DSpace collection: 期刊論文

Toward deployment-oriented long legal document classification: a segmentation-based framework for distributed legal evidence integration

Thu, 14 May 2026 04:05:13 GMT

title: Toward deployment-oriented long legal document classification: a segmentation-based framework for distributed legal evidence integration

A Local Hierarchical LLM Framework for Privacy-Preserving Memory Forensics of Cryptocurrency Wallets

Thu, 16 Apr 2026 04:05:20 GMT

title: A Local Hierarchical LLM Framework for Privacy-Preserving Memory Forensics of Cryptocurrency Wallets abstract: Cryptocurrency-related crime continues to expand worldwide. Chainalysis reports that the global value of illicit cryptocurrency transactions has exceeded USD 50 billion, underscoring an urgent need for more advanced digital forensics. Cryptocurrency investigations are particularly difficult when critical evidence resides in volatile memory. Traditional workflows are time-consuming and heavily manual. Cloud-based large language models (LLMs) also pose unacceptable privacy risks in confidential law-enforcement investigations. This paper presents a multi-layer reasoning framework that integrates LangChain with a locally deployed LLM (LLaMA 3.1-8B). The framework acquires volatile evidence via memory dumping, extracts forensic artifacts via keyword and regular-expression search, and performs three-stage reasoning with a Single-Layer Baseline Architecture, a Dual-Layer Supervisor Architecture, and a Tri-Layer RAG-Decider Architecture. We evaluate the framework on 100 purpose-built crypto-wallet forensic questions. The Tri-Layer architecture achieves an average human-evaluation total score of 11.29, which is an 8.9% improvement over the Single-Layer baseline. It also reaches a BERT F1 score of 0.84 in automated metrics, improving by 15.1%. Notably, the local Tri-Layer system performs very close to the commercial cloud model ChatGPT-4o (only a 0.3% gap overall) and surpasses it on the reasoning dimension. These results demonstrate that local LLM deployment can effectively support memory forensics under strict confidentiality and limited compute resources. The proposed approach offers a practical, low-cost, and privacy-preserving tool for digital investigations. It also shows that optimized lightweight local models can approach the analytical quality of cloud-scale models.

Replacing Batch Normalization with Memory-Based Affine Transformation for Test-Time Adaptation

Thu, 16 Apr 2026 04:05:17 GMT

title: Replacing Batch Normalization with Memory-Based Affine Transformation for Test-Time Adaptation abstract: Batch normalization (BN) has become a foundational component in modern deep neural networks. However, one of its disadvantages is its reliance on batch statistics that may be unreliable or unavailable during inference, particularly under test-time domain shifts. While batch-statistics-free affine transformation methods alleviate this by learning per-sample scale and shift parameters, most treat samples independently, overlooking temporal or sequential correlations in streaming or episodic test-time settings. We propose LSTM-Affine, a memory-based normalization module that replaces BN with a recurrent parameter generator. By leveraging an LSTM, the module produces channel-wise affine parameters conditioned on both the current input and its historical context, enabling gradual adaptation to evolving feature distributions. Unlike conventional batch-statistics-free designs, LSTM-Affine captures dependencies across consecutive samples, improving stability and convergence in scenarios with gradual distribution shifts. Extensive experiments on few-shot learning and source-free domain adaptation benchmarks demonstrate that LSTM-Affine consistently outperforms BN and prior batch-statistics-free baselines, particularly when adaptation data are scarce or non-stationary. Keywords: batch normalization; affine transformation; LSTM; test-time adaptation; memory-based learning; domain adaptation; few-shot learning; normalization-free networks; deep neural networks; feature distribution shift

Physics-Aware Bottleneck-First Target Coverage Scheduling for Solar-Powered Wireless Rechargeable Sensor Networks

Tue, 14 Apr 2026 04:06:41 GMT

title: Physics-Aware Bottleneck-First Target Coverage Scheduling for Solar-Powered Wireless Rechargeable Sensor Networks abstract: Long-term target coverage in solar-powered wireless rechargeable sensor networks (WRSNs) is fundamentally challenged by sensing uncertainty, weather-driven energy variability, and the strong coupling between adjustable sensing ranges and energy consumption. Existing approaches often rely on simplified sensing or harvesting models, which may lead to unstable schedules and degraded coverage at vulnerable targets. This paper proposes Physics-aware Bottleneck-first Target Coverage Scheduling (PBTCS), a unified framework for sustainable target coverage in WRSNs under energy-neutral operation constraints. PBTCS integrates a physics-prior, interpretable day-ahead photovoltaic (PV) forecasting model to derive feasible and auditable energy budgets, and employs a budget-driven time partitioning mechanism to stabilize day-night operations. Based on the probabilistic sensing model, a bottleneck-first scheduling principle is introduced to explicitly prioritize the weakest space-time points, rather than optimizing average coverage metrics. To efficiently realize this objective under adjustable sensing radii, a closed-form marginal-gain decomposition and a budgeted dynamic programming scheme are developed for per-sensor schedule construction. Extensive simulations using real PV and meteorological data demonstrate that PBTCS consistently outperforms state-of-the-art methods in surveillance quality, coverage fairness, and long-term network sustainability across different seasons and network scales.

