Search Results (12,794)

Background: Post-surgical cardiovascular monitoring places a heavy information burden on clinical teams, requiring the rapid synthesis of patient history, intraoperative data, monitoring streams, and surgical outcome evidence. Existing clinical decision support systems handle this integration poorly, and most offer little visibility into their reasoning. We present a Retrieval-Augmented Generation (RAG) architecture designed specifically for this domain, with a focus on evidence traceability and practical workflow integration. Methods: We describe a three-layer RAG architecture comprising a retrieval layer that creates 768-dimensional representations of clinical scenarios; an augmentation layer using a stacking ensemble (Random Forest and XGBoost base learners with a logistic-regression meta-learner) to integrate patient-specific data with retrieved evidence and produce calibrated probability estimates; and a generative layer using a fine-tuned BERT classifier together with Gemini 2.5 Pro to synthesise actionable clinical recommendations. Components were prototyped on publicly available, de-identified data from MIMIC-III and the MIMIC-III-Ext-PPG benchmark to verify pipeline integrity. Proposed Evaluation Framework: This paper presents a system architecture rather than a clinically validated implementation. We outline a structured evaluation framework to assess the technical performance and clinical applicability of the RAG architecture, encompassing the technical validation of system components, expert assessment of clinical workflow integration potential, and analysis of interpretability features essential for healthcare deployment. Specific technical targets include retrieval precision >90% for relevant evidence, query response time <3 s, and a clinical appropriateness rating of >85% from expert review. Conclusions: We describe a RAG architecture for post-surgical cardiovascular monitoring in which every recommendation is linked to retrievable source documents, making the reasoning visible and challengeable. A structured evaluation framework is proposed to guide the system towards clinical validation. Full article

Deep learning models are widely used in autonomous vehicle systems for perception, localization, and decision-making. However, their lack of transparency poses significant challenges in safety-critical environments. This systematic review presents a unified mathematical framework for explainable deep learning which integrates multimodal inputs, graph-theoretic road geometry, uncertainty modeling, and intrinsically interpretable representations. Road-structured priors that include lane topology and spatial constraints are incorporated into learning and optimization processes for ensuring model predictions and explanations to remain physically and semantically grounded. The review synthesizes methods across saliency-based, concept-based, causal, and intrinsic explainability, and extends them to vision-language models. This enables language-grounded, human-interpretable reasoning in autonomous vehicle systems. While vision-language models offer a new paradigm for semantic explainability, their limitations such as hallucinations, misgrounding, and reduced reliability under distribution shifts are also critically examined. Along with the role of road priors in improving alignment and robustness, another key contribution of this work is its quantitative evaluation metrics for road-aware explainability. These evaluation metrics link the explanations to spatial consistency, uncertainty alignment, and graph-structured reasoning. The overall framework connects latent representations, predictions, and explanations within a single formulation, enabling systematic comparison and analysis across models. Based on a PRISMA-guided review of 164 studies, this research identifies gaps in real-world reliability, temporal reasoning, and standardized evaluation, and outlines future directions including human-in-the-loop systems, regulatory readiness, and language-based auditing. Overall, this study advances a mathematically grounded and road-aware perspective on explainable vehicle AI which significantly bridges the gap between high-performance models and transparent, trustworthy autonomous systems. Full article

This paper studies subset construction in NP-complete problems from the perspective of structured exploration of combinatorial search spaces. Classical approaches rely on exhaustive enumeration of subsets, which leads to exponential growth in time and memory requirements. To address this limitation, we introduce an indexed framework based on the correspondence between a finite set and its associated index set. Within this framework, subsets are represented as ordered index sequences, allowing subset construction to be reformulated as a constraint-guided search process over index space. Candidate subsets are characterized by numerical descriptors derived from their indices (referred to as index certificates), which guide and filter the construction process. Subset generation is further organized through admissible index intervals that restrict feasible transitions and reduce the effective search space. The framework is based on an index-based representation and structured traversal of pairwise index combinations. Computational experiments on representative instances illustrate the behavior of the indexed construction procedure and indicate its efficiency relative to classical enumeration-based methods for small and medium-sized instances. The proposed approach provides a structured perspective on combinatorial search and offers a basis for further development of algorithms based on constrained exploration of subset structures. Full article

