Search Results (298)

In order to achieve better accuracy, modern models have become increasingly large, leading to an exponential increase in computational load, making it challenging to apply them to edge computing. Binary neural networks (BNNs) are models that quantize the filter weights and activations to 1-bit. These models are highly suitable for small chips like advanced RISC machines (ARMs), field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), system-on-chips (SoCs) and other edge computing devices. To design a model that is more friendly to edge computing devices, it is crucial to reduce the floating-point operations (FLOPs). Batch normalization (BN) is an essential tool for binary neural networks; however, when convolution layers are quantized to 1-bit, the floating-point computation cost of BN layers becomes significantly high. This paper aims to reduce the floating-point operations by removing the BN layers from the model and introducing the scaled weight standardization convolution (WS-Conv) method to avoid the significant accuracy drop caused by the absence of BN layers, and to enhance the model performance through a series of optimizations, adaptive gradient clipping (AGC) and knowledge distillation (KD). Specifically, our model maintains a competitive computational cost and accuracy, even without BN layers. Furthermore, by incorporating a series of training methods, the model’s accuracy on CIFAR-100 is 0.6% higher than the baseline model, fractional activation BNN (FracBNN), while the total computational load is only 46% of the baseline model. With unchanged binary operations (BOPs), the FLOPs are reduced to nearly zero, making it more suitable for embedded platforms like FPGAs or other edge computers. Full article

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

Missile-borne active phased array antennas have been widely used in missile guidance for their beam agility, multifunctionality, and strong anti-interference capabilities. However, due to space constraints on the platform and difficulty in heat dissipation, the thermal power consumption of the antenna array can easily lead to excessive temperature, causing two primary issues: first, temperature-induced drift in T/R components, resulting in amplitude and phase errors in the feed current; second, temperature-dependent ripple voltage in the array’s secondary power supply, which exacerbates feed errors. Both issues degrade the electromagnetic performance of the array antenna. To mitigate these effects, this paper investigates feed errors and compensation methods in high-temperature environments. First, a synchronous Buck circuit ripple coefficient model is developed, and an electromagnetic–temperature coupling model is established, incorporating temperature-dependent feed current characteristics, and the law of electromagnetic performance changes is analyzed. On this basis, an electromagnetic performance compensation method based on a genetic algorithm is proposed to optimize the quantization compensation amount of the amplitude and phase of each element under the effect of high temperature. Full article

This work investigates the deployment of Deep Learning (DL) models for network intrusion detection on resource-constrained IoT devices, using the public CICIoT2023 dataset. In particular, we consider the following DL models: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), Multi-Layer Perceptron (MLP). Bayesian optimization is employed to fine-tune the models’ hyperparameters and ensure reliable performance evaluation across both binary (2-class) and multi-class (8-class, 34-class) intrusion detection. Then, the computational complexity of each DL model is analyzed—in terms of the number of Multiply–ACCumulate operations (MACCs), RAM usage, and inference time—through the STMicroelectronics Cube.AI Analyzer tool, with models being deployed on an STM32H7S78-DK board. To assess the practical deployability of the considered DL models, a trade-off score (balancing classification accuracy and computational efficiency) is introduced: according to this score, our experimental results indicate that MLP and TCN outperform the other models. Furthermore, Post-Training Quantization (PTQ) to 8-bit integer precision is applied, allowing the model size to be reduced by more than $90\%$ with negligible performance degradation. This demonstrates the effectiveness of quantization in optimizing DL models for real-world deployment on resource-constrained IoT devices. Full article

