Deployment Model Research Articles

A Performance Analysis of You Only Look Once Models for Deployment on Constrained Computational Edge Devices in Drone Applications

Advancements in embedded systems and Artificial Intelligence (AI) have enhanced the capabilities of Unmanned Aircraft Vehicles (UAVs) in computer vision. However, the integration of AI techniques o-nboard drones is constrained by their processing capabilities. In this sense, this study evaluates the deployment of object detection models (YOLOv8n and YOLOv8s) on both resource-constrained edge devices and cloud environments. The objective is to carry out a comparative performance analysis using a representative real-time UAV image processing pipeline. Specifically, the NVIDIA Jetson Orin Nano, Orin NX, and Raspberry Pi 5 (RPI5) devices have been tested to measure their detection accuracy, inference speed, and energy consumption, and the effects of post-training quantization (PTQ). The results show that YOLOv8n surpasses YOLOv8s in its inference speed, achieving 52 FPS on the Jetson Orin NX and 65 fps with INT8 quantization. Conversely, the RPI5 failed to satisfy the real-time processing needs in spite of its suitability for low-energy consumption applications. An analysis of both the cloud-based and edge-based end-to-end processing times showed that increased communication latencies hindered real-time applications, revealing trade-offs between edge (low latency) and cloud processing (quick processing). Overall, these findings contribute to providing recommendations and optimization strategies for the deployment of AI models on UAVs.

Deep Learning Models for Classifying Apple Leaf Diseases: A Review

Accurately and efficiently classifying apple leaf diseases is crucial for early detection, management, and enhanced agricultural productivity. With advancements in deep learning, particularly Convolutional Neural Networks (CNNs), the potential for automated plant disease detection has greatly increased. The research comprehensively reviews deep learning models applied to apple leaf disease classification, focusing on state-of-the-art architectures such as MobileNet, Xception, NASNetLarge, and DenseNet-201. The existing studies on the application of these models in plant disease detection, comparing their strengths, limitations, and suitability for agricultural applications have been critically analyzed. Performance metrics like accuracy, precision, recall, and F1-score are examined to identify the most effective models for detecting various apple leaf diseases under real-world conditions. The research highlights key challenges and opportunities in optimizing deep learning models for practical deployment, contributing to sustainable farming practices and agricultural disease management advancements.

Optimizing deployment model of flexible interconnection systems for new energy 6/10 kV medium-voltage grids

Abstract To improve the voltage instability, overloading, and suboptimal economic performance resultant from the integration of flexible interconnections within new energy 6/10 kV medium-voltage distribution grids, we present a novel methodology for the deployment and optimization of flexible interconnection systems. We incorporate generative adversarial networks to assimilate and delineate prototypical deployment scenarios, mitigating voltage disparities. Then, we formulate a two-tiered deployment model for interconnection systems that synchronizes across various time scales, enabling a comprehensive assessment of safety and cost variables intrinsic to the distribution grid framework. Results proved that our refined deployment strategies and models bolster the operational dependability and economic efficacy.

Knowledge distillation with adapted weight

Although large models have shown a strong capacity to solve large-scale problems in many areas including natural language and computer vision, their voluminous parameters are hard to deploy in a real-time system due to computational and energy constraints. Addressing this, knowledge distillation through Teacher-Student architecture offers a sustainable pathway to compress the knowledge of large models into more manageable sizes without significantly compromising performance. To enhance the robustness and interpretability of this framework, it is critical to understand how individual training data impact model performance, which is an area that remains underexplored. We propose the Knowledge Distillation with Adaptive Influence Weight (KD-AIF) framework which leverages influence functions from robust statistics to assign weights to training data, grounded in the four key SAFE principles: Sustainability, Accuracy, Fairness, and Explainability. This novel approach not only optimizes distillation but also increases transparency by revealing the significance of different data. The exploration of various update mechanisms within the KD-AIF framework further elucidates its potential to significantly improve learning efficiency and generalization in student models, marking a step toward more explainable and deployable Large Models. KD-AIF is effective in knowledge distillation while also showing exceptional performance in semi-supervised learning with outperforms existing baselines and methods in multiple benchmarks (CIFAR-100, CIFAR-10-4k, SVHN-1k, and GLUE).

