine-grained stroke topology, where existing methods often struggle to simultaneously maintain structural integrity, local continuity, and stylistic coherence under sparse-reference conditions. To address this issue, we propose a skeleton-aware deformable alignment framework for few-shot font generation. Specifically, explicit skeleton priors are introduced into the diffusion-based generation process to provide structural supervision during denoising. In addition, a structure-constrained deformable content alignment module is designed to improve local feature correspondence while suppressing unreasonable geometric deformation. We further develop a multi-module content aggregation strategy to jointly model global layout patterns and local stroke details through complementary multi-level representations. Extensive experiments demonstrate that the proposed method consistently outperforms state-of-the-art approaches in both quantitative and qualitative evaluations. The results show that our method provides stronger structural preservation, better perceptual quality, and improved generalization in structurally complex glyph generation and cross-lingual style transfer. Full article

Remote sensing has become a key non-invasive tool in archaeological prospection, particularly in regions where logistical constraints limit sustained fieldwork. This study presents the results from Zar Tepe (1st–5th centuries AD), in the Surkhandarya province of southern Uzbekistan, within northwestern Bactria. The research aimed to document the site’s urban layout, accurately relocate Soviet-era excavation sectors within the present-day topography, and refine the interpretation of earlier interventions that were only partially documented and lacked precise georeferencing. A multiscale and multitemporal methodology was applied, integrating CORONA and WorldView-3 satellite imagery, UAV and terrestrial photogrammetry, GNSS Precise Point Positioning, magnetic prospection, and targeted archaeological verification. The workflow followed an iterative laboratory–field sequence, combining remote-sensing analysis, field checks, data refinement, and systematic ground-truth validation. Fieldwork was conducted during two contrasting phenological periods, in June 2024 and December 2025, to assess seasonal variability in surface and subsurface visibility. The integrated approach enabled the accurate spatial fitting of legacy excavation sectors and supported the cross-validation of optical and salt-efflorescence-related anomalies with geophysical evidence. These results provide a stronger basis for the cautious interpretation of buried architectural features and for refining hypotheses concerning Zar Tepe’s urban organization and occupational dynamics. Full article

Background: Breast cancer remains a leading cause of cancer-related mortality, and reliable computational decision support is increasingly viewed as a complement to expert pathological assessment rather than a replacement for it. Variational quantum classifiers (VQCs) and Quantum Support Vector Machines (QSVMs) have recently been promoted as candidate models for medical classification, yet most published comparisons rely on internal hold-out validation alone and report only a single point estimate of discrimination, omitting calibration, decision-analytic value, and explainability—three ingredients that any clinically credible model must furnish. Methods: We assembled a complete quantum–classical machine learning pipeline and evaluated it under a deliberately stringent protocol designed to expose, rather than conceal, the limitations of current Noisy Intermediate-Scale Quantum (NISQ)-era models. The analytical hypothesis was conservative and stated in advance; in light of saturated classical baselines on this benchmark, we did not anticipate a quantum advantage in raw discrimination, and we framed the study as a methodological probe rather than as a competition. Using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (n = 569) for development and an independent Wisconsin Original (WBC) cohort (n = 683) for external validation, we benchmarked five classical learners (XGBoost, LightGBM, CatBoost, RandomForest, RBF-SVM), two quantum models (an eight-qubit VQC implemented in PennyLane and a ZZ-feature-map QSVM implemented in Qiskit), and a stacked hybrid ensemble. The evaluation framework combined Optuna-driven hyperparameter optimisation, internal–external cross-validation, and external validation on the independent WBC cohort. Robustness and interpretability were then probed through circuit depth and embedding rotation ablation, depolarising noise stress tests, learning curve and feature stability analysis, decision curve analysis, and dual SHAP-based explanations covering both a direct tree-based explanation and a quantum surrogate. Reporting followed the TRIPOD + AI guideline. Results: On the internal test partition, RBF-SVM achieved the highest discrimination (AUC = 0.998), with XGBoost, LightGBM, CatBoost, the hybrid ensemble, and the VQC clustering between 0.992 and 0.996; the QSVM with a ZZ-fidelity kernel underperformed substantially (AUC = 0.727). Pairwise tests for correlated ROC curves indicated that most differences among top models were not statistically significant. On the external WBC cohort, model rankings reorganised, as RBF-SVM (AUC = 0.986, 95% CI 0.946–0.997), RandomForest (0.985, 95% CI 0.945–0.996), VQC (0.983, 95% CI 0.942–0.995), and the hybrid ensemble (0.982, 95% CI 0.941–0.995) all retained near-ceiling discrimination with extensively overlapping confidence intervals. Ablation analysis demonstrated that the choice of embedding rotation is decisive—Z-rotation embeddings collapsed VQC performance to chance levels (AUC ≈ 0.50), whereas X- and Y-rotations preserved it. Depolarising noise up to p = 0.10 had a negligible effect on the VQC, and SHAP analyses converged on worst concave points, mean concave points, and worst area as the dominant predictors across both classical and quantum models. Decision curve analysis showed positive net benefit for both classical and hybrid models across the clinically meaningful threshold range, exceeding both the treat-all and treat-none reference strategies throughout. Conclusions: In the present regime, the principal contribution of QML is not raw discrimination—modern classical learners are already at the data ceiling—but the construction of a rigorous, reproducible, externally validated, and interpretable benchmarking framework in which quantum models can be fairly compared with their classical counterparts. Because evaluation was confined to curated benchmark datasets rather than real-world clinical populations, the interpretability and net benefit findings reported here should be read as benchmark-level evidence and not as a demonstration of readiness for clinical deployment. Full article