Existing inspection systems for railway point machines often suffer from high latency and poor interpretability, which impedes the real-time detection of critical mechanical anomalies, thereby increasing the risks of derailment and leading to cascading schedule delays. Addressing these challenges, this study proposes a lightweight computer vision-based detection framework deployed on the RK3588S edge platform. First, to overcome the accuracy degradation of segmentation networks on constrained edge NPUs, a Sensitivity-Aware Mixed-Precision Quantization and Heterogeneous Scheduling (SMPQ-HS) strategy is proposed. Second, a Multimodal Semantic Diagnostic Framework is constructed. By integrating geometric engagement depths—calculated via perspective rectification—with visual features, a Hard-Constrained Knowledge Embedding Paradigm is designed for the Qwen2.5-VL model. This approach constrains the stochastic reasoning of the Qwen2.5-VL model into standardized diagnostic conclusions. Experimental results demonstrate that the optimized model achieves an inference speed of 38.5 FPS and an mIoU of 0.849 on the RK3588S, significantly outperforming standard segmentation models in inference speed while maintaining high precision. Furthermore, the average depth-estimation error remains approximately 3%, and the VLM-based fault identification accuracy reaches 88%. Overall, this work provides a low-cost, deployable, and interpretable solution for intelligent point machine maintenance under edge-computing constraints. Full article

Multi-fish tracking and behavior analysis in deep-sea cages face two critical challenges: first, the homogeneity of fish appearance and low image quality render appearance-based association unreliable; second, standard linear motion models fail to capture the complex, nonlinear swimming patterns (e.g., turning) of fish, leading to frequent identity switches and fragmented trajectories. To address these challenges, we propose SOD-SORT, which integrates a Constant Turn-Rate and Velocity (CTRV) motion model within an Extended Kalman Filter (EKF) framework into DeepOCSORT, a recent observation-centric tracker. Through systematic Bayesian optimization of the EKF process noise (Q), observation noise (R), and ReID weighting parameters, we achieve harmonious integration of advanced motion modeling with appearance features. Evaluations on the DeepBlueI validation set show that SOD-SORT attains IDF1 = 0.829 and reduces identity switches by 13% (93 vs. 107) compared to the DeepOCSORT baseline, while maintaining comparable MOTA (0.737). Controlled ablation studies reveal that naive integration of CTRV-EKF with default parameters degrades performance substantially (IDs: 172 vs. 107 baseline), but careful parameter optimization resolves this motion-appearance conflict. Furthermore, we introduce a statistical quantization method that converts variable-length trajectories into fixed-length feature vectors, enabling effective unsupervised classification of normal and abnormal swimming behaviors in both the Fish4Knowledge coral reef dataset and real-world Deep Blue I cage videos. The proposed approach demonstrates that principled integration of advanced motion models with appearance cues, combined with high-quality continuous trajectories, can support reliable behavior modeling for aquaculture monitoring applications. Full article

The integration of mobile devices into aviation powering electronic flight bags, maintenance logs, and flight planning tools has created a critical and expanding cyber-attack surface. Security for these systems must be not only effective but also transparent, resource-efficient, and certifiable to meet stringent aviation safety standards. This paper presents AVI-SHIELD, a novel framework for developing high-assurance, on-device threat detection. Our methodology, grounded in the MITRE ATT&CK^® framework, models credible aviation-specific threats to generate the AviMal-TinyX dataset. We then design and optimize a set of compact, interpretable detection algorithms through quantization and pruning for deployment on resource-constrained hardware. Evaluation demonstrates that AVI-SHIELD achieves 97.2% detection accuracy on AviMal-TinyX while operating with strict resource efficiency (<1.5 MB model size, <35 ms inference time and <0.1 Joules per inference) on both Android and iOS. The framework provides crucial decision transparency through integrated, on-device analysis of detection results, adding a manageable overhead (~120 ms) only upon detection. Its successful deployment on both Android and iOS demonstrates that AVI-SHIELD can provide a uniform security posture across heterogeneous device fleets, a critical requirement for airline operations. This work provides a foundational approach for deploying certifiable, edge-based security that delivers the mandatory offline protection required for safety critical mobile aviation applications. Full article