Deep Spatiotemporal Forecasting and Reinforcement Optimization for Ambulance Allocation

Wed, 11 Mar 2026 04:06:10 GMT

title: Deep Spatiotemporal Forecasting and Reinforcement Optimization for Ambulance Allocation

An Effective Learning Model for Multi-label Melon Classification based on Ensemble Learning

Wed, 11 Mar 2026 04:06:06 GMT

title: An Effective Learning Model for Multi-label Melon Classification based on Ensemble Learning abstract: Melons are a popular fruit with various textures and types. Categorizing them before sale enables consumers make informed choices and enhances product appeal. While humans can classify melons by sight, this process is highly inefficient for large quantities. Automated deep-learning agricultural systems offer solutions to reduce costs and increase productivity. Therefore, this study addresses the problem of reduced recognition accuracy caused by multiple textures in a single instance. Specifically, it proposes a multi-label classification method. Four models were trained on our dataset: a custom CNN, VGG16, InceptionV3, and a decision tree. Using voting aggregation techniques, we combined their strengths to produce multi-label outputs. Our method achieved impressive results, with an accuracy of 94%, a precision of 94%, and an F1 score of 93%. Additionally, this work introduced a specialized CNN model for melon rind recognition, further improving accuracy by integrating existing techniques with voting ensemble learning. This advances in automated agriculture and inspires future research.

MeTa Learning-Based Optimization of Unsupervised Domain Adaptation Deep Networks

Wed, 11 Mar 2026 04:06:01 GMT

title: MeTa Learning-Based Optimization of Unsupervised Domain Adaptation Deep Networks

Heterogeneous Multi-AAV Cooperation With Data Gathering, Offloading, and Wireless Charging in Wireless Sensor Networks

Wed, 11 Mar 2026 04:05:55 GMT

title: Heterogeneous Multi-AAV Cooperation With Data Gathering, Offloading, and Wireless Charging in Wireless Sensor Networks

Content-Preserving Image Style Transfer via Reversible Networks with Meta ActNorm

Tue, 10 Mar 2026 04:09:21 GMT

title: Content-Preserving Image Style Transfer via Reversible Networks with Meta ActNorm abstract: Image style transfer aims to synthesize visually compelling images by blending the structural content of one image with the artistic style of another. While arbitrary style transfer methods such as AdaIN and WCT offer flexibility, they often suffer from content distortion and style leakage, particularly in complex or cross-domain scenarios. Recent approaches like ArtFlow address these issues through reversible architectures, effectively reducing distortion and leakage while providing consistent reconstruction. However, ArtFlow’s reliance on fixed normalization parameters limits adaptability across diverse content–style pairs, motivating further improvement. In this paper, we propose ISTMAF (Image Style Transfer based on Meta ArtFlow), a scalable and adaptive reversible framework that incorporates Meta ActNorm—a meta-network that dynamically generates input-specific normalization parameters. To further improve the integration of content and style, we introduce an algebraic–geometric parameter fusion strategy in the reverse process, along with a hierarchical aligned style loss to reduce artifacts and enhance visual coherence. Experiments on MS-COCO, WikiArt, and face datasets demonstrate that ISTMAF achieves superior content preservation and style consistency compared to recent state-of-the-art methods. Quantitative evaluations using SSIM and Gram difference further confirm its effectiveness. ISTMAF provides a flexible, high-fidelity solution for style transfer and shows strong generalization potential, paving the way for future extensions in multi-style fusion, video stylization, and 3D applications.

GNN-RM: A trajectory completion algorithm based on graph neural networks and regeneration modules

Tue, 10 Mar 2026 04:09:17 GMT

title: GNN-RM: A trajectory completion algorithm based on graph neural networks and regeneration modules abstract: Data about vehicle trajectories assumes a crucial role in applications such as intelligent connected vehicles. However, missing values resulting from sensors and other factors frequently affect real trajectory data. Currently, it is challenging to utilize trajectory completion methods to generate accurate real-time results at an affordable computing cost. This paper proposes GNN-RM, a trajectory completion algorithm based on graph neural networks and regeneration modules, encompassing feature extraction, subgraph construction, spatial interaction graph, and trajectory regeneration modules. The feature extraction algorithm extracts influential data as feature vectors based on certain conditions and organizes these feature vectors into different subgraphs according to categories. The spatial interaction graph constructed through graph neural networks extracts spatial interaction features between vehicles and the environment, while the regeneration modules constructed by multi-head attention mechanisms extract temporal features of vehicles, thereby completing the missing trajectories. The experimental results demonstrate that GNN-RM can achieve higher trajectory completion accuracy with fewer input parameters than multiple baseline models.

Cooperative Multi-User Task Allocation in Social-Based Crowdsensing Platform

Tue, 10 Mar 2026 04:09:11 GMT

title: Cooperative Multi-User Task Allocation in Social-Based Crowdsensing Platform abstract: Mobile crowdsensing (MCS) technologies usher in new distributed services that utilize user-provided resources to execute tasks in ubiquitous network environments. Conventional studies mainly emphasize user recruitment for simple task allocation with neither information exchange nor collaboration between users. Our study aims to investigate the new issue of collaborative multi-user (CMU) task allocation, while many users cooperate in complex tasks that require multi-fold resources from different users. To form a group of collaborative users, we adopt the social-tie notion to represent three group types, including basic, bridging, and linked user groups, in MCS contexts. By quantifying the strength of social connections between users, we formulate the measures of service capacity and cost with respect to any user group. Then, the fittingness of each group to a task can be calculated, and then a subset of user groups can be chosen to perform CMU tasks. We propose a group-based task allocation scheme, briefly named GTA, which can evaluate the cost-effectiveness of task processing by any particular group and thus allocate appropriate groups in accordance with different task requirements. Performance results by simulation manifest that the GTA scheme is promising in sustaining lower service time with only a minor influence on service cost, as compared with two typical schemes, MinCost and GoCC.

