Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Energy-Efficient Sector-Based Routing in 3D Wireless Sensor Networks Using Midpoint-Aware Cluster Head Selection
No full textEnergy efficiency is a key concern in three-dimensional Wireless Sensor Networks (3D WSNs), where irregular node distribution can shorten network lifetime. This study introduces a novel sector-based clustering protocol that partitions the sensing field azimuthally and selects Cluster Heads (CHs) based on residual energy and distance to each sector’s geometric midpoint. The core innovation lies in an adaptive threshold formula that incorporates both energy and spatial distance, enabling fairer CH rotation and better intra-cluster communication. This approach ensures more balanced energy depletion and improved load distribution. Simulations in a 3D environment showed that the proposed protocol outperformed LEACH-Classic, LEACH-GA, and LEACH-KMeans in FND, HND, LND, residual energy, and throughput. Notably, it improves FND by up to 59.2% compared to LEACH and increases throughput by 37.8%, confirming the benefits of azimuth-based clustering and midpoint-aware CH selection
Leveraging Machine Learning to Predict Academic Specialization Pathways in Higher Education
No full textThis study developed a machine learning-based model to predict academic concentration selection among information systems students at Universitas Multimedia Nusantara (UMN). A survey of 125 students from the 2024 cohort revealed that 90% experienced difficulties in choosing a specialization, primarily due to limited information on course relevance, unclear academic pathways, and career uncertainty. While the survey provides a contextual background, the predictive model was trained using historical academic performance data from the 2021–2023 cohorts. The three classification algorithms, Decision Tree, Random Forest, and Extreme Gradient Boosting (XGBoost) were implemented following the CRISP-ML methodology. To address class imbalance in the dataset, the Synthetic Minority Over-sampling Technique (SMOTE) was applied, followed by hyperparameter tuning and feature selection. The Random Forest model demonstrated superior performance, achieving an accuracy of 78.08% on the 2021–2022 cohort data, outperforming Decision Tree and XGBoost across all experimental settings. This result highlights Random Forest's robustness in this context, particularly after the integration of SMOTE and optimization procedures. The main contribution of this study lies in the application of machine learning for academic pathway prediction in an Indonesian higher education setting, providing a data-driven decision support tool to assist students in making informed and personalized specialization choices
A Hybrid Round-Robin Scheduler for GPU Batch Rendering in Constrained Cloud Environments
No full textCreating high-quality 2D and 3D assets is essential for digital content, but inefficient scheduling and inaccurate time estimates often hamper the rendering process. Traditional methods, which assume rendering time is directly proportional to frame count, fail to account for variations in scene complexity, resulting in severe estimation errors averaging 97.0% across all tasks. We propose a Hybrid Round-Robin Scheduler (HRRS) that intelligently manages batch rendering tasks through complexity-aware classification. Our method first categorizes tasks by complexity (Low, Medium, High) and routes them to appropriate queues with tiered quantum allocations. It then employs non-linear time estimation models and dynamically adjusts processing priorities based on real-time performance metrics. We evaluated our scheduler against standard algorithms—First-Come-First-Served (FCFS), Shortest Job First (SJF), and Round Robin (RR)—using 21 diverse rendering tasks with frame counts ranging from 10 to 420 frames. The results demonstrate that our approach reduces average waiting time by 45.9% (from 29.63s to 16.02s) and cuts bottleneck-induced delays by 78% (from 41s to 9s), while maintaining optimal CPU utilization at 85% and limiting context switches to only nine occurrences. A key finding reveals that complexity, rather than frame count, is the primary driver of processing time; high-complexity tasks required significantly longer processing (averaging 238.27 seconds) compared to medium-complexity tasks (averaging 34.52 seconds), representing a 6.9-fold performance differential. Our hybrid framework effectively overcomes the primary limitations of existing algorithms: it prevents bottlenecks from large tasks (FCFS), avoids the parallelism issues of SJF, and minimizes the performance overhead from frequent switching in Round Robin. This work provides a robust foundation for intelligent resource allocation in cloud rendering environments where task demands are variable and difficult to predict, establishing that effective scheduling requires complexity-aware algorithms rather than universal approaches
Real-Time Hand Gesture-Based Virtual Mouse System Using ESP32-CAM and OpenCV