This study proposes an end-to-end (E2E) optimization methodology for a white blood cell (WBC) cell segmentation and counting (CSC) pipeline with a focus on deployment to low-power Artificial Intelligence of Things (AIoT) devices. The proposed framework addresses not only the selection of the segmentation model but also the corresponding loss function design, watershed threshold optimization for cell counting, and model compression strategies to balance accuracy, latency, and model size in embedded AIoT applications. For segmentation model selection, UNet, UNet++, ResUNet, EffUNet, FPN, BiFPN, PFPN, Cell-ViT, Evit-UNet and MAXVitUNet were employed, and three types of loss functions—binary cross-entropy (BCE), focal loss, and Dice loss—were utilized for model training. For cell-counting accuracy optimization, a distance transform-based watershed algorithm was applied, and the optimal threshold value was determined experimentally to lie within the range of 0.4 to 0.9. Quantization and pruning techniques were also considered for model compression. Experimental results demonstrate that using an FPN model trained with focal loss and setting the watershed threshold to 0.65 yields the optimal configuration. Compared to the latest baseline techniques, the proposed CSC E2E pipeline achieves a 21.1% improvement in cell-counting accuracy while reducing model size by 74.5% and latency by 16.8% through model compression. These findings verify the effectiveness of the proposed optimization strategy as a lightweight and efficient solution for real-time biomedical applications on low-power AIoT devices. Full article

We present a hybrid end-to-end learned image compression framework that combines a CNN-based variational autoencoder (VAE) with an efficient hierarchical Swin Transformer to address the limitations of existing entropy models in capturing global dependencies under computational constraints. Traditional VAE-based codecs typically rely on CNN-based priors with localized receptive fields, which are insufficient for modelling the complex, high-dimensional dependencies of the latent space, thereby limiting compression efficiency. While fully global transformer-based models can capture long-range dependencies, their high computational complexity makes them impractical for high-resolution image compression. To overcome this trade-off, our approach couples a CNN-based VAE with a patch-based hierarchical Swin Transformer hyperprior that employs shifted window self-attention to effectively model both local and global contextual information while maintaining computational efficiency. The proposed framework tightly integrates this expressive entropy model with an end-to-end differentiable quantization module, enabling joint optimization of the complete rate-distortion objective. By learning a more accurate probability distribution of the latent representation, the model achieves improved bitrate estimation and a more compact latent representation, resulting in enhanced compression performance. We validate our approach on the widely used Kodak, JPEG AI, and CLIC datasets, demonstrating that the proposed hybrid architecture achieves superior rate-distortion performance, delivering higher visual quality at lower bitrates compared to methods relying on simpler CNN-based entropy priors. This work demonstrates the effectiveness of integrating efficient transformer architectures into learned image compression and highlights their potential for advancing entropy modelling beyond conventional CNN-based designs. Full article

The automation of quality control in agricultural food processing, particularly the detection of incomplete peeling in taro, constitutes a critical frontier for ensuring food safety and optimizing production efficiency in the Industry 4.0 era. However, this domain is fraught with significant technical challenges, primarily stemming from the inherent visual characteristics of residual peel: extremely minute scales relative to the vegetable body, highly irregular morphological variations, and the dense occlusion of objects on industrial conveyor belts. To address these persistent impediments, this study introduces a comprehensive solution comprising a specialized dataset and a novel detection architecture. We established the Taro Peel Industrial Dataset (TPID), a rigorously annotated collection of 18,341 high-density instances reflecting real-world production conditions. Building upon this foundation, we propose MDB-YOLO, a lightweight, multi-dimensional bionic detection model evolved from the YOLOv8s architecture. The MDB-YOLO framework integrates a synergistic set of innovations designed to resolve specific detection bottlenecks. To mitigate the conflict between background texture interference and tiny target detection, we integrated the C2f_EMA module with a Wise-IoU (WIoU) loss function, a combination that significantly enhances feature response to low-contrast residues while reducing the penalty on low-quality anchor boxes through a dynamic non-monotonic focusing mechanism. To effectively manage irregular peel shapes, a dynamic feature processing chain was constructed utilizing DySample for morphology-aware upsampling, BiFPN_Concat2 for weighted multi-scale fusion, and ODConv2d for geometric preservation. Furthermore, to address the issue of missed detections caused by dense occlusion in industrial stacking scenarios, Soft-NMS was implemented to replace traditional greedy suppression mechanisms. Experimental validation demonstrates the superiority of the proposed framework. MDB-YOLO achieves a mean Average Precision (mAP50-95) of 69.7% and a Recall of 88.0%, significantly outperforming the baseline YOLOv8s and advanced transformer-based models like RT-DETR-L. Crucially, the model maintains high operational efficiency, achieving an inference speed of 1.1 ms on an NVIDIA A100 and reaching 27 FPS on an NVIDIA Jetson Xavier NX using INT8 quantization. These findings confirm that MDB-YOLO provides a robust, high-precision, and cost-effective solution for real-time quality control in agricultural food processing, marking a significant advancement in the application of computer vision to complex biological targets. Full article

