Search Results (971)

The growing diversity of wireless standards and complex real-world channel effects render automatic modulation classification (AMC) increasingly challenging for spectrum monitoring and edge intelligence. However, most competitive deep-learning-based AMC networks still require ${10}^5$–${10}^6$ parameters, exceeding the memory available on resource-constrained edge platforms. We propose LGP-Net, a lightweight gated-fusion network that pairs a physics-informed expert branch with a compact temporal encoder built from depthwise separable convolution (DSConv), squeeze-and-excitation (SE) attention, and a single-layer gated recurrent unit (GRU). Specifically, unlike other dual-branch structures that directly concatenate the outputs of both pathways, this work designs a lightweight gating unit that requires no external signal-to-noise ratio (SNR) labels and adaptively reweights the two pathways according to signal-quality degradation. With fewer than 40 K parameters, a peak activation footprint of 26.00 KB and an amortised inference latency of 9.7 $\mathsf{\mu }$s per sample under GPU acceleration, LGP-Net attains 65.00% overall accuracy on RadioML 2016.10B (91.48% at 0 dB) and 62.76% on RadioML 2016.10A, placing it in a competitive accuracy–efficiency regime relative to architectures consuming 5× to 500× more parameters. These characteristics support deployment-oriented feasibility under memory-constrained edge settings and high-throughput spectrum-monitoring pipelines. Full article

With the rapid deployment of deep neural networks (DNNs) on edge devices, traditional hardware accelerators face significant challenges in terms of data security, computational redundancy caused by sparsity, and uneven utilization of on-chip resources. This paper proposes SaE-FPGA, a secure and efficient DNN accelerator designed specifically for edge FPGA platforms. The architecture introduces three core innovations: (1) Hash-Bypass Processing Unit (HBPU): Integrating a high-speed SHA-256 hardware engine with a hash-sparse bitmap mechanism, it enables real-time data integrity verification within a single clock cycle while skipping computations for redundant zero-value data. (2) Flexible Mixed-Precision Processing Element (FMP): By reconfiguring idle BRAM and LUT resources into an active lookup table multiplication engine, it overcomes the physical bit-width limitations of DSP blocks and supports INT8/INT6/INT4 mixed-precision multiplication. (3) Multi-mode Reconfigurable Streaming Frame (MRSF): A sparse-aware, elastic load balancing and data routing mechanism designed to mask long memory access latencies and ensure high hardware resource utilization. Experimental results on the Zynq 7045 platform demonstrate that SaE-FPGA reduces redundant computations by 23.2% while maintaining high precision and minimizing precision loss. The system effectively mitigates the risk of physical tampering. When tested on ResNet-50, it achieved a 27.2% improvement in energy efficiency and a 2.97× speedup compared to DSP-based FPGA solutions. Furthermore, by fully exploiting the hybrid BRAM-LUT and DSP configuration, the proposed accelerator achieves a remarkable peak throughput of 782.4 GOPS. Full article

Aiming at the problems of weak background adaptive ability, high dependence on edge features, high computational complexity of some traditional license plate location algorithms, high deployment cost and strong training dependence of location model based on deep learning, this paper proposes a fast license plate location algorithm based on statistical color features. The algorithm uses the HSV color space as the main processing channel, and quantifies the regional color distribution characteristics by constructing the hue histogram and calculating its standard deviation and other statistics, which significantly improves the discrimination and illumination adaptability of the license plate mask in complex background. Compared with the lightweight deep learning models such as “You Only Look Once Version 12 Nano”, this algorithm does not need GPU acceleration and model loading, eliminates the need for data training, significantly reduces the deployment cost and complexity, and can run efficiently on the general computing platform. The experimental results show that compared with the YOLOv12n model, the average processing time of this algorithm is shortened by 30.81% (when YOLOv12n is evaluated with GPU) or 48.42% (when YOLOv12n is evaluated with CPU) at the cost of sacrificing about 5.8% positioning accuracy. The positioning accuracy still reaches 93.7%, demonstrating high processing efficiency and excellent platform adaptability. The algorithm has the advantages of being lightweight, efficient and interpretable, and is especially suitable for intelligent parking lots, edge devices and other scenes sensitive to real time, cost and energy consumption. Full article

