Search Results (39)

Real-time visual inference on resource-constrained hardware remains a core challenge for edge computing and embedded artificial intelligence systems. Recent deep learning architectures, particularly Vision Transformers (ViTs) and Detection Transformers (DETRs), achieve high detection accuracy but impose substantial computational and memory demands that limit their deployment on low-power edge platforms such as NVIDIA Jetson and Raspberry Pi devices. This paper presents a systematic review of model compression and optimization strategies—specifically quantization, pruning, and knowledge distillation—applied to lightweight object detection architectures for edge deployment. Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, peer-reviewed studies were analyzed from Scopus, IEEE Xplore, and ScienceDirect to examine the evolution of efficient detectors from convolutional neural networks to transformer-based models. The synthesis highlights a growing focus on real-time transformer variants, including Real-Time DETR (RT-DETR) and low-bit quantized approaches such as Q-DETR, alongside optimized YOLO-based architectures. While quantization enables substantial theoretical acceleration (e.g., up to 16× operation reduction), aggressive low-bit precision introduces accuracy degradation, particularly in transformer attention mechanisms, highlighting a critical efficiency-accuracy tradeoff. The review further shows that Quantization-Aware Training (QAT) consistently outperforms Post-Training Quantization (PTQ) in preserving performance under low-precision constraints. Finally, this review identifies critical open research challenges, emphasizing the efficiency–accuracy tradeoff and the high computational demands imposed by Transformer architectures. Future directions are proposed, including hardware-aware optimization, robustness to imbalanced datasets, and multimodal sensing integration, to ensure reliable real-time inference in practical agricultural edge computing environments. Full article

This paper reviews Artificial Intelligence techniques for distributed energy management, focusing on integrating machine learning, reinforcement learning, and multi-agent systems within IoT-Edge-Cloud architectures. As energy infrastructures become increasingly decentralized and heterogeneous, AI must operate under strict latency, privacy, and resource constraints while remaining transparent and auditable. The study examines predictive models ranging from statistical time series approaches to machine learning regressors and deep neural architectures, assessing their suitability for embedded deployment and federated learning. Optimization methods—including heuristic strategies, metaheuristics, model predictive control, and reinforcement learning—are analyzed in terms of computational feasibility and real-time responsiveness. Explainability is treated as a fundamental requirement, supported by model-agnostic techniques that enable trust, regulatory compliance, and interpretable coordination in multi-agent environments. The review synthesizes advances in MARL for decentralized control, communication protocols enabling interoperability, and hardware-aware design for low-power edge devices. Benchmarking guidelines and key performance indicators are introduced to evaluate accuracy, latency, robustness, and transparency across distributed deployments. Key challenges remain in stabilizing explanations for RL policies, balancing model complexity with latency budgets, and ensuring scalable, privacy-preserving learning under non-stationary conditions. The paper concludes by outlining a conceptual framework for explainable, distributed energy intelligence and identifying research opportunities to build resilient, transparent smart energy ecosystems. Full article

High-Performance Computing (HPC) has become essential for training medium- and large-scale Artificial Intelligence (AI) models, yet two bottlenecks remain under-exploited: the semantic coherence of training data and the interaction between distributed deep learning runtimes and heterogeneous HPC architectures. Existing work tends to optimise multi-node, multi-GPU training in isolation from data semantics or to apply semantic technologies to data curation without considering the constraints of large-scale training on modern clusters. This paper introduces SemanticHPC, an experimental framework that integrates ontology and Resource Description Framework (RDF)-based semantic preprocessing with distributed AI training (Horovod/PyTorch Distributed Data Parallel) and hardware-aware optimisations for Non-Uniform Memory Access (NUMA), multi-GPU and high-speed interconnects. The framework has been evaluated on 1–8 node configurations (4–32 GPUs) on a production-grade cluster. Experiments on a medium-size Open Images V7 workload show that semantic enrichment improves validation accuracy by 3.5–4.4 absolute percentage points while keeping the additional end-to-end overhead below 8% and preserving strong scaling efficiency above 79% on eight nodes. We argue that bringing semantic technologies into the training workflow—rather than treating them as an offline, detached phase—is a promising direction for large-scale AI on HPC systems. We detail an implementation based on standard Python libraries, RDF tooling and widely adopted deep learning runtimes, and we discuss the limitations and practical hurdles that need to be addressed for broader adoption. Full article