Accurate crop row detection is identified as a critical prerequisite for autonomous agricultural navigation, yet it remains challenging in complex field environments. To achieve a balance between segmentation accuracy, robustness, and real-time performance, an improved crop row segmentation and navigation method based on the DeepLabV3+ framework was developed. MobileNetV2 was adopted as the backbone to minimize computational costs, while feature representation was enhanced through integrated attention mechanisms and multi-scale fusion. Specifically, split-attention convolution was integrated into the backbone, a DenseASPP + SP module was employed for multi-scale contextual capture, and a Convolutional Block Attention Module (CBAM) was added to refine feature responses. Experimental results demonstrated that the proposed method outperformed mainstream models, achieving a mean Intersection over Union (mIoU) of 93.42% and an f1-score of 96.8%. The model maintained a lightweight architecture with 8.35 M parameters and a real-time speed of 32 FPS. Furthermore, crop row anchor points were extracted and processed via DBSCAN clustering and RANSAC fitting to generate high-precision navigation lines. Validation showed that the middle crop row yielded the highest fitting accuracy with minimal angular and lateral errors. This study provides an efficient visual perception solution for intelligent field operations. Full article

Optical and Synthetic Aperture Radar (SAR) sensors capture complementary and consistent information, and their fusion enhances remote sensing image quality. Existing pyramid decomposition-based methods suffer from insufficient texture–edge discrimination. Additionally, the manual setting of parameters during pyramid decomposition introduces uncertainty in the fusion results. To address this problem, we propose an optical and SAR image fusion framework based on local extrema adaptive pyramid decomposition (LEAPFusion), which enhances edge preservation and improves parameter adaptability. Specifically, by leveraging the edge-preserving properties of local extrema, we introduce them into the image pyramid decomposition framework to construct complementary local extrema and Laplacian pyramids. Then, we introduce an explicit parameter adaptation strategy in which the decomposition levels and local extrema kernel sizes are automatically determined from image size and pyramid scale, enabling consistent multi-scale representation and reducing parameter sensitivity compared to empirically tuned settings. Finally, by exploiting the complementary properties of the two pyramids, we implement a multi-type fusion strategy: weighted averaging for low-frequency components and parameter-adaptive pulse-coupled neural network (PAPCNN) for high-frequency details. Our decomposition framework seamlessly integrates three representative edge-preserving filters—a median filter, a guided filter, and a rolling guidance filter—demonstrating strong generalization capability across different filtering paradigms. Extensive experiments on two benchmark datasets demonstrate that our method outperforms seven state-of-the-art algorithms, achieving the best results across diverse scenes with improvements of up to 13.38% in SF and 18.90% in SCD compared to the second-best methods. Full article

Scene graph generation (SGG) provides a structured representation for visual understanding. However, most existing methods are designed to optimize global triplet recall rather than retrieve relation instances specified by a user query. In query-driven visual relation retrieval, two major challenges arise: the target relation must compete with a highly redundant candidate space, and query semantics are not incorporated before relation classification. To address these challenges, we propose a Query-Driven Candidate Relation Screening (QCRS) module, which injects query semantics into the candidate screening process. Specifically, QCRS encodes the query and candidate visual relation features, and then filters query-relevant candidates through relevance scoring. By reducing interference from irrelevant candidates and avoiding redundant computation, QCRS improves the final exact triplet hit performance and enhances the interpretability of query-specific relations, thereby facilitating query-driven visual relation retrieval. Built upon the strong EGTR baseline, QCRS learns query relevance to prioritize relation instances matching the target query, enabling precise triplet retrieval. Extensive ablation studies and analyses on the VG150 benchmark validate the effectiveness of the proposed approach: when integrated with EGTR, QCRS improves PairR@50 from 61.52% to 80.06% and ETR@50 from 30.54% to 47.07%, achieving absolute gains of over 16 percentage points in both correct object pair retention and end-to-end target relation retrieval performance. Full article