Transitioning predictive maintenance from expensive, high-frequency piezoelectric sensors to affordable, edge-deployed MEMS sensors poses a significant challenge in industrial tool condition monitoring (TCM). Both technologies differ in signal quality, frequency capability, robustness, and reliability, which would affect how accurately machine faults can be detected. This work presents a systematic framework to bridge this gap, enabling real-time tool wear prediction and cross-sensor transferability. The methodology employs unsupervised Wavelet Packet Decomposition (WPD) and dynamic programming on high-resolution vibration signals to establish ground-truth wear phases: initial, steady-state, and accelerated. Multi-resolution time-frequency features are extracted and globally ranked using a multi-metric scoring system. A multi-task Bidirectional Long Short-Term Memory (Bi-LSTM) network is then trained to simultaneously predict a continuous wear index and classify discrete wear zones. To ensure model portability, Canonical Correlation Analysis (CCA) is utilised to align the high-fidelity piezoelectric feature space with the lower-frequency MEMS domain. The optimised multi-task Bi-LSTM architecture achieved up to 97.9% zone classification accuracy and a mean absolute error of 0.042 for wear index regression. Furthermore, CCA-based domain adaptation successfully transferred a model trained on piezoelectric data to classify unseen low-cost MEMS sensor data, maintaining a robust 87% accuracy. Combining optimised WPD features with CCA effectively overcomes hardware and sampling rate discrepancies, proving the viability of using low-cost sensors for reliable industrial retrofitting and real-time degradation tracking. Full article

The picking of tea leaves in tea gardens requires multiple batches in the short and valuable tea harvest period. To realize timely and efficient tea plucking, it is feasible to use unmanned aerial vehicles (UAV) for tea shoot detection in large tea gardens. For the typical small targets of tea buds in unmanned aerial vehicle (UAV) aerial images, it is necessary to design an efficient tea buds detection model. In order to improve the accuracy and the speed of the tea buds detection in the UAV images, we designed the SH-CoordMapping hash space mapping algorithm to accelerate the remerging of the detection results into the original image. The C2PSA-BI module and the CARAFE upsampling module are applied to improve detail preservation during feature fusion. A lightweight detection head is further used to reduce redundant computation in the detection stage. By comparing with the traditional detection methods, it can be proved that the SWO sections are necessary for UAV-scale tea shoots detection. Based on the accuracy and the number of model parameters, the YOLO11n model with slice size as 640 and overlap rate as 0.1 performs the best. The TSDet-UAV was deployed on the NVIDIA Jetson Orin NX chip to construct an inspection system capable of real-time acquisition and detection. The experimental results demonstrate that the proposed TSDet-UAV achieves excellent performance, recording an mAP₅₀ of 52.9% on the constructed UAV-TS dataset while maintaining high efficiency. With a parameter size of 2.4 M and a total processing time of 1.32 s per high-resolution image under TensorRT FP16, the processing speed is highly suitable for real-time edge deployment on agricultural UAV platforms. The UAV image-based tea garden shoot inspection platform proposed in this paper can effectively confirm the growth status of tea shoots, assisting farm management in formulating precise picking plans. Full article

This paper presents a comprehensive review on near-infrared (NIR) spectroscopy applied in food analysis, systematically elaborating its core principles, widespread industrial applications, advanced chemometric strategies, and cutting-edge technological progress. NIR spectroscopy (760–2500 nm), characterized by rapid, non-destructive detection and minimal sample preparation, has been widely implemented in quality evaluation and safety monitoring of grains, meat, fruits and vegetables, dairy, fermented products, tea, coffee, and other processed foods, realizing quantitative analysis of nutrients, freshness assessment, texture prediction, adulteration identification, origin tracing, and rapid preliminary screening of toxin/pesticide residues. A series of chemometric methods, including spectral preprocessing (SNV, MSC, S-G smoothing), feature extraction, and variable selection (CARS, PSO-CMW, ICPA), as well as linear/nonlinear modeling algorithms (PLS, SVM, BP-ANN, fuzzy clustering) significantly boost the accuracy and robustness of spectral analysis. Meanwhile, portable NIR devices and online monitoring systems promote on-site and real-time detection in food supply chains. Despite existing challenges such as calibration transfer, matrix interference, and model generalization, innovations like multimodal data fusion, deep learning integration, and intelligent algorithm optimization offer effective solutions. This review not only summarizes the latest research advances of NIR technology in the food field but also emphasizes its significant advantages as a rapid, non-destructive complementary tool to traditional destructive detection methods, providing theoretical support and technical reference for accelerating the industrial translation and standardized application of NIR spectroscopy, and ultimately safeguarding global food quality and safety. Full article