Wide-area visual monitoring of construction sites is constrained by the reliable detection of safety-critical targets that appear small, low-resolution, and weakly textured under elevated or distant camera views. To address this problem, this study proposes Scale-Gated Edge Detection Transformer (SGE-DETR), a safety-oriented end-to-end detector for large open construction scenes. The framework incorporates scale-aware residual edge modulation to preserve weak contours and local structures, density-guided context-adaptive fusion to balance multi-level features according to contextual and edge-density responses, and spatial gated reparameterized feature refinement to suppress redundant background textures. Experiments were conducted on SODA and STWD using COCO-style scale-sensitive metrics and efficiency indicators. On SODA, SGE-DETR achieved AP50, APS, APM, and APL values of 0.8748, 0.2157, 0.4577, and 0.6013, respectively, with 32.5 GFLOPs, 14.5 M parameters, and 83.4 FPS. On STWD, it obtained the highest AP50, APS, APM, and APL among the compared methods, reaching 0.7936, 0.8132, 0.8594, and 0.9253, respectively. Ablation results further showed that the full model improved mAP50 and mAP50–95 over RT-DETR-r18 by 4.15 and 2.93 percentage points while reducing computational complexity. These results indicate that SGE-DETR improves safety-oriented small-object detection and multi-scale robustness while retaining a relatively low parameter count. Full article

Background: The human visual system can extract object-category information from extremely brief visual input, and animate categories often show behavioral advantages over inanimate categories in rapid categorization, visual search, and change-detection tasks. Motivated by these behavioral findings, the present study asked whether, at the representational level, animate categories elicit more category-selective neural patterns than inanimate categories in the ventral visual cortex under brief input. Methods: Using fMRI and correlation-based multivoxel pattern analysis (MVPA), we examined whether activity patterns elicited by animate categories in the ventral occipitotemporal cortex under a 33 ms brief-presentation condition corresponded more selectively to same-category patterns under a 500 ms extended-viewing condition than did patterns elicited by inanimate categories. During scanning, participants viewed animate and inanimate stimuli, each comprising four basic-level subcategories, and performed a noise-detection task that did not require explicit category judgments. Results: Across multiple ROI-definition strategies, animate categories showed significantly higher cross-duration category information than inanimate categories. This effect also remained significant after excluding the human-head category, which contained human-face information. Stimulus-level image-feature analyses further showed that within-category visual homogeneity explained part of the variance in cross-duration category information, particularly in the full stimulus set that included human heads. However, a composite visual homogeneity index derived from HOG, Gabor, and ResNet50 features did not fully account for the higher cross-duration category information observed for animate categories. Conclusions: Overall, these results suggest that, when visual input is highly limited, animate categories elicit VOTC multivoxel patterns that correspond more selectively to same-category patterns under extended viewing. Full article