Accurate direction-of-arrival (DOA) estimation with high resolution is fundamental to many array sensing applications. In practice, however, sparse arrays with missing sensors and snapshot-limited observations often lead to incomplete and noisy second-order statistics, which substantially degrades the performance of conventional eigendecomposition-based estimators. In this paper, we propose a two-stage Transformer framework for sparse-array DOA estimation that explicitly separates correlation recovery from angle inference. The first stage operates in the correlation domain and learns to reconstruct a clean and complete correlation vector from partially observed measurements using masking-aware tokenization and global-context modeling. The recovered representation can be further converted into a structured covariance matrix, providing an interpretable interface to classical signal processing back-ends. Based on the recovered features, the second stage adopts a Transformer regressor to directly predict multi-source DOAs. Extensive simulations on a large-scale dataset with SNRs from −5 to 10 dB and various snapshot numbers demonstrate that the proposed method delivers robust accuracy and improved stability in low-SNR and snapshot-limited regimes, while maintaining competitive performance at higher SNRs. Additional evaluations with an ESPRIT back-end further confirm that the recovery-based covariance yields more reliable DOA estimation than conventional difference–coarray processing, with particularly evident gains under challenging noise conditions. Full article

Seismic inversion is a key technique for quantitative characterization of subsurface elastic parameters and detailed reservoir description. However, due to the limited bandwidth of seismic signals and the strong heterogeneity of complex reservoirs, conventional inversion methods struggle to simultaneously achieve high vertical resolution and lateral continuity. To address these challenges, an intelligent elastic parameter inversion method based on kernel density estimation within a Bayesian framework is proposed. First, kernel density estimation is introduced to augment the training samples, thereby alleviating data scarcity. Second, a hybrid architecture integrating convolutional modules, Mamba, and cross-attention mechanisms is constructed to achieve collaborative modeling of local spatial features and long-range temporal dependencies. The cross-attention mechanism is further employed to adaptively weight and fuse multi-source features, thus enhancing the representation capability of the model. Subsequently, by designing a joint loss function, the strengths of deterministic inversion and data-driven approaches are effectively integrated, ensuring physical consistency while enhancing data adaptability, thereby improving the stability and accuracy of the inversion results. Furthermore, the neural network outputs are used as the initial model for Bayesian inversion to construct a probabilistic inversion framework for elastic parameter inversion. Finally, experimental results demonstrate that the proposed method improves the R² values of inversion results by more than 8.0% and 5.0% compared with conventional methods in thin interbedded models and real data experiments, respectively. Full article

Change detection in open-pit mining areas from remote sensing imagery is of great importance for mining supervision, ecological monitoring, and restoration planning. Nevertheless, mining-related changes usually exhibit multi-scale patterns, irregular boundaries, and fragmented spatial distributions, which make accurate detection difficult. Existing CNN- and Transformer-based methods often cannot effectively balance global context perception and local detail preservation, resulting in incomplete boundary extraction and insufficient sensitivity to subtle changes. To overcome these limitations, we propose GLMECD-Net, a Global–Local Multi-scale Cross-fusion Enhanced Change Detection Network for remote sensing image change detection in open-pit mining areas. Specifically, a Siamese encoder is used to extract hierarchical bi-temporal features, while a Global–Local Feature Mixing Embedding (GLME) module is introduced to jointly capture long-range contextual information and local spatial details. Furthermore, multi-scale feature aggregation and cross-temporal feature fusion are employed to improve change representation and boundary recovery. Experimental results on mining area datasets show that the proposed method achieves 71.66% Precision, 83.78% OA, 77.53% F1-score, and 53.82% IoU. The results demonstrate that GLMECD-Net provides effective and robust performance for detecting complex and subtle changes in open-pit mining areas. Full article