Walnut leaf necrosis causes leaf desiccation and premature abscission, substantially reducing photosynthetic efficiency, impairing fruit development, and ultimately leading to yield loss and quality deterioration. In severe cases, it accelerates branch senescence or even whole-tree mortality, resulting in considerable economic damage to the walnut industry. Rapid and accurate monitoring of this disease is therefore essential for sustainable production. This study aimed to characterize the different stages of walnut leaf necrosis using spectral analysis and develop classification models for stage-specific identification. Leaf samples representing healthy leaves and the early, middle, and late stages of necrosis were analyzed for spectral responses. Sensitive bands were identified using the variable importance in projection (VIP), successive projections algorithm (SPA), and the combined VIP-SPA method, and corresponding vegetation indices were constructed. The selected features were incorporated into classification models based on random forest (RF), extreme gradient boosting (XGBoost), and convolutional neural networks (CNNs). Results revealed that the red-edge (640–700 nm) and near-infrared (720–1000 nm) regions were identified as key diagnostic spectral ranges. Among the vegetation indices evaluated, the Simple Ratio Index (SRI) calculated from reflectance at 705.7 nm and 707.1 nm, the Normalized Difference Index (NDI) using the same band pair, and the Difference Index (DI) derived from 417.1 nm and 638.7 nm emerged as the most sensitive indicators of disease severity. Classification accuracies for different necrosis stages reached 0.9583, 0.9583, and 0.9333, respectively. These findings demonstrate that the identified spectral bands and vegetation indices provide robust tools for monitoring the progression of walnut leaf necrosis. Full article

Edge computing on unmanned aerial vehicles (UAVs) enables low-latency wildfire monitoring by performing visual inference onboard; however, practical deployment is constrained by limited labeled data and resource budgets that often preclude reliance on large GPU servers. This work investigates transfer learning (TL) for UAV-based wildfire smoke and flame detection and evaluates its impact on both detection accuracy and edge deployment performance. We introduce the Aerial Fire and Smoke Essential (AFSE) dataset (282 aerial-view images; classes—smoke and fire), compiled from publicly available wildfire footage and FLAME2. Lightweight YOLO models are fine-tuned using heterogeneous (MS COCO) and homogeneous (FASDD) source pretraining and are assessed using mAP@0.5 together with frames per second (FPS), average inference power, energy consumption, and the normalized energy–delay product (EDP) on an edge computing platform. Results show that TL substantially improves detection accuracy on AFSE, achieving up to 79.2% mAP@0.5, while reducing training time, and improving cross-validation stability. On the tested edge platform, TL does not materially change inference speed or energy use, indicating that accuracy gains from TL do not automatically translate to improved efficiency without additional optimization. Among the evaluated lightweight detectors, YOLOv5n achieves the best mAP@0.5 while maintaining the highest edge device throughput, processing images nearly twice as fast as YOLO11n without hardware acceleration. More broadly, the measured throughput and energy differences among lightweight YOLO variants show that edge model selection should be guided by application-specific accuracy, latency, and energy constraints. Full article

With the 2030 deadline for the United Nations Sustainable Development Goals (SDGs) approaching, there is a growing global urgency to identify innovative, scalable, and inclusive AI-based or AI-enabled solutions capable of accelerating progress across sectors. Yet the benefits of AI remain unevenly distributed, particularly in low-resource settings where limited infrastructure, cost barriers, and unequal access to skills constrain adoption. This paper explores how Tiny Machine Learning (TinyML)—a low-power, low-cost edge AI paradigm—offers a concrete technological pathway aligned with the principles of Frugal AI, providing accessible, energy-efficient, and context-adapted tools for sustainable development. We evaluate how participatory citizen science, when combined with TinyML, enables communities to co-create AI applications that address locally defined challenges in environmental monitoring, agriculture, and public health. Drawing on early outcomes from workshops, collaborative projects, and innovation competitions, the paper examines how TinyML-enabled participatory approaches cultivate technical skills, stimulate grassroots entrepreneurship, and generate prototypes suited to low-resource environments. Using a qualitative multiple-case study of 50 participatory TinyML initiatives across 22 countries, we analyse how frugal edge-AI practices support skills formation, prototype development, and early entrepreneurial engagement. The analysis identifies the pedagogical, technical, and institutional frameworks that support successful participatory AI initiatives, emphasizing open educational resources, cross-sector partnerships, and community-driven problem formulation. We introduce the Frugal Edge AI Lean Canvas to help innovators identify novelty, ethical implications, and measurable impact. TinyML-based participatory innovation offers a promising route for accelerating SDG progress by expanding who can create, deploy, and benefit from AI. Full article