Hypergraph neural networks have shown strong potential for node classification due to their ability to capture high-order relationships and multi-granularity structural patterns. However, real-world hypergraphs are often sparse, which limits interaction modeling through node–hyperedge incidence and, in turn, weakens reliable attribute propagation and global dependency capture. To address this issue, we propose DDEC, a Dual Dependency-Enhanced Contrastive learning framework for sparse hypergraph node classification. To compensate for relational information lost under sparse structures, DDEC introduces an attribute view to complement the structural view. Since attribute information can be noisy and unreliable, we first design an entropy-guided feature recalibration mechanism to estimate node uncertainty and emphasize trustworthy attribute interactions. Building upon this, DDEC performs dual dependency enhancement from both structural and attribute perspectives. Specifically, we exploit the duality between a hypergraph and its line graph to perform line-graph transformation in both views, thereby constructing a shared dual relational space for interaction enhancement under sparse topologies. Within this dual space, we perform attention-based dependency enhancement in both views, so that the structural view captures explicit topological dependencies among hyperedges, while the attribute view uncovers latent semantic correlations beyond sparse incidence relations. The resulting representations from the two views are then adaptively fused, and collaborative contrastive learning is further performed at both the node and hyperedge levels to enforce multi-granularity semantic consistency. Experiments on eight public datasets demonstrate that DDEC consistently outperforms competitive baselines, validating its effectiveness and robustness. Full article

Accurate extraction of transmission towers from Unmanned Aerial Vehicle (UAV) Light Detection and Ranging (LiDAR) point clouds is a prerequisite for overhead transmission line (OTL) acceptance. This task remains challenging because tower points are heavily entangled with ground, vegetation, conductors, and insulators, especially in complex terrain. To address this issue, we propose a shape prior-guided coarse-to-fine framework for tower extraction from UAV LiDAR point clouds. First, candidate tower regions are localized from the scene point cloud through preprocessing, near-ground suppression, and density-based clustering. Second, the least-disturbed central body of each candidate tower is identified in a slice-wise manner and used to estimate the tower orientation and four principal structural axes. Third, side-view and front-view structural envelopes are progressively inferred to suppress non-tower points around the tower body and tower head. Finally, a base-constrained filtering strategy is introduced to remove residual ground and low-vegetation points within the tower footprint. Experiments conducted on multiple OTL datasets acquired in different regions of China, including plains and mountainous areas, demonstrate that the proposed method achieves robust and efficient tower extraction across diverse scenarios. The results indicate that explicit structural priors offer a promising complement to feature-driven and data-intensive approaches, particularly in scenarios with limited annotated data and strict real-time requirements. The proposed method processes scene point clouds containing tens to hundreds of millions of points, with an average extraction time of approximately 100 to 300 s per tower depending on scene density. Full article

Rotating machinery is a key component in modern industry, and its operating condition directly affects equipment safety and production reliability. However, discrepancies among different machines cause source–target distribution shifts, while fault annotation for target machines is costly, limiting the performance of deep learning-based diagnosis under cross-machine scenarios with limited labels. To address these issues, this paper proposes a multi-scale time–frequency semantic consistency model based on self-supervised transferable representation learning, termed MTFSC. First, augmented waveform views and multi-scale frequency-domain views are constructed from unlabeled source-domain vibration signals for self-supervised pre-training without source labels. Then, a time-domain impulse-aware feature extractor and a time–frequency decoupled spectral feature extractor are designed to enhance local impulsive responses and emphasize fault-sensitive time–frequency patterns. Furthermore, a semantic-aware soft contrastive loss is developed to mine potential semantic neighbors from multi-scale frequency-domain structural similarity, reducing false-negative effects in conventional hard-label contrastive learning. Finally, the pre-trained time-domain extractor is transferred to the target machine and fine-tuned with limited labeled samples. Experimental results show that MTFSC outperforms comparison methods under different labeled sample ratios and achieves an average accuracy of 97.5% across four cross-machine diagnostic tasks. Full article

The high dropout rate in massive open online courses (MOOCs) continues to limit their potential in promoting inclusive and sustainable learning. Although many prediction models have been used to identify potential dropouts, most studies view dropout as a static classification problem and fail to clearly reveal the dynamic trajectory of learner participation over time. Therefore, this study introduces a phased analysis perspective, treating MOOC dropout as a process that continuously evolves at different stages. On the basis of the KDDCUP2015 dataset, we constructed behavioral characteristics at three time points: the first week, the third week, and the fifth week. By combining robust feature analysis and interpretable models, we systematically examined the changing patterns of dropout modes. The results revealed significant differences across the different stages. In the early stage of the course, dropout was related mainly to the unstable interaction behaviors of learners, such as restricted access to resources and irregular participation rhythms. In the middle and late stages, task-oriented behaviors, especially those related to video-based learning activities, gradually became key factors. Notably, high-frequency video participation does not always reduce the risk of dropout; when video activity is high but the overall interaction rate is low, it is more likely to indicate an increase in the risk of dropout. These results indicate that the combination of behaviors is more crucial than mere activity levels. By revealing the changing characteristics of behaviors at different stages, this study helps support the design of more practical early warning methods. Full article