No full textThis research develops a virtual mouse control system that uses real-time hand gesture recognition implemented on an ESP32-CAM–based Internet of Things (IoT) platform. By leveraging OpenCV for image processing, the system translates hand gestures into corresponding mouse actions, including cursor movement, clicking, and scrolling. The study evaluates system performance under different lighting conditions and Wi-Fi speeds. Results show that higher Wi-Fi speeds significantly reduce latency, enabling smoother real time gesture recognition and high definition video output, while lower speeds lead to noticeable delays and reduced accuracy. The system successfully enables remote cursor control through camera captured hand gestures, supporting functions such as left click, right click, scrolling, and dragging. In latency tests performed with an internet speed of approximately 60 Mbps, the system achieved an average delay of about 50 milliseconds. Under optimal lighting conditions with minimal background interference, it accurately tracked hand movements and recognized gestures such as pointing, clicking, dragging, and scrolling in real time, achieving an accuracy rate of 95%. Despite its lower resolution compared to conventional webcams, the ESP32-CAM proves to be an effective solution for virtual mouse control, particularly in scenarios where high-resolution imaging is unnecessary. Its IoT capabilities support remote operation, allowing users to control the virtual mouse from a distance as long as both the ESP32-CAM and the computer remain connected. Overall, the findings highlight the ESP32-CAM based IoT platform as a viable alternative for gesture based interaction in real applications, although further enhancements are needed to improve performance in challenging environments
Enhanced RegNetY-400MF for Fruit Fly Species Classification: Fine-Tuning Strategies and Data Balancing for Improved Accuracy
No full textFruit fly infestations pose a significant threat to agricultural productivity, especially in chili plantations, which can cause substantial yield losses. Accurate and rapid species classification is crucial for implementing targeted pest control strategies. This study developed a computationally efficient fruit fly species classification model using a deep learning approach that focused on improving accuracy with fine tuning and class balancing strategies. The dataset consists of 1049 images across 4 fruit fly species, captured in a natural plantation environment and available at www.inaturalist.org. The model evaluated several lightweight Convolutional Neural Network (CNN) architectures, including MobileNetV3-Small, RegNetY-400MF, and SqueezeNet among others, with RegNetY-400MF emerging as the best performing model, achieving a validation accuracy of 96.10% and a macro F1 score of 95.70%. The models tested in this study included several lightweight Convolutional Neural Network (CNN) architectures, including MobileNetV3-Small, RegNetY-400MF, and SqueezeNet, among others. RegNetY-400MF proved to be the best performing model, achieving a validation accuracy of 96.10% and a macro F1 score of 95.70%. Compared to other state-of-the-art models, RegNetY-400MF demonstrated higher accuracy while maintaining a lower number of parameters (8.3 million) and reduced computational complexity (0.41 GFLOPs). This makes the model highly suitable for real-time applications in resource-constrained agricultural environments. The model offers a practical solution for fruit fly species detection, enabling early and accurate identification of pest infestations in chili plantations, thereby reducing the risk of crop failure. By providing an efficient and scalable pest control tool, the model supports precision pest management, improves yield stability, and contributes to sustainable agriculture
Multi-Class Semantic Segmentation of Oil Palm Areas Using a VGG-19 U-Net Improvement
No full textUAV imagery-based semantic segmentation is crucial for mapping tropical agricultural areas such as oil palm plantations. The main challenges are overlapping vegetation objects, unclear boundaries, and spectral similarities between classes, which reduce the accuracy of conventional models. This study proposes a modified U-Net architecture with a VGG-19 backbone, achieved through hyperparameter tuning (M7) and the integration of residual blocks (M8), to enhance multi-class segmentation performance. Experiments were conducted on aerial imagery with two resolutions (512×512 and 256×256) using four-class and three-class scenarios. The results show that M7 and M8 consistently outperform the baseline model (M2) in terms of accuracy, precision, recall, and average Intersection over Union (IoU). In the 512x512 four-class scenario, M8 achieved the highest accuracy (87.40%), precision (88.32%), recall (86.32%), and MIoU (0.132). M7 reached similar accuracy (>86%) but trained significantly faster than the baseline. In the 256x256 scenario, M8 maintained strong performance with 86.44% accuracy and 0.302 MIoU. For the three-class experiment, M8 reached a top MIoU of 0.178. Accuracy, precision, and recall were all above 87%, showing improved recognition of minority classes such as waterways. Confusion matrix analysis confirmed that M8 provided more balanced class predictions. It also reduced false negatives for oil palm vegetation. M7 showed slight fluctuations, suggesting possible overfitting. These findings support M8 as a robust solution for UAV-based oil palm mapping and large-scale monitoring
A Hybrid Framework Combining U-Net, Ant Colony Optimization, and CNN for Rice Leaf Disease Classification under Class Imbalance
No full textAccurate classification of rice plant diseases is essential for early intervention and precision agriculture. However, real-world datasets often suffer from complex backgrounds, high-dimensional features, and severe class imbalances, which compromise classification performance. This study proposes an integrated framework combining image segmentation using U-Net, feature selection via Ant Colony Optimization (ACO), hybrid sampling to handle class imbalance, and final classification using a Convolutional Neural Network (CNN). Segmentation isolates disease-affected areas, ACO optimizes feature subsets, and hybrid sampling balances class distribution using undersampling and SMOTE. The proposed method was tested on four rice leaf disease datasets—Brown Spot, Leaf Blast, Leaf Blight, and Leaf Scald—exhibiting significant class imbalance. Experimental results show that the proposed approach outperforms baseline models (SegNet, PspNet, and E-Net) across multiple metrics: Accuracy, IoU, Precision, and Recall. This indicates the framework’s robustness and potential for real-world deployment in precision agriculture. Future work will focus on model compression and real-time implementation in IoT systems