A 3D Spatial–Spectral–Temporal Deep Regression Model for Improving Mangrove Canopy Height Estimation Through Fusion of Optimized Red-Edge Sentinel-2 Bands and Sentinel-1 SAR Data

Tue, 10 Mar 2026 04:09:08 GMT

title: A 3D Spatial–Spectral–Temporal Deep Regression Model for Improving Mangrove Canopy Height Estimation Through Fusion of Optimized Red-Edge Sentinel-2 Bands and Sentinel-1 SAR Data

Elastic-Trust Hybrid Federated Learning

Tue, 10 Mar 2026 04:08:59 GMT

title: Elastic-Trust Hybrid Federated Learning

Bi-Phase LSTM: A LSTM-Based Autoencoder Architecture for Dynamic Social Network Prediction

Tue, 10 Mar 2026 04:08:54 GMT

title: Bi-Phase LSTM: A LSTM-Based Autoencoder Architecture for Dynamic Social Network Prediction abstract: In recent years, social networks have grown in popularity, with most people actively engaging on these platforms. These networks hold valuable insights into users' values and interests, allowing us to analyse relationships between connected individuals and even predict potential friendships. However, social networks are dynamic, and their structure evolves over time. To account for this, we employed a dual approach using a bi-phase LSTM autoencoder and a bi-phase LSTM predictor. These tools capture the changing characteristics of social networks and predict future graph structures. We rigorously tested our model on three datasets and compared its performance with other models. The bi-phase LSTM consistently delivered strong results across all datasets. Additionally, the model's hyperparameters were fine-tuned to improve predictive accuracy, demonstrating its reliability in forecasting the evolution of social network structures.

CARES: A Hybrid Caregivers Recommendation System Using Deep Learning and Knowledge Graphs

Tue, 10 Mar 2026 04:08:50 GMT

title: CARES: A Hybrid Caregivers Recommendation System Using Deep Learning and Knowledge Graphs abstract: Recommendation systems have prospered by leveraging user-item interactions and their features for personalized recommendations. Recent advancements in deep learning further enhance these recommendation systems with powerful backbones for learning from user-item data. However, solely depending on these interactions often leads to the cold-start problem, where items lacking historical data cannot be effectively recommended. Additionally, the issue of high similarity between user and item features frequently goes unresolved. This paper introduces a Hybrid Caregiver Recommendation mechanism, called CARES, designed to recommend suitable caregivers for postpartum women using deep learning and knowledge graphs. Initially, the proposed CARES utilizes Extreme Gradient Boosting (XGBoost) to identify important features, addressing the issue of feature similarity. Then it employs K-Means clustering to group postpartum women and caregivers based on similar features. Subsequently, it utilizes a Deep & Cross Network (DCN) to automatically learn feature interactions and constructs knowledge graphs to tackle the cold start problem. The proposed CARES also integrates exploration and exploitation strategies to balance the accuracy and diversity of recommendations. The proposed CARES compares with existing mechanisms on real datasets, and the simulation results demonstrate its effectiveness in terms of precision, recall, and F1-Score.

A Multiple Mobile Chargers Collaboratative Recharging Scheme for Enhanced Data Quailty and Sustainable Lifetime in WRSNs

Tue, 10 Mar 2026 04:08:44 GMT

title: A Multiple Mobile Chargers Collaboratative Recharging Scheme for Enhanced Data Quailty and Sustainable Lifetime in WRSNs abstract: With the rapid advancement of Internet of Things (IoT) technology, wireless sensor networks (WSNs) have become increasingly pivotal across various domains. However, the limited battery life of sensor nodes remains a critical bottleneck affecting their development and sustainability. In wireless rechargeable sensor networks (WRSNs), existing multiple mobile chargers (MCs) approach face challenges, including uneven node distribution, suboptimal charging station placement, and poor coordination, resulting in diminished recharging efficiency and degraded surveillance quality. Therefore, this study proposes a cooperative recharging mechanism for WRSNs based on multiple MCs. This mechanism first uses attractive-repulsive forces and Voronoi diagrams to achieve sensor relocation, resulting in a more balanced distribution of sensor nodes and improved spatial surveillance quality. Second, by optimizing the number and location of recharging stations (RSs) and dividing the surveillance area into multiple subregions, each served by a single MC, recharging efficiency is enhanced while satisfying the energy capacity constraints of MCs. Finally, the batch recharging strategy is adopted, and the sensing frequency of sensors are adjusted based on the recharging interval to improve temporal surveillance quality. Simulation results demonstrate that the proposed mechanism outperforms existing algorithms in terms of energy consumption, recharging efficiency, and surveillance quality, effectively extending the lifetime of WRSNs and improving surveillance quality, providing a new perspective for the sustainable development and application of WRSNs.