High-resolution images of non-cooperative spacecraft are essential for on-board autonomous operations. Hardware bandwidth limits and continuously changing observation distances mean that a practical super-resolution (SR) system must handle arbitrary, non-integer magnification factors without retraining, a setting known as arbitrary-scale SR (ASSR). Recent 2D Gaussian splatting (2DGS) methods represent image content with explicit anisotropic Gaussian primitives and render at any continuous coordinate, offering substantially faster inference than implicit neural representation (INR) approaches. Yet spacecraft imagery presents a structural mismatch for uniform 2DGS regression: the target occupies a small, densely structured region within a vast, featureless deep-space background, so a network that minimizes average reconstruction loss inevitably over-invests capacity in the irrelevant background and smears the fine edges of antennas and solar panels. We propose S³R-GS, a saliency-guided framework that embeds semantic spatial priors into the 2DGS pipeline at three levels: an encoder-level module that suppresses background noise before it reaches the splatting stage; a discrete Gaussian routing mechanism that assigns each spatial location to a semantically appropriate kernel group and reformulates Gaussian modeling as semantic prototype selection; and a saliency-weighted training strategy that concentrates the optimization gradient on the spacecraft target. Experiments on the SPEED and SPEED+ benchmarks show that S³R-GS achieves strong PSNR performance, competitive SSIM, and improved perceptual quality across scale factors from $\times 2$ to $\times 12$; additional ablation, extreme-lighting, and efficiency analyses further support the robustness and practicality of the proposed design. Full article

Traffic accidents pose a persistent challenge for urban mobility, public safety, and sustainable development in smart cities, particularly in rapidly growing urban environments. This study presents a data-driven spatiotemporal analysis of traffic accidents in Ecuador, aimed at supporting evidence-based urban traffic management and road safety planning. Using large-scale historical accident records, the proposed approach combines spatial clustering and temporal forecasting techniques to characterize accident concentration patterns and temporal dynamics at national and metropolitan scales. Spatial accident hotspots are identified using Density-Based Spatial Clustering of Applications with Noise (DBSCAN), enabling the detection of high-risk zones without imposing assumptions on cluster shape or size. This analysis reveals strong spatial concentration of accidents, with a limited number of clusters accounting for a substantial proportion of fatalities and injuries. Complementary temporal analysis is conducted using a multi-model ensemble framework to examine accident trends and seasonal patterns. This approach integrates SARIMA for linear stochastic modeling and Prophet for additive trend analysis, alongside two Long Short-Term Memory (LSTM) architectures: a direct 12-month vector output and a recursive horizon-3 model. By synthesizing these statistical and neural network-based methods through inverse-RMSE weighting, the study captures both stable seasonal cycles and non-linear, short-to-medium-term variations in accident frequency. Results show that traffic accidents in Ecuador exhibit stable diurnal and seasonal structures, alongside pronounced spatial heterogeneity across urban regions. The combined spatial and temporal insights provide a coherent representation of accident risk patterns, facilitating the prioritization of critical zones and high-risk periods. The resulting hotspot maps and multi-model forecasting horizons offer actionable information for smart city stakeholders, supporting targeted infrastructure interventions, adaptive enforcement strategies, and data-informed urban mobility policies. This work contributes to the broader understanding of traffic safety analytics as a core component of smart city decision-support systems. Full article

Real-time simulation plays an important role in the development and verification of modular multi-level converter (MMC) systems, especially for the rapid and low-risk evaluation of control and protection functions in medium- and high-voltage applications. However, MMC validation often requires simulation models with different fidelity levels for different testing purposes, while detailed device-level representation further imposes stringent constraints on computational efficiency. To address these issues, this paper develops a multi-level real-time modeling framework for MMCs, in which switch models of different accuracy can be incorporated within a unified architecture and flexibly selected according to the target test scenario. On this basis, a device-level real-time simulation method is further established to capture the nonlinear switching transients of MMCs under the proposed framework. By combining network decoupling with FPGA-oriented implementation, the framework can achieve a minimum simulation step of 50 ns under fully parallel hardware allocation. Considering FPGA resource optimization, the prototype implemented in this work is validated with a 100 ns time-step. A three-phase MMC with four submodules per arm is used as the validation case and implemented on an FPGA platform. Both waveform comparisons and quantitative error analysis demonstrate close agreement between the proposed real-time model and offline reference models. In addition, closed-loop real-time experiments are conducted to further confirm the effectiveness of the developed MMC model in realistic real-time simulation-based testing applications. Full article