Network function virtualization (NFV) is an emerging technology that is gaining popularity for network function migration. NFV converts a network function from a dedicated hardware device into a virtual network function (VNF), thereby improving the agility of network services and reducing management costs. A complex network service can be expressed as a service function chain (SFC) request, which consists of an ordered sequence of VNFs. Given the inherent heterogeneity and dynamic nature of network services, effective SFC deployment encounters significant unpredictable challenges. Machine learning-based methods offer the flexibility to predict and select the optimal next action based on existing data models. In this paper, we propose a deep reinforcement learning-based approach for bandwidth-aware service function chaining (DRL-BSFC). Aiming to simultaneously improve the acceptance ratio of SFC requests and maximize the total revenue for Internet service providers, DRL-BSFC integrates a graph convolutional network (GCN) for feature extraction of the underlying physical network, a sequence-to-sequence (Seq2Seq) model for capturing the order information of an SFC request, and a modified A3C (Asynchronous Advantage Actor–Critic) algorithm of deep reinforcement learning. To ensure efficient resource utilization and a higher acceptance ratio of SFC requests, the bandwidth cost for deploying an SFC is explicitly incorporated into the A3C’s reward function. The effectiveness and superiority of DRL-BSFC compared to the existing DRL-SFCP scheme are demonstrated via simulations. The performance measures include the acceptance ratio of SFC requests, the average bandwidth cost, the average remaining link bandwidth, and the average revenue-to-cost ratio under different SFC request arrival rates. Full article

We propose a Hybrid Parallel Temporal–Spatial CNN-LSTM (HPTS-CL) architecture for optimized indoor environment modeling in sports halls, addressing the computational and scalability challenges of high-resolution spatiotemporal data processing. The sports hall is partitioned into distinct zones, each processed by dedicated CNN branches to extract localized spatial features, while hierarchical LSTMs capture both short-term zone-specific dynamics and long-term inter-zone dependencies. The system integrates model and data parallelism to distribute workloads across specialized hardware, dynamically balanced to minimize computational bottlenecks. A gated fusion mechanism combines spatial and temporal features adaptively, enabling robust predictions of environmental parameters such as temperature and humidity. The proposed method replaces monolithic CNN-LSTM pipelines with a distributed framework, significantly improving efficiency without sacrificing accuracy. Furthermore, the architecture interfaces seamlessly with existing sensor networks and control systems, prioritizing critical zones through a latency-aware scheduler. Implemented on NVIDIA Jetson AGX Orin edge devices and Google Cloud TPU v4 pods, HPTS-CL demonstrates superior performance in real-time scenarios, leveraging lightweight EfficientNetV2-S for CNNs and IndRNN cells for LSTMs to mitigate gradient vanishing. Experimental results validate the system’s ability to handle large-scale, high-frequency sensor data while maintaining low inference latency, making it a practical solution for intelligent indoor environment optimization. The novelty lies in the hybrid parallelism strategy and hierarchical temporal modeling, which collectively advance the state of the art in distributed spatiotemporal deep learning. Full article

State-of-charge (SoC) balancing across multiple battery energy storage systems (BESS) is a central challenge in renewable-rich mini-grids. Heterogeneous battery capacities, differing states of health, stochastic renewable generation, and variable loads create a high-dimensional uncertain control problem. Conventional droop-based SoC balancing strategies are decentralized and computationally light but fundamentally reactive and limited, whereas model predictive control (MPC) is insightful but computationally intensive and prone to modeling errors. This paper proposes a Hierarchical Predictive–Adaptive Control (HPAC) framework for SoC balancing in mini-grids using deep reinforcement learning. The framework consists of two synergistic layers operating on different time scales. A long-horizon Predictive Engine, implemented as a federated Transformer network, provides multi-horizon probabilistic forecasts of net load, enabling multiple mini-grids to collaboratively train a high-capacity model without sharing raw data. A fast-timescale Adaptive Controller, implemented as a Soft Actor-Critic (SAC) agent, uses these forecasts to make real-time charge/discharge decisions for each BESS unit. The forecasts are used both to augment the agent’s state representation and to dynamically shape a multi-objective reward function that balances SoC, economic performance, degradation-aware operation, and voltage stability. The paper formulates SoC balancing as a Markov decision process, details the SAC-based control architecture, and presents a comprehensive evaluation using a MATLAB-(R2025a)-based digital-twin simulation environment. A rigorous benchmarking study compares HPAC against fourteen representative controllers spanning rule-based, MPC, and various DRL paradigms. Sensitivity analysis on reward weight selection and ablation studies isolating the contributions of forecasting and dynamic reward shaping are conducted. Stress-test scenarios, including high-volatility net-load conditions and communication impairments, demonstrate the robustness of the approach. Results show that HPAC achieves near-minimal operating cost with essentially zero SoC variance and the lowest voltage variance among all compared controllers, while maintaining moderate energy throughput that implicitly preserves battery lifetime. Finally, the paper discusses a pathway from simulation to hardware-in-the-loop testing and a cloud-edge deployment architecture for practical, real-time deployment in real-world mini-grids. Full article