This expository paper provides a unified and pedagogical introduction to optimal quantization for probability measures supported on spherical curves and discrete subsets of the sphere, emphasizing both continuous and discrete settings. We first present a detailed geometric and analytical foundation for intrinsic quantization on the unit sphere, including definitions of great and small circles, spherical triangles, geodesic distance, Slerp interpolation, the Fréchet mean, spherical Voronoi regions, centroid conditions, and quantization dimensions. Building upon this framework, we develop explicit continuous and discrete quantization models on spherical curves, namely great circles, small circles, and great circular arcs—supported by rigorous derivations and pedagogical exposition. For uniform continuous distributions, we compute optimal sets of n-means and the associated quantization errors on these curves; for discrete distributions, we analyze antipodal, equatorial, tetrahedral, and finite uniform configurations, illustrating convergence to the continuous model. The central conclusion is that for a uniform probability distribution supported on a one-dimensional geodesic subset of total length L, the optimal n-means form a uniform partition and the quantization error satisfies $V_n=L^2/\left(12n^2\right)$.The exposition emphasizes geometric intuition, detailed derivations, and clear step-by-step reasoning, making it accessible to beginning graduate students and researchers entering the study of quantization on manifolds. This article is intended as an expository and tutorial contribution, with the main emphasis on geometric reformulation and pedagogical clarity of intrinsic quantization on spherical curves, rather than on the development of new asymptotic quantization theory. Full article

Passenger-gathering behavior often triggers safety incidents such as stampedes due to overcrowding, posing significant challenges to public order maintenance and passenger safety. Traditional early warning algorithms for passenger-gathering behavior typically perform only global modeling of image appearance, neglecting the analysis of individual passenger actions in practical 3D physical space, leading to high false-alarm and missed-alarm rates. To address this issue, we decompose the modeling process into two stages: human pose estimation and gathering behavior recognition. Specifically, the pose of each individual in 3D space is first estimated from images, and then fused with global features to complete the early warning. This work focuses on the former stage and aims to develop an accurate and efficient human pose estimation model capable of real-time inference on resource-constrained devices. To this end, we propose a 3D human pose estimation framework that integrates a hybrid spatio-temporal Transformer with three collaborative agents. First, a reinforcement learning-based architecture search agent is designed to adaptively select among Global Self-Attention, Window Attention, and External Attention for each block to optimize the model structure. Second, a feedback optimization agent is developed to dynamically adjust the search process, balancing exploration and convergence. Third, a quantization agent is employed that leverages quantization-aware training (QAT) to generate an INT8 deployment-ready model with minimal loss in accuracy. Experiments conducted on the Human3.6M dataset demonstrate that the proposed method achieves a mean per joint position error (MPJPE) of 42.15 mm with only 4.38 M parameters and 19.39 GFLOPs under FP32 precision, indicating substantial potential for subsequent gathering behavior recognition tasks. Full article