EBAQ: An Entropy-Based Bit Allocation Framework for Lightweight Autoencoder Models
No full textAutoencoder-based models have shown strong potential for anomaly detection in complex time-series data; however, they often assume equal importance across latent dimensions, resulting in inefficiencies and reduced precision. This study addresses this limitation by introducing the Entropy-Based Bit Allocation Quantizer (EBAQ), a novel quantization framework that adaptively allocates bits to each latent dimension based on its entropy, preserving more precision where information content is highest. The primary objective is to enhance representational efficiency and anomaly detection performance without increasing model complexity or computational cost. EBAQ is implemented as a plug-and-play module within a standard autoencoder architecture, requiring no retraining or architectural modification. The method was evaluated using a publicly available ECG dataset, where reconstruction-based anomaly detection was employed to assess its performance. Results show that EBAQ outperforms the standard autoencoder baseline, achieving higher accuracy (94.9%), precision (99.4%), and recall (91.4%), while also demonstrating more apparent separation between normal and anomalous data in latent space visualizations. These findings confirm that entropy-aware quantization improves both fidelity and interpretability in unsupervised anomaly detection. Overall, this work presents a theoretically grounded and practically efficient solution that bridges information theory and deep learning, offering a human-centered approach to developing more intelligent and efficient AI systems for real-world applications
Benchmarking Transformer Architectures for Chest X-ray Classification
No full textLung diseases remain a major global health concern, necessitating accurate and timely diagnosis. Chest X-ray (CXR) imaging is widely used but challenging to interpret due to overlapping radiographic features and subjective variability among radiologists. Deep learning approaches, particularly Convolutional Neural Networks (CNNs), have shown promise but are limited in capturing global spatial dependencies. Vision Transformers (ViTs) overcome this limitation through self-attention, making them increasingly attractive for medical image analysis. This study systematically evaluates 13 Transformer-based architectures across three CXR datasets with distinct tasks: Pneumonia (3-class: Normal, Bacterial, Viral), COVID-QU-Ex (3-class: Normal, Non-COVID Pneumonia, COVID-19), and Tuberculosis (2-class: Normal, Tuberculosis). All models were trained under a unified setup with consistent preprocessing, augmentation, and evaluation protocols. To improve robustness, a soft voting ensemble of the top five models was also implemented. Results demonstrate that Transformer-based models provide highly competitive performance. On the Pneumonia dataset, the ensemble achieved an accuracy of 0.8743 and F1-score of 0.8615, surpassing several single models such as DeiT-Base (F1 = 0.8725). On COVID-QU-Ex, the ensemble soft voting obtained 0.9593 accuracy and 0.9582 F1-score, effectively balancing precision and recall. On Tuberculosis, ViT-B/16 and MobileViT-S achieved perfect performance (F1 = 1.0), likely influenced by dataset imbalance. These findings highlight the clinical potential of Transformer-based models, particularly when combined through ensembles, for robust and accurate CXR classification
Web-Based Deepfake Detection Using VERITAS: Integrating Vision-Based Excitation with Transformer-Driven Intelligence
No full textThis study proposes a web-based deepfake detection system that integrates Vision-Based Excitation technology and Transformer-based intelligence, called VERITAS (Vision-based Excitation and Robust Intelligence for Transformer-Assisted Deepfake Detection). The system is designed to automatically detect manipulated images and videos by leveraging the Vision Transformer (ViT) model architecture, equipped with the Grad-CAM mechanism for interpretability of detection results. The study conducted a series of tests to measure the system's performance in various scenarios and ensure its reliability in dealing with various types of input. Load testing results showed that up to 30 simultaneous users, the system can operate with good responsiveness (average response time of 130 ms) without experiencing errors. However, when the number of users reaches 40 or more, the system performance drops drastically with a very high error rate, reflecting limitations in handling server load. Real-world testing showed the system can detect deepfakes with an accuracy of 73.61%, with results varying depending on the quality of the tested images. Furthermore, unit functional testing and coverage analysis demonstrated an excellent test pass rate (85%), with all major functions running smoothly and error handling needed to be fixed in some code sections. Overall, the VERITAS system demonstrates strong potential for web-based deepfake detection, with high reliability under low load and adequate performance in functional testing. However, further optimization is needed to handle higher user loads