Biometric authentication on smartphones has advanced rapidly in recent years, with face recognition becoming the dominant modality due to its convenience and easy integration with modern mobile hardware. However, despite these developments, smartphone-based facial recognition systems remain vulnerable to a broad spectrum of attacks. This survey provides an updated and comprehensive examination of the evolving attack landscape and corresponding defense mechanisms, incorporating recent advances up to 2025. A key contribution of this work is a structured taxonomy of attack types targeting smartphone facial recognition systems, encompassing (i) 2D and 3D presentation attacks; (ii) digital attacks; and (iii) dynamic attack patterns that exploit acquisition conditions. We analyze how these increasingly realistic and condition-dependent attacks challenge the robustness and generalization capabilities of modern face anti-spoofing (FAS) systems. On the defense side, the paper reviews recent progress in liveness detection, deep-learning- and transformer-based approaches, quality-aware and domain-generalizable models, and emerging unified frameworks capable of handling both physical and digital spoofing. Hardware-assisted methods and multi-modal techniques are also examined, with specific attention to their applicability in mobile environments. Furthermore, we provide a systematic overview of commonly used datasets, evaluation metrics, and cross-domain testing protocols, identifying limitations related to demographic bias, dataset variability, and controlled laboratory conditions. Finally, the survey outlines key research challenges and future directions, including the need for mobile-efficient anti-spoofing models, standardized in-the-wild evaluation protocols, and defenses robust to unseen and AI-generated spoof types. Collectively, this work offers an integrated view of current trends and emerging paradigms in smartphone-based face anti-spoofing, supporting the development of more secure and resilient biometric authentication systems. Full article

Physical AI enables reliable and timely operations of autonomous systems such as robots and smart manufacturing equipment under diverse and dynamic execution environments. In these environments, computing resources are often limited, shared among tasks, and fluctuate over time. This makes it difficult to guarantee that tasks meet timing constraints. As a result, resource-aware execution time prediction becomes essential for efficient resource management in physical AI systems. However, existing methods typically assume specific environments or static resource usage and often fail to generalize to new environments. In this paper, we propose CARE-D (Calibration-Assisted Resource-aware Execution time prediction), which trains a deep neural network to model the nonlinear relationships among hardware characteristics, resource levels, and task features across environments. The model predicts the execution time of tasks under diverse hardware and dynamically allocated computing resources, using a few execution records from new environments. CARE-D applies few-history-based calibration using only 1 to k execution records from target environments to adjust predictions without retraining the model. Experiments show that CARE-D improves prediction accuracy by about 7.3% over zero-history predictors within a 10% relative error and outperforms regression and deep learning baselines, using only one to five records per target environment. Full article

Accurate sky–obstacle segmentation in hemispherical fisheye imagery is essential for solar irradiance forecasting, photovoltaic system design, and environmental monitoring. However, existing methods often rely on expensive all-sky imagers and region-specific training data, produce coarse sky–obstacle boundaries, and ignore the optical properties of fisheye lenses. We propose a low-cost segmentation framework designed for fisheye imagery that combines synthetic data generation, lens-aware augmentation, and a hybrid deep-learning pipeline. Synthetic fisheye training images are created from publicly available street-view panoramas to cover diverse environments without dedicated hardware, and lens-aware augmentations model fisheye projection and photometric effects to improve robustness across devices. On this dataset, we train a convolutional neural network (CNN) and refine its output with gradient-boosted decision trees (GBDT) to sharpen sky–obstacle boundaries. The method is evaluated on real fisheye images captured with smartphones and low-cost clip-on lenses across multiple sites, achieving an Intersection over Union (IoU) of 96.63% and an F1 score of 98.29%, along with high boundary accuracy. An additional evaluation on an external panoramic baseline dataset confirms strong cross-dataset generalization. Together, these results show that the proposed framework enables accurate, low-cost, and widely deployable hemispherical sky segmentation for practical solar and environmental imaging applications. Full article

This review surveys the past five years of research on energy-aware path optimization for both solar-powered and battery-only UAVs. First, the energy constraints of these two platforms are contrasted. Next, advanced computational frameworks—including model predictive control, deep reinforcement learning, and bio-inspired metaheuristics—are examined along with their hardware implementations. Recent studies show that hybrid methods combining neural networks with bio-inspired search can boost net energy efficiency by 15–25% while maintaining real-time feasibility on embedded GPUs or FPGAs. Among the remaining challenges are federated learning at the edge, multi-UAV coordination under partial observability, and field trials on ultra-long-endurance platforms like the Airbus Zephyr HAPS. Addressing these issues will accelerate the deployment of truly persistent and green aerial services. Full article