Cloud-Based License Plate Recognition: A Comparative Approach Using You Only Look Once Versions 5, 7, 8, and 9 Object Detection

Cloud-based license plate recognition (LPR) systems have emerged as essential tools in modern traffic management and security applications. Determining the best approach remains paramount in the field of computer vision. This study presents a comparative analysis of various versions of the YOLO (You Only Look Once) object detection models, namely, YOLO 5, 7, 8, and 9, applied to LPR tasks in a cloud computing environment. Using live video, we performed experiments on YOLOv5, YOLOv7, YOLOv8, and YOLOv9 models to detect number plates in real time. According to the results, YOLOv8 is reported the most effective model for real-world deployment due to its strong cloud performance. It achieved an accuracy of 78% during cloud testing, while YOLOv5 showed consistent performance with 71%. YOLOv7 performed poorly in cloud testing (52%), indicating potential issues, while YOLOv9 reported 70% accuracy. This tight alignment of results shows consistent, although modest, performance across scenarios. The findings highlight the evolution of the YOLO architecture and its impact on enhancing LPR accuracy and processing efficiency. The results provide valuable insights into selecting the most appropriate YOLO model for cloud-based LPR systems, balancing the trade-offs between real-time performance and detection precision. This research contributes to advancing the field of intelligent transportation systems by offering a detailed comparison that can guide future implementations and optimizations of LPR systems in cloud environments.

A Lightweight Model for Shine Muscat Grape Detection in Complex Environments Based on the YOLOv8 Architecture

Automated harvesting of “Sunshine Rose” grapes requires accurate detection and classification of grape clusters under challenging orchard conditions, such as occlusion and variable lighting, while ensuring that the model can be deployed on resource- and computation-constrained edge devices. This study addresses these challenges by proposing a lightweight YOLOv8-based model, incorporating DualConv and the novel C2f-GND module to enhance feature extraction and reduce computational complexity. Evaluated on the newly developed Shine-Muscat-Complex dataset of 4715 images, the proposed model achieved a 2.6% improvement in mean Average Precision (mAP) over YOLOv8n while reducing parameters by 36.8%, FLOPs by 34.1%, and inference time by 15%. Compared with the latest YOLOv11n, our model achieved a 3.3% improvement in mAP, with reductions of 26.4% in parameters, 14.3% in FLOPs, and 14.6% in inference time, demonstrating comprehensive enhancements. These results highlight the potential of our model for accurate and efficient deployment on resource-constrained edge devices, providing an algorithmic foundation for the automated harvesting of “Sunshine Rose” grapes.

Internal validation of a convolutional neural network pipeline for assessing meibomian gland structure from meibography.

Optimal meibography utilization and interpretation are hindered due to poor lid presentation, blurry images, or image artifacts and the challenges of applying clinical grading scales. These results, using the largest image dataset analyzed to date, demonstrate development of algorithms that provide standardized, real-time inference that addresses all of these limitations. This study aimed to develop and validate an algorithmic pipeline to automate and standardize meibomian gland absence assessment and interpretation. A total of 143,476 images were collected from sites across North America. Ophthalmologist and optometrist experts established ground-truth image quality and quantification (i.e., degree of gland absence). Annotated images were allocated into training, validation, and test sets. Convolutional neural networks within Google Cloud VertexAI trained three locally deployable or edge-based predictive models: image quality detection, over-flip detection, and gland absence detection. The algorithms were combined into an algorithmic pipeline onboard a LipiScan Dynamic Meibomian Imager to provide real-time clinical inference for new images. Performance metrics were generated for each algorithm in the pipeline onboard the LipiScan from naive image test sets. Individual model performance metrics included the following: weighted average precision (image quality detection: 0.81, over-flip detection: 0.88, gland absence detection: 0.84), weighted average recall (image quality detection: 0.80, over-flip detection: 0.87, gland absence detection: 0.80), weighted average F1 score (image quality detection: 0.80, over-flip detection: 0.87, gland absence detection: 0.81), overall accuracy (image quality detection: 0.80, over-flip detection: 0.87, gland absence detection: 0.80), Cohen κ (image quality detection: 0.60, over-flip detection: 0.62, and gland absence detection: 0.71), Kendall τb (image quality detection: 0.61, p<0.001, over-flip detection: 0.63, p<0.001, and gland absence detection: 0.67, p<001), and Matthews coefficient (image quality detection: 0.61, over-flip detection: 0.63, and gland absence detection: 0.62). Area under the precision-recall curve (image quality detection: 0.87 over-flip detection: 0.92, gland absence detection: 0.89) and area under the receiver operating characteristic curve (image quality detection: 0.88, over-flip detection: 0.91 gland absence detection: 0.93) were calculated across a common set of thresholds, ranging from 0 to 1. Comparison of predictions from each model to expert panel ground-truth demonstrated strong association and moderate to substantial agreement. The findings and performance metrics show that the pipeline of algorithms provides standardized, real-time inference/prediction of meibomian gland absence.