Teaching Authentic Sign Language Through Multiple Representation Learning

Mon, 09 Mar 2026 04:06:04 GMT

title: Teaching Authentic Sign Language Through Multiple Representation Learning abstract: Sign language recognition plays a crucial role in bridging the communication gap between hearing-impaired and hearing individuals. However, traditional teaching systems often rely on expert systems or rule-based approaches for recognition, which struggle to meet the needs of learners at different levels. This paper introduces a novel Sign Language Teaching and Scoring System (SLTS) based on multi-model collaboration, aimed at improving learning efficiency and accuracy for diverse learners. The proposed SLTS employs teaching strategies suitable for both beginners and advanced learners, offering a comprehensive solution for sign language education through multiple representation learning. Specifically, for beginners, it uses an improved Siamese Long Short-Term Memory (LSTM) module to facilitate passive learning. This approach analyzes individual gestures by comparing them to conventional sign language, allowing novices to focus on mimicking movements and establishing a solid foundation in sign language norms. For advanced learners, the proposed SLTS implements an active learning approach using an enhanced Convolutional LSTM (ConvLSTM) module to handle more complex sign language vocabulary. The system captures both spatial and temporal features of gestures, enhancing learners' fluency and expressiveness in real communication scenarios. The experimental results in real-world environments demonstrate that the proposed SLTS significantly outperforms existing methods in recognition accuracy, proving its effectiveness and advanced nature.

Toward Interpretable Multimodal Violence Detection with Knowledge Distillation and Modality-Aligned Preprocessing

Mon, 09 Mar 2026 04:05:59 GMT

title: Toward Interpretable Multimodal Violence Detection with Knowledge Distillation and Modality-Aligned Preprocessing abstract: Social violence presents a compelling challenge to public safety, yet existing multimodal detection systems exhibit excessive reliance on RGB image semantics and opaque decision-making processes. Despite leveraging visual and auditory data, current models demonstrate RGB bias in feature prioritization, as evidenced by explainability analyzes, thereby limiting their generalization for behavioral understanding. Additionally, modality inconsistency and inefficient fusion mechanisms impair model transparency and training stability. To bridge these gaps, this study proposes modality-aligned preprocessing (VAJ) that structurally unifies visual-auditory features through conflict resolution and input optimization, explicitly suppressing color dominance while enhancing interpretable feature representations. Complementing this, we design DTVDS, an interpretable detection framework integrating knowledge distillation to transfer distilled behavioral insights from a cumbersome teacher network to an efficient student model. This dual strategy not only addresses computational overhead but also clarifies decision logic through simplified inference pathways. Evaluations on XD-Violence and UCF-Crime benchmarks demonstrate superior performance, with AP (89.64%) and AUC (88.35%) outperforming existing methods. Qualitative evaluations further validate interpretability, revealing modality-coherent attention maps and human-aligned rationale visualization. The proposed method advances violence detection by addressing persistent shortcomings in multimodal alignment and model explainability.

Dynamic Weather-Adaptive Enhanced Barrier Coverage with Adjustable-Range Sensors for WRSNs

Mon, 09 Mar 2026 04:05:56 GMT

title: Dynamic Weather-Adaptive Enhanced Barrier Coverage with Adjustable-Range Sensors for WRSNs abstract: Barrier coverage is vital for wireless sensor networks (WSNs). Traditional approaches using battery-powered, fixed-radius sensors under the Boolean Sensing Model (BSM) struggle to ensure long-term, high-quality monitoring. This paper proposes BCRAS, a barrier coverage algorithm based on solar-powered sensors with adjustable sensing radii and the Probabilistic Sensing Model (PSM). It addresses three key challenges: (1) To cope with solar power uncertainty, a CNN-LSTM model predicts next-day PV energy to support energy-aware scheduling; (2) To manage varying energy consumption across sensing ranges, each sensor selects its sensing radius based on predicted energy gain and usage balance; (3) To enhance coverage under PSM, sensors are scheduled according to their cooperative detection probability at bottleneck points. Experiments show that BCRAS improves surveillance quality, energy utilization, and long-term stability compared to existing methods.

MMDL: A Multi-Modal Deep Learning for Video Highlight Detection in Sports

Mon, 09 Mar 2026 04:05:51 GMT

title: MMDL: A Multi-Modal Deep Learning for Video Highlight Detection in Sports abstract: With the growing interest in sports events, the ability to capture highlights has become increasingly important. Traditionally, the process of editing these highlights required significant time and manpower. To address this challenge, this paper introduces an innovative multi-modal deep learning method for highlight detection (MMDL). The proposed MMDL integrates information from multiple modalities, including subtitles, static skeletal features, and video content, to gain a deep understanding of specific behaviors and identify sub-videos containing those highlights. Additionally, the proposed MMDL employed Siamese networks to accurately capture different aspects of behavior by comparing the similarity between input and training videos across different modalities. Experiments conducted on two datasets, MLB-YouTube and ELTA, demonstrate that the proposed MMDL significantly outperforms existing models, achieving at least a 5% improvement in F1-Score compared to the baseline models, such as I3D and NPL.