Adversarial camouflage has attracted growing research attention owing to its ability to execute multi-view, persistent attacks in real physical environments, outperforming conventional single-view adversarial patches. However, most existing methods are confined to non-targeted attacks, which induce arbitrary incorrect detection results without specifying target categories. This ambiguity weakens attack destructiveness and stealthiness, posing limitations for security evaluation of real-world vision systems. To address this gap, we present TACT, an approach built upon the full-coverage physical camouflage pipeline. By replacing the original category supervision with a predefined target class, TACT redirects the optimization gradient to guide 3D texture toward the target category features. Such a scheme only employs the inherent feature alignment mechanism of off-the-shelf object detectors, without redesigning network modules, defining novel loss functions, or modifying the rendering pipeline. Extensive experiments across digital and physical domains validate its effectiveness: on seven mainstream general-purpose object detectors, TACT-person achieves an average targeted attack success rate of 51.91%, and delivers cross-architecture and cross-version transferability. In physical tests, TACT-bird reduces mAP50-95 by 59.87% on YOLOv8, yet a TCER–TASR gap suggests that the physical pipeline acts as a low-pass filter: coarse-grained target classes transfer robustly while fine-grained ones suffer feature collapse. These results confirm the viability of native supervision redirection and reveal an empirical pattern: coarse-grained target classes transfer more robustly through the physical pipeline than fine-grained ones, suggesting that target class feature granularity consistently influences physical-domain attack effectiveness. Full article

In precision livestock farming, 3D point clouds provide important data support for analyzing pig behavior and monitoring their health. However, due to environmental occlusions, limited sensor viewpoints, and mutual shielding between pigs, the acquired point clouds are often severely partial, which affects the accuracy of body shape modeling and behavior recognition. To address these challenges, this study constructed a pig pose point cloud dataset using multi-view depth camera acquisition and point cloud registration techniques. Based on this dataset, an improved point cloud completion model, IIR-PoinTr, is proposed to enhance the reconstruction of geometric and topological structures in pig bodies. By strengthening local geometric perception and high-dimensional feature representation, the model improves the reconstruction quality of partial pig point clouds and produces more structurally consistent pig body shapes. Experimental results show that, on the self-constructed pig posture dataset, the proposed method reduces Chamfer Distance (CD-L1) by 3.6%, CD-L2 by 6.9%, and Earth Mover’s Distance (EMD) by 2.0%, while improving the F-score by 5.4% compared with the baseline model. In single-view point cloud completion tasks, the method is capable of reconstructing geometrically consistent pig body structures and increases downstream classification accuracy by 34.9%. These results indicate that the proposed method can improve the reconstruction quality of partial pig point clouds and provide preliminary technical support for posture analysis under occlusion. Full article

Compared with hardware-dependent methods, light field (LF) reconstruction algorithms enable a more economical and convenient acquisition of densely sampled LF (DSLF). Existing learning-based LF reconstruction methods suffer from limited flexibility, as they rely on fixed sampling patterns and predefined angular resolutions. In this paper, we propose a flexible deep learning framework, which can reconstruct DSLF with arbitrary angular resolution from randomly distributed sparse input views of an arbitrary quantity. The proposed framework consists of two core stages, namely the SAI Synthesis and the LF Refinement. The SAI Synthesis adopts Plane Sweep Volume (PSV) to cope with randomly sampled input views, and leverages the Multi-Scale Attention (MSA) module to compute per-view weights for adaptive feature fusion and support arbitrary numbers of input views. The LF Refinement stage integrates intermediate results and fully exploits LF parallax structures to further improve reconstruction quality. Experimental results demonstrate that our method achieves superior flexibility and reconstruction quality, and outperforms most state-of-the-art LF reconstruction methods. Full article