Improving LV networks hosting capacity via the use of voltage stabilizers – lab testing and simulation models for de-risking field deployment

Improving LV networks hosting capacity via the use of voltage stabilizers – lab testing and simulation models for de-risking field deployment

Majoritarianism and Monoculture (M&amp;M)

Abstract Research in decentralized computing, specifically in consensus algorithms, has focused on providing resistance to an adversary with a minority stake. This has resulted in systems that are majoritarian in the extreme, ignoring valuable lessons learned in law and politics over centuries. In this article, we first detail this phenomenon of majoritarianism and point out how minority protections in the nondigital world have been implemented. We motivate adding minority protections to collaborative systems with examples. We also show how current software deployment models exacerbate majoritarianism, highlighting the problem of monoculture in client software in particular. We conclude by giving some suggestions on how to make decentralized computing less hostile to those in the minority.

THE IMPACT OF CLOUD COMPUTING ON CONSTRUCTION PROJECT DELIVERY ABUJA NIGERIA

Cloud computing is the delivery of computing services, such as storage, processing power, and software applications, via the internet. Cloud computing offers various advantages and opportunity to construction industry, such as cost reduction, improving quality, enhancing productivity, and fostering innovation. Nigeria is lagging behind in this technological advancement due to lack of awareness, resistant to knowledge, which lead to lack of empirical evidence on the impact of cloud computing on project delivery and performance in Nigeria. The adoption and application of cloud computing in the construction industry is still low and limited in Nigeria. This study is aim at filling the gap by conducting a comprehensive study on the concept, adoption, impact, and perceive benefits of cloud computing in construction project delivery in Nigeria. To achieve this, mix-method of analysis was used, to asses, identify, determine, and evaluate the level of adoption of cloud computing in construction industry in Nigeria. From the findings, the level of adoption of cloud computing in the construction industry in Nigeria was moderate. The most adopted cloud service model was software as a service SaaS, followed by platform as a service, PaaS and infrastructure as a service IaaS. The most adopted cloud deployment model was public cloud, followed by hybrid cloud and private cloud. The factors that influenced the adoption and usage of cloud computing were organizational factors, individual factors, and environmental factors. Cloud computing has a positive and significant impact on the project delivery, in terms of time, cost, quality, and sustainability, as it enables the project managers to deliver the project faster, cheaper, better, and greener, by using the cloud services and solutions that suit their project needs and demands. The study recommend, The increase of the awareness and understanding of cloud computing among the construction project site managers, and to provide them with the necessary training and education, to enhance their knowledge and skills of cloud computing, and to increase their confidence . Keywords: Cloud computing, Empirical Evidence,Project Delivery

Performance and Efficiency Comparison of U-Net and Ghost U-Net in Road Crack Segmentation with Floating Point and Quantization Optimization