As global population aging intensifies, fall-related injuries among the elderly have become a critical public health concern. Existing fall detection methods based on wearable IMUs all extract features in the sensor’s body frame, failing to exploit the information embedded in sensor signals. Some higher-performing methods incorporate magnetometer-fused Euler angles to enrich features, but their dependence on specific hardware and fusion algorithms makes exact replication during deployment difficult. In contrast, the proposed MadgwickFall-Net relies on acceleration and angular velocity, and, to the best of our knowledge, for the first time introduces the Madgwick algorithm into fall detection to transform inertial signals into a gravity-aligned global coordinate system. A four-branch parallel architecture processes signals from both coordinate frames, fully exploiting the complementarity between dual-frame signals. Cross-validation on the KFall dataset using 5-fold subject-independent stratification demonstrates an F1-Score of 0.9824 and accuracy of 98.36%, specifically, four main evaluation indicators outperform all comparison models. With only 59.7 KB parameters, the model is suitable for edge device deployment. Rolling inference experiments demonstrate a median pre-impact lead time of 390 ms. MadgwickFall-Net offers a practical and deployable solution for real-world wearable fall detection systems, demonstrating strong potential for protecting elderly individuals in daily life scenarios. Full article

As a critical component of sustainable energy systems, the efficient maintenance of photovoltaic (PV) panels is essential. While deep learning is an important approach for PV panel defect detection, the high complexity of existing models and their substantial computational demand make deployment on edge platforms difficult. This paper studies an acceleration method for photovoltaic panel defect detection on the Zynq-7020 heterogeneous platform. We design DARE-YOLO, a lightweight network for photovoltaic panel defect detection, together with a Zynq-based accelerator. In DARE-YOLO, we introduce RepConv and a lightweight single-path backbone to reduce the memory bandwidth overhead caused by multi-branch structures. We further design a Dilated Context Block (DCB) and a Dual-scale Decoupled Head (DDH), which effectively improve the detection accuracy of DARE-YOLO. On the Zynq platform, we develop the accelerator through a mixed fixed-point quantization strategy, a custom convolution IP core, and pipeline unrolling. These optimizations reduce data access latency, improve computational parallelism, and increase computational throughput. Experimental results show that DARE-YOLO achieves 93.84% mAP@0.5 with only 6.4 M parameters. The accelerator has a total on-board power consumption of only 1.95 W, while delivering a throughput of 37.5 GOPS, an energy efficiency of 19.23 GOPS/W. The image inference latency is 661.3 ms. This low-power, high-efficiency co-design paradigm ensures the long-term reliability of renewable energy facilities. Full article

The spread of multimodal misinformation demands verification that is both accurate and fast while keeping knowledge current. Large language models are powerful but costly and slow, and their static knowledge can lag behind events. We introduce EC-MFR, a hierarchical framework that divides work between edge and the cloud. The system first optionally decomposes the claim into a few targeted sub-claims to guide retrieval, retrieves text and image evidence, and then compresses it into a small set of question–answer items using a lightweight, quantized multimodal language model deployed at the edge. A compact verifier on the edge predicts a label with calibrated confidence. If confidence is high, the decision is returned immediately. If confidence is low, the claim is sent to the cloud where retrieval can be expanded and the reasoning can be redone by a stronger verifier. This design offers three core benefits. It makes reasoning explicit through question–answer items, which shortens prompts and improves auditability. It improves retrieval recall via a light decomposition step that produces targeted sub-queries. Finally, it lets most easy claims finish on the edge to reduce cost and latency while preserving accuracy on difficult claims by allowing the cloud to broaden evidence and refine reasoning. Experiments on MOCHEG and AVERITEC validate the approach. Notably, EC-MFR achieves highly competitive accuracy of 54.10% on the multimodal MOCHEG dataset, and reaches 68.80% on AVERITEC under realistic retrieval settings, outperforming the GPT-4o cloud-only baseline by 6.6 percentage points. Furthermore, system-level profiling on edge hardware demonstrates that EC-MFR reduces processing costs by 51.8% and accelerates inference latency by 2.4× for edge-resolved claims, confirming a highly favorable accuracy–efficiency trade-off compared to existing multimodal fact-checking systems. We also formalize routing and efficiency and analyze calibration and retrieval. Full article