Quantitative structure–activity relationship and quantitative structure–property relationship (QSAR/QSPR)-based molecular toxicity prediction provides an in silico strategy for prioritizing environmental contaminants when longer-duration bioassay data are sparse. However, many Simplified Molecular-Input Line-Entry System (SMILES)-based machine learning models treat exposure duration as an unconstrained numerical covariate and provide limited information on whether predictions are supported by the observed temporal domain. Here, we evaluated an applicability-domain-aware chemoinformatics framework that combines transformer-derived molecular representations with toxicokinetic-informed temporal encoding and evidential uncertainty estimation. The approach replaces conventional ${log}_{10}$-transformed time encoding with a bounded first-order toxicokinetic saturation feature and combines this representation with Deep Evidential Regression to support a joint chemical–temporal view of the QSAR/QSPR applicability domain. Using experimentally derived U.S. EPA Ecotoxicology Knowledgebase (ECOTOX) fish EC50 mortality records, models were trained on 48,728 acute-duration observations and evaluated retrospectively on 2090 temporally separated longer-duration observations. The combined toxicokinetic and evidential model reduced temporal extrapolation error relative to conventional time encoding while maintaining comparable within-domain validation performance. The learned population-level timescale converged to 221 ± 3 h, consistent with accumulation timescales extending beyond standard acute fish test durations. Epistemic uncertainty was positively associated with absolute prediction error across all 10 folds, suggesting that the uncertainty estimates retained sample-level information relevant to applicability-domain-aware molecular toxicity screening. Cross-species analyses further showed that model behavior depended on training time coverage, with greater convergence when available assays covered a larger fraction of the learned timescale. These results suggest that toxicokinetic-informed temporal encoding can improve uncertainty-aware QSAR/QSPR modeling of environmental contaminant toxicity and support prioritization of compounds for further testing, while complementing rather than replacing chronic bioassays. Full article

Background/Objectives: Parkinson’s disease (PD) is increasingly recognized as a multisystem disorder affecting perceptual, emotional, and reward-related processes. While arts-based interventions in PD have primarily focused on active creative arts engagement, it remains unclear whether receptive arts engagement with visual art—how artworks are perceived and evaluated—is altered. Our objective is to determine whether aesthetic evaluation of visual artworks differs in individuals with PD compared to age-matched healthy controls. We further examine whether emotional interpretation, color-emotion associations, and experiential responses to art viewing are altered. Methods: In a cross-sectional case–control study, individuals with PD (n = 87) and age-matched healthy controls (n = 49) completed two online assessments. Participants evaluated 36 artworks from the Vienna Art Picture System in terms of liking, beauty, and subjective art attributes. Objective image-derived features were computed for each artwork. Interpretable machine learning models were used to test whether evaluation patterns predicted diagnostic group and to identify determinants of aesthetic judgments. Participants further completed a color-emotion association task using ambiguous expressive portraits and reported perceived changes in cognitive, emotional, motivational, and physical states following art viewing. Results: Aesthetic evaluation patterns did not support reliable classification of PD status, indicating no systematic group differences in liking, beauty, or attribute-based judgments between PD and controls. Instead, aesthetic judgments were robustly predicted by individual differences and objective artwork properties, including art-historical style, symmetry, complexity, and color-related features, whereas diagnostic group, gender, and age did not contribute to predictions. Emotional interpretation and color-emotion associations were largely comparable between groups, with a single specific deviation in color-emotion mapping. Positive emotions were less frequently associated with pink in people with PD. Self-reported experiential responses to art viewing did not differ significantly between groups. Conclusions: Aesthetic evaluation of visual artworks appears largely preserved in people with PD. These findings suggest that, in digital viewing contexts, substantial adaptation of visual content to make it accessible for people with PD may not be necessary, although subtle perceptual and emotional differences may still be relevant. Efforts may instead be better directed toward addressing practical barriers to visual art engagement. Full article

This paper extends a previous classification study by examining clustering methods on the same synthetic datasets and comparing their behavior with the previously obtained classification results. This study investigates the performance of selected time series clustering methods under controlled changes in noise level and class complexity. Six clustering methods representing distance-based, feature-based, and deep learning approaches were evaluated on 82 balanced synthetic datasets. The datasets contained from two to six classes, different levels of additive Gaussian noise, 200 time series per dataset, and 1000 observations per time series. The analysis focused on clustering quality, comparative behavior with classification models, and computational cost in terms of training time and peak memory usage. Clustering quality was assessed mainly using Adjusted Rand Index and V-measure, while accuracy after Hungarian label matching was used as an auxiliary measure for comparison with classification models. The results show that distance-based methods, and particularly TimeSeriesKMedoids, achieved the most robust and consistent clustering performance across the considered settings. Clustering quality decreased with both the number of classes and the noise level, but the effect of noise was clearly stronger. Feature-based and deep learning-based clustering methods were generally more sensitive to noise, while deep models were also associated with substantially higher computational cost. In terms of memory usage, classical clustering methods remained below 50 MiB, whereas deep learning-based clustering methods required substantially more memory. This study further shows that accuracy computed after Hungarian label matching may provide an overly optimistic view of clustering quality. Accuracy after Hungarian label matching is reported only as an auxiliary metric, while the main interpretation of clustering quality is based on structure-sensitive measures such as Adjusted Rand Index and V-measure. Overall, the findings highlight the importance of robust distance-based approaches and of using structure-sensitive evaluation measures when analyzing time series clustering. Full article