This study presents a comprehensive comparison of U-Net and Ghost U-Net for road crack segmentation, emphasizing their performance and memory efficiency across various data representation formats, including FP32, FP16, and INT8 quantization. A dataset of 12,480 images was used, with preprocessing steps such as binarization and normalization to improve segmentation accuracy. Results show that Ghost U-Net achieved a marginally higher performance, with an IoU of 0.5041 and a Dice coefficient of 0.6664, compared to U-Net’s IoU of 0.5034 and Dice coefficient of 0.6662. Ghost U-Net also demonstrated significant memory efficiency, reducing GPU usage by up to 60% in FP16 and INT8 formats. However, a sharp decline in performance was observed for Ghost U-Net in the INT8 format, where the IoU dropped to 0.2038 and the Dice coefficient to 0.3227, whereas U-Net maintained stable performance across all formats. These findings suggest that Ghost U-Net is preferable for applications prioritizing memory efficiency and inference speed, while U-Net may be better suited for tasks requiring consistent accuracy across different quantization levels. This study underscores the importance of considering both performance stability and memory efficiency when selecting models for deployment in real-world applications.

Virtual Voice Assistance for Partially Dumb People Using Audio Processing with Deep Learning

This paper presents a novel virtual voice assistance system designed specifically for partially dumb individuals, leveraging advanced audio processing and deep learning techniques. Traditional voice assistants struggle to accurately interpret partial or unclear speech, creating significant communication barriers for individuals with speech impairments. To address this, our system employs robust preprocessing methods, including noise reduction and Mel-Frequency Cepstral Coefficients (MFCC) extraction, to process incomplete speech signals. ARecurrent Neural Network (RNN) model, trained on a diverse dataset of partial speech inputs, enables accurate interpretation and contextual understanding of spoken words. The system integrates real-time processing with a text-to- speech (TTS) module, allowing it to generate meaningful responses tailored to the user's intent. Extensive testing shows the proposed system achieves 91% recognition accuracy and significantly reduces Word Error Rate (WER) compared to baseline methods. Additionally, case studies demonstrate its effectiveness in enhancing communication for users in real-world scenarios.This work contributes to inclusive technology by addressing a critical gap in voice- assisted communication. Future work includes expanding the dataset to accommodate diverse accents and optimizing the model for deployment on mobile and wearable devices, thereby extending its accessibility andimpact.

D3-YOLOv10: Improved YOLOv10-BasedLightweight Tomato Detection Algorithm Under Facility Scenario

Accurate and efficient tomato detection is one of the key techniques for intelligent automatic picking in the area of precision agriculture. However, under the facility scenario, existing detection algorithms still have challenging problems such as weak feature extraction ability for occlusion conditions and different fruit sizes, low accuracy on edge location, and heavy model parameters. To address these problems, this paper proposed D3-YOLOv10, a lightweight YOLOv10-based detection framework. Initially, a compact dynamic faster network (DyFasterNet) was developed, where multiple adaptive convolution kernels are aggregated to extract local effective features for fruit size adaption. Additionally, the deformable large kernel attention mechanism (D-LKA) was designed for the terminal phase of the neck network by adaptively adjusting the receptive field to focus on irregular tomato deformations and occlusions. Then, to further improve detection boundary accuracy and convergence, a dynamic FM-WIoU regression loss with a scaling factor was proposed. Finally, a knowledge distillation scheme using semantic frequency prompts was developed to optimize the model for lightweight deployment in practical applications. We evaluated the proposed framework using a self-made tomato dataset and designed a two-stage category balancing method based on diffusion models to address the sample class-imbalanced issue. The experimental results demonstrated that the D3-YOLOv10 model achieved an mAP0.5 of 91.8%, with a substantial reduction of 54.0% in parameters and 64.9% in FLOPs, compared to the benchmark model. Meanwhile, the detection speed of 80.1 FPS more effectively meets the demand for real-time tomato detection. This study can effectively contribute to the advancement of smart agriculture research on the detection of fruit targets.