RealExp: Decoupling Correlation Bias in Shapley Values for Faithful Model Interpretations

Mon, 09 Mar 2026 04:05:46 GMT

title: RealExp: Decoupling Correlation Bias in Shapley Values for Faithful Model Interpretations abstract: Deep learning has achieved significant success in handling unstructured data but remains limited by its "black box" nature, especially in sensitive applications. Existing interpretable machine learning methods partially address this issue but often overlook feature correlations and provide inadequate assessments of model decision paths. To tackle these challenges, this paper introduces Real Explainer (RealExp), a novel interpretability method that decouples the Shapley Value into individual feature importance and feature correlation importance. By integrating feature similarity computations, RealExp enhances interpretability by precisely quantifying both individual contributions and interactions, leading to more reliable explanations. Furthermore, a new interpretability evaluation criterion is proposed, focusing on decision path analysis beyond accuracy-based assessments. Experiments on image classification and sentiment analysis demonstrate RealExp's superiority in interpretability. Case studies highlight its practical benefits: RealExp aids in selecting pre-trained models for pneumonia detection, emphasizes pre-trained video segmentation for anomaly detection, and optimizes text models, achieving performance to RoBERTA finetuning model without pre-trained embeddings.

Hierarchical Knowledge Graph-Based QA System with Retrieval-Augmented Generation

Fri, 06 Mar 2026 04:06:43 GMT

title: Hierarchical Knowledge Graph-Based QA System with Retrieval-Augmented Generation abstract: Hierarchical knowledge graphs (KGs) are vital to question-answering (QA) systems for complex queries, integrating structured and unstructured knowledge. This study introduces a QA system combining a hierarchical KG, graph convolutional networks (GCNs), and retrieval-augmented generation (RAG) to enhance reasoning, retrieval, and response generation. The KG organises information into title, subtitle, and content layers for structured, efficient retrieval; GCNs aggregate local and global relations across layers; RAG incorporates external sources (e.g., Wikipedia) for contextually accurate answers. On standard benchmarks, the system outperformed strong baselines in precision, recall, and F1-score, offering an effective solution for complex queries and advancing QA design.

A Machine Learning-Based Model for Classifying the Shape of Tomato

Fri, 06 Mar 2026 04:06:39 GMT

title: A Machine Learning-Based Model for Classifying the Shape of Tomato abstract: Most fruit classification studies rely on color-based features, but shape-based analysis provides a promising alternative for distinguishing subtle variations within the same variety. Tomato shape classification is challenging due to irregular contours, variable imaging conditions, and difficulty in extracting consistent geometric features. In this study, we propose an efficient and structured workflow to address these challenges through contour-based analysis. The process begins with the application of a Mask Region-based Convolutional Neural Network (Mask R-CNN) model to accurately isolate tomatoes from the background. Subsequently, the segmented tomatoes are extracted and encoded using Elliptic Fourier Descriptors (EFDs) to capture detailed shape characteristics. These features are used to train a range of machine learning models, including Support Vector Machine (SVM), Random Forest, One-Dimensional Convolutional Neural Network (1D-CNN), and Bidirectional Encoder Representations from Transformers (BERT). Experimental results observe that the Random Forest model achieved the highest accuracy of 79.4%. This approach offers a robust, interpretable, and quantitative framework for tomato shape classification, reducing manual labor and supporting practical agricultural applications.

A Dynamic Recommendation System Integrated Long Short-Term Memory (LSTM) and Matrix Factorization

Thu, 05 Mar 2026 04:06:47 GMT

title: A Dynamic Recommendation System Integrated Long Short-Term Memory (LSTM) and Matrix Factorization abstract: Matrix factorization (MF) technique has been widely utilized in recommendation systems due to the precise prediction of users’ interests. Prior MF-based methods adapt the overall rating to make the recommendation by extracting latent factors from users and items. However, in real applications, people’s preferences usually vary with time; the traditional MF-based methods could not properly capture the change of users’ interests. In this paper, by incorporating the recurrent neural network (RNN) into MF, we developed a novel recommendation system, M-RNN-F, to effectively describe the preference evolution of users over time. A learning model is proposed to capture the evolution pattern and predict the user preference in the future. The experimental results show that M-RNN-F performs better than other state-of-the-art recommendation algorithms. In addition, we conduct experiments on real world dataset to demonstrate the practicability.

Replacing Batch Normalization with Memory-Based Affine Transformation for Test-Time Adaptation

Thu, 27 Nov 2025 04:05:28 GMT

A Batch-Statistics-Free Adaptive Normalization Method for Robust Few-Shot Learning and Domain Adaptation

Thu, 27 Nov 2025 04:05:18 GMT

title: A Batch-Statistics-Free Adaptive Normalization Method for Robust Few-Shot Learning and Domain Adaptation abstract: Batch Normalization (BN) has been widely adopted in deep neural networks for its ability to stabilize training and improve convergence. However, BN relies on batch-wise mean and variance estimates, which can become inaccurate during inference, particularly in Few-shot Learning (FSL) and domain adaptation scenarios where the test distribution differs from training or the available batch size is small. This dependency often causes performance degradation due to mismatched or outdated statistics. In this work, we introduce Meta Affine Transformation (MetaAFN), a batch-statistics-free normalization strategy that replaces the normalization step in BN with a meta-network-generated affine transformation. By entirely removing the reliance on batch statistics, MetaAFN avoids mismatched training-set statistics and instead uses a lightweight meta-network to dynamically produce scale ( γ ) and shift ( β ) parameters conditioned on the current input features. This design enables the model to adaptively modulate representations without explicit BN, improving robustness to distribution shifts. We evaluate MetaAFN on two representative tasks — FSL and source-free domain adaptation — using multiple benchmark datasets. Experimental results show that MetaAFN consistently outperforms or matches BN and MetaBN, with clear advantages under significant distributional shifts. These findings highlight MetaAFN as an effective and practical alternative to BN, offering improved adaptability and generalization in heterogeneous data scenarios.