Auditability and regulatory compliance are increasingly required for deploying large language models (LLMs). Prior work typically targets isolated stages such as training or unlearning and lacks a unified mechanism for verifiable accountability across model updates. This paper presents AuditableLLM, a lightweight framework that decouples update execution from an audit-and-verification layer and records each update as a hash-chain-backed, tamper-evident audit trail. The framework supports parameter-efficient fine-tuning such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA), full-parameter optimization, continual learning, and data unlearning, enabling third-party verification without access to model internals or raw logs. Experiments on LLaMA-family models with LoRA adapters and the MovieLens dataset show negligible utility degradation (below 0.2% in accuracy and macro-F1) with modest overhead (3.4 ms/step; 5.7% slowdown) and sub-second audit validation in the evaluated setting. Under a simple loss-based membership inference attack on the forget set, the audit layer does not increase membership leakage relative to the underlying unlearning algorithm. Overall, the results indicate that hash-chain-backed audit logging can be integrated into practical LLM adaptation, update, and unlearning workflows with low overhead and verifiable integrity. Full article

This study investigates bit error rate (BER)-constrained power allocation for uplink non-orthogonal multiple access (NOMA) systems in which a base station employs one-bit analog-to-digital converters. Although one-bit quantization significantly reduces hardware costs and receiver power consumption, it also introduces severe nonlinear distortions that degrade detection performance. To address this challenge, a pairwise error probability expression is first derived for the one-bit quantized uplink NOMA model, from which an analytical upper bound on the BER is obtained. Based on this characterization, a fairness-driven max–min power allocation strategy is formulated to minimize the BER of the worst-performing user. A closed-form solution for the optimal power allocation is obtained under binary phase-shift keying (BPSK) signaling. Simulation results verify the tightness of the analytical BER bound and demonstrate that the proposed power allocation scheme provides noticeable BER improvements that compensate for the performance degradation caused by one-bit quantization. Full article

To address the issues of efficiency and robustness in UAV trajectory planning under complex environments, this paper proposes a Graph Multi-Head Attention Policy Optimization Network (G-MAPONet) algorithm that integrates Graph Attention (GAT), Multi-Head Attention (MHA), and Group Relative Policy Optimization (GRPO). The algorithm adopts a three-layer architecture of “GAT layer for local feature perception–MHA for global semantic reasoning–GRPO for policy optimization”, comprehensively achieving the goals of dynamic graph convolution quantization and global adaptive parallel decoupled dynamic strategy adjustment. Comparative experiments in multi-dimensional spatial environments demonstrate that the Gat_Mha combined mechanism exhibits significant superiority compared to single attention mechanisms, which verifies the efficient representation capability of the dual-layer hybrid attention mechanism in capturing environmental features. Additionally, ablation experiments integrating Gat, Mha, and GRPO algorithms confirm that the dual-layer fusion mechanism of Gat and Mha yields better improvement effects. Finally, comparisons with traditional reinforcement learning algorithms across multiple performance metrics show that the G-MAPONet algorithm reduces the number of convergence episodes (NCE) by an average of more than 19.14%, increases the average reward (AR) by over 16.20%, and successfully completes all dynamic path planning (PPTC) tasks; meanwhile, the algorithm’s reward values and obstacle avoidance success rate are significantly higher than those of other algorithms. Compared with the baseline APF algorithm, its reward value is improved by 8.66%, and the obstacle avoidance repetition rate is also enhanced, which further verifies the effectiveness of the improved G-MAPONet algorithm. In summary, through the dual-layer complementary mode of GAT and MHA, the G-MAPONet algorithm overcomes the bottlenecks of traditional dynamic environment modeling and multi-scale optimization, enhances the decision-making capability of UAVs in unstructured environments, and provides a new technical solution for trajectory planning in intelligent logistics and distribution. Full article