Music emotion recognition (MER) can be formulated as a multimodal optimization problem that predicts an emotion label from coupled audio and lyric sequences. Existing methods typically perform unimodal learning or coarse global fusion, which overlooks fine-grained temporal-token correspondences between musical dynamics and lyric semantics. We propose MELT (Music Emotion Learning with Temporal token-level fusion), an optimization-driven framework with four modules: a BERT-based lyrics semantic encoder (LSE), a segment temporal encoder (STE) that models audio-segment dependencies via a Transformer, a token-level temporal fusion (TTF) module with gated cross-attention, and an emotion mood head (EMH) for four-class prediction. Training is conducted end-to-end by jointly minimizing a supervised classification term and an auxiliary cross-modal contrastive alignment term, yielding a unified objective that improves both class separability and representation consistency. On the MoodyLyrics benchmark, MELT achieves 87.6% weighted F1 for four-class emotion recognition (angry, happy, relaxed, sad), outperforming unimodal baselines and representative early/late fusion strategies. Ablation results further verify that temporal encoding, gated token-level fusion, and joint optimization each contribute to the final performance. Full article

With the increasing scale and complexity of network traffic, anomaly detection faces significant challenges, particularly under the scarcity of labeled data in real-world environments. Although graph neural networks (GNNs) effectively model relational structures, most existing approaches rely on supervised learning, limiting their applicability in weakly labeled or unlabeled scenarios. To address these limitations, this paper proposes a self-supervised graph neural network framework, termed EGSCA, for network traffic anomaly detection. The framework employs a GNN to jointly model node and edge information, enabling the learning of discriminative representations. On this basis, a graph contrastive learning strategy is designed, where diverse subgraphs are generated via breadth-first search (BFS) to effectively capture local structural patterns. Meanwhile, a hybrid contrastive loss based on Wasserstein distance and Gromov–Wasserstein distance is introduced to achieve collaborative optimization between feature-space alignment and structural consistency under unlabeled conditions. Experimental results on multiple benchmark datasets demonstrate that the proposed method achieves competitive performance. Notably, it achieves the best results on datasets NF-BoT-IoT and NF-BoT-IoT-v2, with average improvements of approximately 3.2% in F1-score and 1.7% in DR over the strongest baseline. Further analysis indicates that the model yields more pronounced performance gains in scenarios with high class separability. Full article

Underwater image enhancement is critical for marine robotic perception, yet existing methods often face a persistent trade-off between restoration quality, structural reliability, and deployment efficiency. Although lightweight enhancement networks are attractive for resource-constrained underwater platforms, many of them mainly rely on empirical architectural simplification and appearance-oriented objectives, with limited mathematical analysis of complexity reduction, semantic regularization, and optimization coordination. To address this issue, this paper proposes MobileGAN, a lightweight underwater image enhancement framework equipped with dual-reference regularization and a theoretical analysis module. The proposed generator adopts a compact encoder–bottleneck–decoder architecture based on depthwise separable convolutions, which substantially reduces convolutional redundancy while preserving effective restoration capability. A dual-reference feature consistency formulation is introduced to simultaneously constrain the enhanced image toward the high-quality target representation and the degraded-input semantic anchor. In addition, an edge-aware regularization term and a stage-wise dynamic weighting mechanism are incorporated to improve local structure recovery and multi-objective optimization behavior. Beyond architectural design, we provide a mathematical analysis of the proposed framework from three aspects: computational complexity reduction, geometric interpretation of dual-reference regularization, and piecewise optimization properties of stage-wise weighted training. Extensive experiments on the UIEB benchmark demonstrate that MobileGAN achieves a favorable trade-off between enhancement quality and computational efficiency. The proposed method maintains real-time inference with a compact model size while providing competitive structural consistency and detail restoration. These results indicate that MobileGAN is not only a practical deployment-oriented enhancement framework but also an interpretable optimization model with analyzable structural properties. Full article