Content-adaptive reversible data hiding with multi-stage prediction schemes

Thu, 23 Oct 2025 04:05:33 GMT

title: Content-adaptive reversible data hiding with multi-stage prediction schemes abstract: With the proliferation of image-capturing and display-enabled IoT devices, ensuring the authenticity and integrity of visual data has become increasingly critical, especially in light of emerging cybersecurity threats and powerful generative AI tools. One of the major challenges in such sensor-based systems is the ability to protect privacy while maintaining data usability. Reversible data hiding has attracted growing attention due to its reversibility and ease of implementation, making it a viable solution for secure image communication in IoT environments. In this paper, we propose reversible data hiding techniques tailored to the content characteristics of images. Our approach leverages subsampling and quadtree partitioning, combined with multi-stage prediction schemes, to generate a predicted image aligned with the original. Secret information is embedded by analyzing the difference histogram between the original and predicted images, and enhanced through multi-round rotation techniques and a multi-level embedding strategy to boost capacity. By employing both subsampling and quadtree decomposition, the embedding strategy dynamically adapts to the inherent characteristics of the input image. Furthermore, we investigate the trade-off between embedding capacity and marked image quality. Experimental results demonstrate improved embedding performance, high visual fidelity, and low implementation complexity, highlighting the method’s suitability for resource-constrained IoT applications.

Multiple parents crossover operators: A new approach removes the overlapping solutions for sequencing problems

Wed, 01 Oct 2025 04:05:47 GMT

title: Multiple parents crossover operators: A new approach removes the overlapping solutions for sequencing problems abstract: This paper presents a novel multiple parents crossover operator approach for solving sequencing problems, particularly addressing the issue of overlapping solutions in genetic algorithms. The proposed method enhances solution diversity and convergence performance in scheduling and optimization problems.

Artificial Immune Network with Feature Selection for Bank Term Deposit Recommendation

Wed, 01 Oct 2025 04:05:41 GMT

title: Artificial Immune Network with Feature Selection for Bank Term Deposit Recommendation abstract: This paper proposes an artificial immune network approach combined with feature selection techniques for bank term deposit recommendation. The method utilizes immune system principles to identify optimal feature subsets and improve prediction accuracy in banking marketing campaigns. The proposed algorithm demonstrates superior performance compared to traditional machine learning methods in terms of both accuracy and computational efficiency. Experimental results on real banking datasets show significant improvements in customer targeting and conversion rates.

Verifiable (2, n) Image Secret Sharing Scheme Using Sudoku Matrix

Wed, 01 Oct 2025 04:05:37 GMT

title: Verifiable (2, n) Image Secret Sharing Scheme Using Sudoku Matrix abstract: This paper presents a novel verifiable (2, n) image secret sharing scheme using Sudoku matrix properties. The proposed method enhances traditional secret sharing by incorporating verification capabilities through Sudoku matrix structures. The scheme ensures both security and authenticity of shared image secrets while maintaining computational efficiency. Experimental results demonstrate the effectiveness of our approach in various image types and security scenarios.

A Random Forest-Enhanced Genetic Algorithm for Order Acceptance Scheduling with Past-Sequence-Dependent Setup Times

Wed, 01 Oct 2025 04:05:32 GMT

title: A Random Forest-Enhanced Genetic Algorithm for Order Acceptance Scheduling with Past-Sequence-Dependent Setup Times abstract: This paper proposes a novel hybrid approach combining Random Forest machine learning techniques with Genetic Algorithms to solve the order acceptance and scheduling problem with past-sequence-dependent setup times. The Random Forest component helps predict optimal scheduling patterns, while the Genetic Algorithm optimizes the overall solution. Experimental results demonstrate superior performance compared to traditional methods in terms of computational efficiency and solution quality.

Revised NMS-Driven Pipeline for Heart Valve Regurgitation and Kawasaki Disease Coronary Aneurysm Localization

Wed, 01 Oct 2025 04:05:25 GMT

title: Revised NMS-Driven Pipeline for Heart Valve Regurgitation and Kawasaki Disease Coronary Aneurysm Localization

A 3D Spatial–Spectral–Temporal Deep Regression Model for Improving Mangrove Canopy Height Estimation Through Fusion of Optimized Red-Edge Sentinel-2 Bands and Sentinel-1 SAR Data

Mon, 22 Sep 2025 04:06:58 GMT

title: A 3D Spatial–Spectral–Temporal Deep Regression Model for Improving Mangrove Canopy Height Estimation Through Fusion of Optimized Red-Edge Sentinel-2 Bands and Sentinel-1 SAR Data abstract: Mangroves are vital blue carbon ecosystems with high carbon storage, where canopy height is a key parameter for estimating above-ground biomass. This study integrates Sentinel-1 SAR time-series and Sentinel-2 optical imagery and focused on the investigation of Red-Edge (RE) bands for mangrove canopy height estimation. A new RE-based spectral index named REMCH (RE Mangrove Canopy Height) index was developed for improving mangrove canopy height estimation. To improve the estimation results, this study proposed the 3DSST-RECLT model, a 3D spatial–spectral–temporal deep learning regression model that combining ConvLSTM, hybrid 3D–2D convolution, and Swin Transformer. Airborne LiDAR canopy height data served as target data. Results show fusing Sentinel-1 time-series and Sentinel-2 data using the proposed 3DSST-RECLT model achieved satisfactory performance, with the inclusion of RE bands and the REMCH index enhancing the model performance with an average mean absolute error of 1.648 m on the test dataset and outperforming the other models. This study produced mangrove canopy height maps of the coastal zone of South and Southwest Florida for 2017 and 2020 and found an increase in mangrove canopy height between 2017 and 2020. The produced mangrove canopy height map for 2020 was compared with three global canopy height maps, with the map generated in this study exhibiting higher accuracy. This finding indicates the advantage of integrating Sentinel-1 time-series and Sentinel-2 RE bands with a deep learning regression model to improve mangrove canopy height mapping and monitoring.