Sentiment analysis is essential for understanding consumer opinions, yet selecting the optimal models and embedding methods remains challenging, especially when handling ambiguous expressions, slang, or mismatched sentiment–rating pairs. This study provides a comprehensive comparative evaluation of sentiment classification models across three paradigms: traditional machine learning, pre-transformer deep learning, and transformer-based models. Using the Amazon Magazine Subscriptions 2023 dataset, we evaluate a range of embedding techniques, including static embeddings (GloVe, FastText) and contextual transformer embeddings (BERT, DistilBERT, etc.). To capture predictive confidence and model uncertainty, we include categorical cross-entropy as a key evaluation metric alongside accuracy, precision, recall, and F1-score. In addition to detailed quantitative comparisons, we conduct a systematic qualitative analysis of misclassified samples to reveal model-specific patterns of uncertainty. Our findings show that FastText consistently outperforms GloVe in both traditional and LSTM-based models, particularly in recall, due to its subword-level semantic richness. Transformer-based models demonstrate superior contextual understanding and achieve the highest accuracy (92%) and lowest cross-entropy loss (0.25) with DistilBERT, indicating well-calibrated predictions. To validate the generalisability of our results, we replicated our experiments on the Amazon Gift Card Reviews dataset, where similar trends were observed. We also adopt a resource-aware approach by reducing the dataset size from 25 K to 20 K to reflect real-world hardware constraints. This study contributes to both sentiment analysis and sustainable AI by offering a scalable, entropy-aware evaluation framework that supports informed, context-sensitive model selection for practical applications. Full article

Recent space exploration roadmaps from China, the United States, and Russia highlight the establishment of Mars bases as a major objective. Future deep-space missions will span the inner solar system and extend beyond the asteroid belt, demanding network control systems that sustain reliable communication and efficient scheduling across vast distances. Current centralized or regionalized technologies, such as the Deep-Space Network and planetary relay constellations, are limited by long delays, sparse visibility, and heterogeneous onboard resources, and thus cannot meet these demands. To address these challenges, we propose a dual-center architecture, DualSynNet, anchored at Earth and Mars and enhanced by Lagrange-point relays and a minimal heliocentric constellation to provide scalable multi-mission coverage. On this basis, we develop a federated multi-agent reinforcement learning framework with graph attention (Fed-GAT-MADDPG), integrating centralized critics, decentralized actors, and interplanetary parameter synchronization for adaptive, resource-aware scheduling. A unified metric system: Reachability, Rapidity, and Availability, is introduced to evaluate connectivity, latency, and resource sustainability. Simulation results demonstrate that our method increases task completion to 52.4%, reduces deadline expiration, constrains rover low-state-of-charge exposure to approximately 0.8%, and maintains consistently high hardware reliability across rover and satellite nodes. End-to-end latency is reduced, with a shorter tail distribution due to fewer prolonged buffering or stagnation periods. Ablation studies confirm the essential role of graph attention, as removing it reduces completion and raises expiration. These results indicate that the integration of a dual-center architecture with federated graph reinforcement learning yields a robust, scalable, and resource-efficient framework suitable for next-generation interplanetary exploration. Full article