GBiDC-PEST: A novel lightweight model for real-time multiclass tiny pest detection and mobile platform deployment

GBiDC-PEST: A novel lightweight model for real-time multiclass tiny pest detection and mobile platform deployment

A Deep Learning Approach to Detecting Advanced Persistent Threats in Cybersecurity

Advanced Persistent Threats (APTs) represent one of the most sophisticated and insidious forms of cyber-attacks, often eluding traditional detection methods due to their stealthy and prolonged nature. This paper presents a novel approach to detecting APTs by leveraging the power of deep learning. We propose a hybrid model that combines Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to capture both the spatial and temporal features inherent in APT behaviors. The model was trained and validated on a comprehensive dataset, demonstrating an accuracy of 98.5% in detecting APT activities, significantly outperforming traditional machine learning models. The proposed approach not only enhances detection accuracy but also reduces false positive rates, making it a robust solution for real-time cybersecurity applications. Our findings highlight the potential of deep learning to revolutionize APT detection, offering a scalable and adaptive framework for securing critical systems against evolving cyber threats. Future work will focus on refining the model for deployment in diverse operational environments and incorporating adaptive learning techniques to keep pace with the rapidly changing threat landscape

Employment and deployment of additional roles staff in general practice: a realist evaluation of what works for whom, how, and why.

The Additional Roles Reimbursement Scheme (ARRS) was introduced in England in 2019 to alleviate workforce pressures in general practice by funding additional staff such as clinical pharmacists, paramedics, first-contact physiotherapists, and from 1 October 2024 the scheme funds recently qualified GPs. However, the employment and deployment models of ARRS staff present ongoing complexities and challenges that require further exploration. To explore the decision-making processes behind primary care networks (PCNs) and general practice staffing choices, and how these choices influence the operationalisation of ARRS. This was a qualitative case study across four PCNs in England using a realist evaluation framework. Data collection took place between September 2022 and November 2023. Semi-structured interviews were conducted with PCN clinical directors, GPs, practice managers, and ARRS staff (n = 42). Transcripts were analysed using a realist evaluation framework to identify the context-mechanism-outcome configurations. Direct employment models fostered staff development and retention, contingent on established trust among practices. Subcontracting was favoured to mitigate employment risks but could lead to unintended consequences such as conflicting accountabilities and less integration with existing GP practice staff. The optimal deployment model involved rotations across a limited number of GP practices, ideally two, with one serving as a base, ensuring consistency in training and management. This study provides novel insights into the complexities of different employment and deployment models of ARRS staff. These findings will be invaluable for creating a sustainable GP practice workforce and informing future workforce strategies as the scheme expands to include recently qualified GPs.

Transit search algorithm based on oscillation exploitation factor and Roche limit for wireless sensor network deployment optimization

To optimize the deployment of nodes in Wireless Sensor Networks (WSN) and effectively control network node energy consumption, thereby improving the quality of perception services, a Transit search algorithm based on oscillation exploitation factor and Roche limit is proposed. The Roche limit-inspired approach enhances the stellar phase of the algorithm, accelerating the convergence rate in the mid-to-late stages of iteration while ensuring adequate exploration of the solution space. Subsequently, five weakening oscillation development factors are introduced to refine the algorithm’s exploitation phase and improve its fine-tuning accuracy. To validate the effectiveness of these strategies, various approaches are applied to optimize the coverage, waste rate and energy consumption in two models of WSN deployment, with connectivity recorded. The comparison reveals the optimal improved algorithm, SEROTS, which enhances coverage by 1.34% in the obstacle-free model compared to the original TS algorithm, with waste and energy consumption rates reduced by 2.05% and 0.00016%, respectively. In the obstacle model, coverage increases by 1.49%, while waste and energy consumption rates decrease by 6.96% and 0.0004%, respectively. To demonstrate the efficiency of the improved algorithm in optimizing WSN deployment, SEROTS is compared with four optimization algorithms: Egret Swarm Optimization Algorithm (ESOA), Honey Badger Algorithm (HBA), Sparrow Search Algorithm (SSA) and Differential Evolution (DE). Two models are selected, integrating the three objectives into a single objective function. Simulation results indicate that SEROTS performs best in both models, with an improvement of 0.53% and 0.79% over the second-best algorithm, respectively. Furthermore, the proposed strategies are compared with simulation results from five other studies, achieving higher coverage rates by 1.57%, 3.33%, 0.87%, 3.81% and 0.21%, respectively. Finally, experiments discuss the application in large-scale scenarios, verifying the feasibility and efficiency of the SEROTS algorithm in WSN deployment optimization.