Sanitizing diffusion-generated images via fingerprint removal and adversarial perturbation for forensic evasion

Thu, 18 Sep 2025 04:06:30 GMT

title: Sanitizing diffusion-generated images via fingerprint removal and adversarial perturbation for forensic evasion abstract: Diffusion models have rapidly advanced the realism of synthetic image generation, posing new challenges for forensic detectors. This paper proposes a two-stage forensic evasion framework designed to undermine the detectability of diffusion-generated images. In the first stage, a spectrum-aware generative adversarial network (GAN) removes frequency-domain fingerprints that are commonly exploited by forensic models. In the second stage, adversarial perturbations are applied using the Iterative Fast Gradient Sign Method (I-FGSM) to further mislead detectors while preserving visual fidelity. Experiments conducted on COCO-based datasets demonstrate that our method significantly reduces detection accuracy across multiple state-of-the-art forensic models, including UniFD, DIGBD, and SSIP. Furthermore, we show that combining fingerprint removal with adversarial perturbation achieves stronger evasion than either method alone. Ablation studies also highlight the benefits of adaptive perturbation strengths and data augmentation for enhancing cross-model evasion. This work reveals critical vulnerabilities in current forensic approaches and underscores the need for more robust detection systems against adaptive evasion

Convergence of Technological Social Changes in the Development of Intelligent Technology Innovation System in Taiwan

Wed, 17 Sep 2025 04:05:28 GMT

title: Convergence of Technological Social Changes in the Development of Intelligent Technology Innovation System in Taiwan abstract: How do intelligent technology innovation systems develop? How do institutional factors influence the development of innovation systems? This study adopted the National/Sectoral Innovation System as the main framework to study the dynamics of intelligent technology innovation system in Taiwan. Combining scientometric mapping, social network analysis, Multi Criteria Decision-Making (MCDM), and 33 interviews to understand the interactions between key stakeholders and the institutions in the intelligent technology sector in Taiwan. The results show that a relatively large proportions of foreign technologies are widely adopted into the system through international business networks. Regarding commercial applications, firms prefer to introduce foreign technology through international business networks. This illustrates the needs of commercializations of domestic technological research and development. The future policy should promote academia-industry collaborations to enhance the commercializations of scientific research outcomes. This research contributes to enhance the knowledge flow in the innovation system, thus, to upgrade the national capabilities of small emerging economies in East Asia.

SMRT: Surveillance Monitoring and Recognition Techniques for Analyzing Service Behavior in Blurred and Unsteady Video

Mon, 15 Sep 2025 04:06:13 GMT

title: SMRT: Surveillance Monitoring and Recognition Techniques for Analyzing Service Behavior in Blurred and Unsteady Video abstract: With the rapid development of the service industry and increasing customer expectations, traditional mystery shopper audit methods face several challenges, such as time-consuming manual analysis, significant subjective bias, and difficulty in accurately quantifying complex service behaviors. To overcome these limitations, this paper introduces an innovative approach called Surveillance Monitoring and Recognition Techniques (SMRT) for analyzing service behavior. The proposed SMRT achieves precise classification of service behaviors through a two-phase process: coarse-grained and fine-grained analysis. In the coarse-grained phase, the proposed SMRT preprocesses blurred video to extract and emphasize relevant external features, specifically detecting and capturing ‘person’ objects in video frames, thereby effectively filtering out irrelevant frames and reducing computational load. In the fine-grained phase, it performs spatiotemporal feature extraction and utilizes Transformer models to conduct a detailed comparison of target behavioral features across video segments. Simulation results demonstrate that the proposed SMRT significantly enhances recognition performance in terms of accuracy, and F1-score compared to existing methods.

AI-Based Multimodal Anomaly Detection for Industrial Machine Operations

Mon, 15 Sep 2025 04:06:07 GMT

title: AI-Based Multimodal Anomaly Detection for Industrial Machine Operations abstract: In the manufacturing process involving grinding wheels, challenges in fine-tuning grinding machines are typically addressed by craftsmen through subjective observations of sparks and sounds. However, most current anomaly detection methods mainly aim at a single modality, whereas existing multimodal methods cannot effectively cope with a common issue. To address this, this paper introduces an innovative mechanism, AI-Based Multimodal Anomaly Detection (AMAD), designed to optimize the efficiency and accuracy of grinding wheel production lines. The proposed AMAD includes data preprocessing and multimodal anomaly detection, accurately identifying anomalies in grinding wheel operation videos. In the data preprocessing phase, the proposed AMAD utilizes Mel Frequency Cepstral Coefficients (MFCC) and AutoEncoder for audio processing and segmentation for video processing. In the multimodal anomaly detection phase, the proposed AMAD employs Convolutional Neural Networks (CNN) for audio analysis and Convolutional Long Short-Term Memory (ConvLSTM) for video analysis. By combining both audio and video modalities, the proposed AMAD effectively predicts whether the input video represents normal or abnormal grinding wheel operations. This multimodal approach not only improves the accuracy of anomaly detection but also enhances the robustness of the system. Simulation results demonstrate that the proposed AMAD significantly improves performance in anomaly detection in terms of precision, recall, and F1-Score.