Structure detection (SD) has emerged as a critical technology for ensuring the safety and longevity of infrastructure, particularly in housing and civil engineering. Traditional SD methods often rely on manual inspections, which are time-consuming, labor-intensive, and prone to human error, especially in complex environments such as dense urban settings or aging buildings with deteriorated materials. Recent advances in autonomous systems—such as Unmanned Aerial Vehicles (UAVs) and climbing robots—have shown promise in addressing these limitations by enabling efficient, real-time data collection. However, challenges persist in accurately detecting and analyzing structural defects (e.g., masonry cracks, concrete spalling) amidst cluttered backgrounds, hardware constraints, and the need for multi-scale feature integration. The integration of machine learning (ML) and deep learning (DL) has revolutionized SD by enabling automated feature extraction and robust defect recognition. For instance, RepConv architectures have been widely adopted for multi-scale object detection, while attention mechanisms like TAM (Technology Acceptance Model) have improved spatial feature fusion in complex scenes. Nevertheless, existing works often focus on singular sensing modalities (e.g., UAVs alone) or neglect the fusion of complementary data streams (e.g., ground-based robot imagery) to enhance detection accuracy. Furthermore, computational redundancy in multi-scale processing and inconsistent bounding box regression in detection frameworks remain underexplored. This study addresses these gaps by proposing a generalized safety inspection system that synergizes UAV and stair-climbing robot data. We introduce a novel multi-scale targeted feature extraction path (Rep-FasterNet TAM block) to unify automated RepConv-based feature refinement with dynamic-scale fusion, reducing computational overhead while preserving critical structural details. For detection, we combine traditional methods with remote sensor fusion to mitigate feature loss during image upsampling/downsampling, supported by a structural model GIOU [Mathematical Definition: GIOU = IOU − (C − U)/C] that enhances bounding box regression through shape/scale-aware constraints and real-time analysis. By siting our work within the context of recent reviews on ML/DL for SD, we demonstrate how our hybrid approach bridges the gap between autonomous inspection hardware and AI-driven defect analysis, offering a scalable solution for large-scale housing safety assessments. In response to challenges in detecting objects accurately during housing safety assessments—including large/dense objects, complex backgrounds, and hardware limitations—we propose a generalized inspection system leveraging data from UAVs and stair-climbing robots. To address multi-scale feature extraction inefficiencies, we design a Rep-FasterNet TAM block that integrates RepConv for automated feature refinement and a multi-scale attention module to enhance spatial feature consistency. For detection, we combine dynamic-scale remote feature fusion with traditional methods, supported by a structural GIOU model that improves bounding box regression through shape/scale constraints and real-time analysis. Experiments demonstrate that our system increases masonry/concrete assessment accuracy by 11.6% and 20.9%, respectively, while reducing manual drawing restoration workload by 16.54%. This validates the effectiveness of our hybrid approach in unifying autonomous inspection hardware with AI-driven analysis, offering a scalable solution for SD in housing infrastructure. Full article

Small-object detection (SOD) remains an important and growing challenge in computer vision and is the backbone of many applications, including autonomous vehicles, aerial surveillance, medical imaging, and industrial quality control. Small objects, in pixels, lose discriminative features during deep neural network processing, making them difficult to disentangle from background noise and other artifacts. This survey presents a comprehensive and systematic review of the SOD advancements between 2023 and 2025, a period marked by the maturation of transformer-based architectures and a return to efficient, realistic deployment. We applied the PRISMA methodology for this work, yielding 112 seminal works in the field to ensure the robustness of our foundation for this study. We present a critical taxonomy of the developments since 2023, arranged in five categories: (1) multiscale feature learning; (2) transformer-based architectures; (3) context-aware methods; (4) data augmentation enhancements; and (5) advancements to mainstream detectors (e.g., YOLO). Third, we describe and analyze the evolving SOD-centered datasets and benchmarks and establish the importance of evaluating models fairly. Fourth, we contribute a comparative assessment of state-of-the-art models, evaluating not only accuracy (e.g., the average precision for small objects (AP_S)) but also important efficiency (FPS, latency, parameters, GFLOPS) metrics across standardized hardware platforms, including edge devices. We further use data-driven case studies in the remote sensing, manufacturing, and healthcare domains to create a bridge between academic benchmarks and real-world performance. Finally, we summarize practical guidance for practitioners, the model selection decision matrix, scenario-based playbooks, and the deployment checklist. The goal of this work is to help synthesize the recent progress, identify the primary limitations in SOD, and open research directions, including the potential future role of generative AI and foundational models, to address the long-standing data and feature representation challenges that have limited SOD. Full article

Real-time gait analysis is essential for the development of responsive and reliable motorized prosthetics. Deploying advanced deep learning models on resource-constrained embedded systems, however, remains a major challenge. This proof-of-concept study presents a TinyML-based approach for knee joint angle prediction using convolutional neural networks (CNNs) trained on inertial measurement unit (IMU) signals. Gait data were acquired from four healthy participants performing multiple stride types, and data augmentation strategies were applied to enhance model robustness. Multi-objective optimization was employed to balance accuracy and computational efficiency, yielding specialized CNN architectures tailored for short, natural, and long strides. A lightweight classifier enabled real-time selection of the appropriate specialized model. The proposed framework achieved an average RMSE of 2.05°, representing a performance gain of more than 35% compared to a generalist baseline, while maintaining reduced inference latency (16.8 ms) on a $40 embedded platform (Sipeed MaixBit with Kendryte K210). These findings demonstrate the feasibility of deploying compact and specialized deep learning models on low-cost hardware, enabling affordable prosthetic solutions with real-time responsiveness. This work contributes to advancing intelligent assistive technologies by combining efficient model design, hardware-aware optimization, and clinically relevant gait prediction performance. Full article