Advancement of High Temperature Electrochemistry for Decarbonization Applications at Idaho National Laboratory

Decarbonization is vital for addressing climate change, protecting public health, enhancing energy security, fostering economic prosperity, and advancing sustainable development. High temperature electrochemistry is playing an increasingly important role in decarbonization as we reduce energy use and address uncertain supply chains driven by industrial electrification and cheap heat resources. As one of DOE’s premier multi-program energy research laboratories, Idaho National Laboratory’s (INL) mission is to ensure the nation’s energy security with safe, competitive and sustainable energy systems and unique national and homeland security capabilities. INL’s initiative, integrated energy systems (IES) provides carbon-free nuclear and renewable energy for downstream use in manufacturing and transportation. The manufacturing component of IES focuses on integrating clean, carbon-free electrons with waste/renewable heat and includes both manufacturing of electrochemical devices and implementing them in energy-efficient processes. In this talk, we will provide representative examples for the pillar of energy to molecules and materials (E2M2) by developing, validating and demonstrating the high temperature electrochemical technologies towards chemical/fuel/hydrogen production as well as CO2 capture and utilization. INL research portfolio represents an outstanding example that links basic research to applied technology for DOE national laboratory complex and uniquely provide technology deployment models through synergistic collaboration among national labs, universities and industry.

A robust Pareto model for electric vehicle charging station deployment in urban areas considering psychology effects of drivers

A successful transition from gas-powered to electric vehicles (EVs) depends on identifying the most convenient locations for electric vehicle charging stations (EVCS), particularly in urban areas. While EVCS location problems have been addressed in the literature, this study considers the ambiguity of EV drivers' range anxiety and charging demand to explore the EVCS deployment in continuous and discrete solution spaces, representing roads and parking facilities in the real-world. Additionally, our paper highlights the novelty of including the negative psychology effects experienced by both electric vehicle (EV) and fuel vehicle (FV) drivers due to the ICEing problem, where fuel vehicles (FVs) block EVCS access. This paper proposes a comprehensive framework that includes a spatio-temporal Gaussian process model for predicting charging demand, a multi-objective EVCS location model for an EVCS deployment, and a Scenario-based Multi-Objective min-max Robust Pareto (SMORP) model with ambiguous charging demand and drivers' range anxiety for a robust Pareto EVCS deployment. The proposed algorithms identify the optimal and robust Pareto fronts for EVCS deployments. We validate the models using a case study of an urban area. The resulting EVCS deployment enables the selection of optimal EVCS locations among discrete parking facilities and identifies continuous coordinates for curb parking space for EV charging.

Evaluasi Trade-off Akurasi dan Kecepatan YOLOv5 dalam Deteksi Kebakaran pada Edge Devices

Real-time object detection using the YOLO (You Only Look Once) algorithm has shown promising performance in various computer vision applications. However, its application on devices with limited resources is still a challenge due to its high computational requirements. This study aims to optimize the YOLOv5 model for fire and smoke detection on Orange Pi Zero 3 devices using quantization techniques. Using a dataset of 2247 fire and smoke images, this study applies static quantization techniques to improve model efficiency. The methodology includes training of standard YOLOv5 models, conversion to ONNX format, and application of static quantization. Results show a significant improvement in computational efficiency, with a 42.2% reduction in model size and a 65.21% increase in inference speed. Despite a decrease in the mAP value by 25.6%, the optimized model was still able to perform object detection at a significantly higher speed. In conclusion, the quantization technique is effective in optimizing the YOLOv5 model for deployment on edge computing devices, despite the trade-off between speed and accuracy.