Meta Network for Flow-Based Image Style Transfer

Fri, 12 Sep 2025 04:05:34 GMT

title: Meta Network for Flow-Based Image Style Transfer abstract: A style transfer aims to produce synthesized images that retain the content of one image while adopting the artistic style of another. Traditional style transfer methods often require training separate transformation networks for each new style, limiting their adaptability and scalability. To address this challenge, we propose a flow-based image style transfer framework that integrates Randomized Hierarchy Flow (RH Flow) and a meta network for adaptive parameter generation. The meta network dynamically produces the RH Flow parameters conditioned on the style image, enabling efficient and flexible style adaptation without retraining for new styles. RH Flow enhances feature interaction by introducing a random permutation of the feature sub-blocks before hierarchical coupling, promoting diverse and expressive stylization while preserving the content structure. Our experimental results demonstrate that Meta FIST achieves superior content retention, style fidelity, and adaptability compared to existing approaches.

Comparing the Carpal Tunnel Area and Carpal Boundaries in Patients with Carpal Tunnel Syndrome and Healthy Volunteers: A Magnetic Resonance Imaging Study

Fri, 12 Sep 2025 04:05:26 GMT

title: Comparing the Carpal Tunnel Area and Carpal Boundaries in Patients with Carpal Tunnel Syndrome and Healthy Volunteers: A Magnetic Resonance Imaging Study abstract: Background: Carpal tunnel syndrome (CTS) is a common neuropathy caused by compression of the median nerve (MN) within the carpal tunnel, which causes pain, paresthesia, or altered sensation. While a small carpal tunnel area is considered a risk factor for CTS, varying carpal tunnel dimensions in CTS patients have been obtained via axial computed tomography and magnetic resonance imaging (MRI). Methods: In this retrospective study, MR images from 49 CTS patients and 38 healthy controls were analyzed to investigate differences in the carpal tunnel area and carpal boundaries between the groups and to explore the relationships of these parameters with CTS severity. Results: Our findings revealed that compared with the controls, CTS patients presented significantly larger cross-sectional areas (CSAs) of the MN and carpal tunnel and increased MN flattening ratios. The CSAs of the MN showed moderate positive correlations with severity (r = 0.395, p < 0.001), symptom score (r = 0.354, p < 0.001), and disability score (r = 0.300, p < 0.001), while the carpal tunnel area showed weaker but significant correlations with severity (r = 0.268, p = 0.002), symptom score (r = 0.173, p = 0.026), and disability score (r = 0.183, p = 0.018). The ratios of the MN CSA to those of the carpal tunnel, the interior carpal boundary (ICB), the exterior carpal boundary (ECB), and the wrist were disproportionately greater in the CTS patients. Among them, both the MN-to-ICB and MN-to-ECB ratios had fair to good diagnostic values (area under the curve = 0.725 and 0.794, respectively). Conclusions: These results highlight the utility of MRI-derived CSA measurements and ratios in identifying pathophysiological changes in CTS patients, particularly crowding of the MN inside the carpal tunnel. Further studies are recommended to refine MRI-based diagnostic protocols for CTS.

Development and Deployment of a Virtual Water Gauge System Utilizing the ResNet-50 Convolutional Neural Network for Real-Time River Water Level Monitoring: A Case Study of the Keelung River in Taiwan

Mon, 01 Sep 2025 04:05:13 GMT

title: Development and Deployment of a Virtual Water Gauge System Utilizing the ResNet-50 Convolutional Neural Network for Real-Time River Water Level Monitoring: A Case Study of the Keelung River in Taiwan abstract: Climate change has exacerbated severe rainfall events, leading to rapid and unpredictable fluctuations in river water levels. This environment necessitates the development of real-time, automated systems for water level detection. Due to degradation, traditional methods relying on physical river gauges are becoming progressively unreliable. This paper presents an innovative methodology that leverages ResNet-50, a Convolutional Neural Network (CNN) model, to identify distinct water level features in Closed-Circuit Television (CCTV) river imagery of the Chengmei Bridge on the Keelung River in Neihu District, Taiwan, under various weather conditions. This methodology creates a virtual water gauge system for the precise and timely detection of water levels, thereby eliminating the need for dependable physical gauges. Our study utilized image data from 1 March 2022 to 28 February 2023. This river, crucial to the ecosystems and economies of numerous cities, could instigate a range of consequences due to rapid increases in water levels. The proposed system integrates grid-based methods with infrastructure like CCTV cameras and Raspberry Pi devices for data processing. This integration facilitates real-time water level monitoring, even without physical gauges, thus reducing deployment costs. Preliminary results indicate an accuracy range of 83.6% to 96%, with clear days providing the highest accuracy and heavy rainfall the lowest. Future work will refine the model to boost accuracy during rainy conditions. This research introduces a promising real-time river water level monitoring solution, significantly contributing to flood control and disaster management strategies.