Real-time psychological stress detection on wearable and edge devices requires models that are accurate, computationally efficient, and small enough for on-device deployment. This paper proposes a Micro-Attention CNN Hybrid Architecture for stress recognition using wearable bio-signals. The model uses six sensor channels, namely tri-axial acceleration, electrodermal activity, heart rate, and skin temperature, and classifies three stress levels: no stress, low stress, and high stress. This study is conducted on a public wearable sensor dataset collected from 15 nurses during hospital work, providing a realistic benchmark for continuous stress monitoring under practical conditions. The proposed architecture combines one-dimensional and depthwise separable convolutions with a lightweight attention module to emphasize the most informative temporal patterns in short multivariate signal segments. To support deployment on resource-constrained devices, we further apply structured pruning, selective quantization-aware training, and post-training quantization. The full-precision model achieves a Macro-F1 score of 99.63%, while the final compressed model retains 98.03% Macro-F1 with a model size of 1.76 kilobytes and a CPU inference latency of 0.40 ms. Additional analyses show that most residual errors occur near the boundary between low stress and neighboring classes, while simple post-compression calibration improves reliability. These results demonstrate that accurate and low-latency stress detection using wearable bio-signals is feasible on compact edge hardware without transmitting raw sensor streams off-device. Full article

As the number of unmanned aerial vehicles (UAVs) and the volume of computational tasks increase in UAV ad hoc networks (UAVANET), the solution space for task allocation strategies grows exponentially. In practical emergency scenarios with concurrent multi-user access, multi-UAV systems equipped with mobile edge computing (MEC) devices face challenges such as limited computing resources and imbalanced task distribution during task offloading. To address these challenges, this paper proposes an adaptive task allocation algorithm named AUSTA-DQHO (Adaptive UAV Swarm Task Allocation using Deep Q-networks and Genetic Algorithms Hybrid Optimization), which combines Deep Q-Network (DQN) with Genetic Algorithm (GA), aiming to optimize computational task scheduling and minimize both the total task delay and the variance in task delays. First, we introduce a multi-UAV-assisted MEC application framework. In this framework, UAVs equipped with high-performance computing modules are deployed as airborne servers in the target area, providing data offloading and task computation support for IoT devices. Next, to tackle the optimization problem, we replace the random action selection process in DQN with a hybrid strategy that incorporates heuristic methods—specifically, GA and greedy algorithms—to perform global search and make more effective decisions for optimal task allocation for each offloading request. Furthermore, to accelerate the convergence of the AUSTA-DQHO policy while ensuring global optimality, we introduce a pre-clustering mechanism and a dynamic weighting factor for randomly generated task offloading requests in the target area. These mechanisms effectively reduce the solution space and ensure that optimal actions are learned at different stages of the training process. Experimental results demonstrate that the proposed algorithm achieves a task latency reduction of 18.72% and a load balancing improvement of 98.72%, surpassing the performance of the other algorithms. Additionally, we explore the optimal number of UAVs under given environmental conditions to minimize the waste of computing resources. Full article

Glaciers are critical indicators of global climate change, and their accelerated retreat has profound implications for sea-level rise, water resources, and ecosystem stability. Accurate and timely mapping of glacier extent is essential for monitoring these changes. This review provides a comprehensive overview of the evolution of remote sensing techniques for glacier mapping, charting the progression from traditional spectral indices to the current state-of-the-art machine learning (ML) and deep learning (DL) models. We analyze the strengths and limitations of various methods, including the computational efficiency of indices like the Normalized Difference Snow Index (NDSI), the classificatory power of ML algorithms like Random Forest (RF), and the superior performance of DL architectures, particularly U-Net and its variants, for semantic segmentation of glacier mapping. Our analysis highlights a clear trend towards automated, data-driven approaches that have significantly enhanced the accuracy and scale of glacier delineation. However, progress is slowed by key challenges, most importantly the difficulty in getting accurate ‘ground truth’ data due to a lack of standardized, high-resolution training and validation datasets. Other key limitations include an over-reliance on a few model architectures and the need to bridge the gap between research-level accuracy and operational, real-time monitoring systems. Future progress in the field will depend on community-led efforts to create robust benchmark datasets, explore more diverse and efficient model architecture, develop sophisticated data fusion techniques, and improve model transferability and uncertainty quantification. By integrating cutting-edge AI with improved data practices, the remote sensing community can deliver the crucial data needed to understand and respond to the impacts of a changing climate. Full article