Articles

Access the latest knowledge in applied science, electrical engineering, computer science and information technology, education, and health.

Advanced Search
Filter Icon

Filters article

Years

FAQ Arrow
Single Year
Range Year
0
0

Source Title

FAQ Arrow

Authors

FAQ Arrow

31,377 Article Results

A deep learning-based system for coral reef image segmentation using YOLOv8 with EfficientNet-B0 in Indonesian waters

10.11591/ijeecs.v43.i2.pp628-639
Raden Sutiadi , Sparisoma Viridi , Giyanto Giyanto , Andarta F. Khoir , Rizkie S. Utama , Elsa D. Aulia , Tri A. Hadi , Ludi Parwadani Aji
Manual coral point count with excel extension (CPCe) analysis requires approximately 4–6 hours to process one 50-image station, limiting the scale of coral reef monitoring. This study presents an artificial intelligence-based workflow using YOLOv8 with an EfficientNet-B0 backbone to automate benthic cover estimation. A total of 8,147 underwater images from 151 transects across 39 Indonesian coral reef stations were annotated into eleven benthic categories for model training, while evaluation was conducted using the 2021 Derawan Islands dataset. During validation, YOLOv8 achieved an average mAP@0.5 of 0.58, recall of 0.79, and an average absolute percentage cover error of 9.8% compared with CPCe. The model processed each image in 3.13 seconds, equivalent to 156.47 seconds per 50-image station, representing a 92–138× speedup over manual CPCe analysis. These results show that the proposed workflow can support scalable and near-real-time coral reef monitoring across Indonesia.
Volume: 43
Issue: 2
Page: 628-639
Publish at: 2026-08-01

WVisionBERT-VL: a multimodal model architecture for toxicity classification on social media platforms using large language models

10.11591/ijai.v15.i4.pp3365-3375
Witta Listiya Ningrum , Achmad Benny Mutiara , Diana Ikasari
The increasing prevalence of toxic content on social media, conveyed through text, images, and videos, poses significant challenges for automated content moderation systems. Although prior studies have reported promising results in unimodal and bimodal settings, they often fail to capture implicit and contextual toxicity emerging from interactions across multiple modalities, particularly in non-English environments. This paper proposes WVisionBERT-VL, an end-to-end multimodal framework for toxicity detection that integrates text, image, and video modalities within a unified architecture. The proposed model incorporates modality-specific encoders, bidirectional multi-head cross attention (BMHCA) for cross-modal synchronization, and an adaptive fusion gate to dynamically balance modality contributions. A balanced multimodal dataset is constructed from social media platforms, including X, Instagram, and TikTok, and refined using a model-based labeling strategy with limited human-in-the-loop validation. Experimental results on a custom Indonesian dataset demonstrate strong in-domain performance, achieving an accuracy of 94.12%, a macro-F1 of 0.9407, and a receiver operating characteristic - area under the curve (ROC-AUC) of 0.9721, with robustness further validated through five-fold cross-validation. Cross-dataset evaluation highlights challenges related to domain shift, underscoring the need for future research on robust and domain-adaptive multimodal toxicity detection.
Volume: 15
Issue: 4
Page: 3365-3375
Publish at: 2026-08-01

A machine learning framework for skin cancer classification using texture descriptors

10.11591/ijai.v15.i4.pp3556-3567
Shikha Malik , Vaibhav V. Dixit
Skin cancer remains a critical health and economic concern worldwide. Timely and accurate diagnosis is crucial for improving mortality rate of patients. Although automated machine learning (ML) models assist doctors in clinical assessment of skin cancer from dermoscopic images, their performance often suffers from extreme class imbalance, as benign image samples greatly exceed malignant ones. This leads to false detection and delayed diagnosis of skin cancer. To overcome this issue, the study proposes an efficient and lightweight geometric transformation (GT)–augmented support vector machine (SVM) framework for early diagnosis of skin cancer. It effectively addresses the class imbalance issues present in the datasets and improves the detection of positive cases. The novel pipeline integrates preprocessing, morphology preserving GT, rotation-invariant texture feature extraction, feature validation, and feature scaling for performing binary classification using an optimized SVM framework. This framework provides balanced and accurate lesion classification even if image samples are insufficient. Experimental results have successfully achieved a true positive rate (TPR) of 93% on PH2 and 81.1% on International Skin Imaging Collaboration 2016 (ISIC-2016) dataset, which is better than conventional ML models. These findings prove that proposed GT-SVM framework is a lightweight, interpretable, and computationally efficient approach for early skin cancer diagnosis. Future developments shall explore multi-class lesion classification, validation across diverse clinical datasets, and hybrid feature fusion.
Volume: 15
Issue: 4
Page: 3556-3567
Publish at: 2026-08-01

YOLO-based deep learning for tooth detection, segmentation, and numbering in panoramic radiographs

10.11591/ijai.v15.i4.pp3591-3602
Sri Oktamuliani , Luqyana Mahdiyah , Haritsul Haq , Nesa Perdana Putri , Wulandani Liza Putri
Panoramic dental radiographs are essential for diagnosis and treatment planning, but are difficult to interpret due to overlapping anatomical structures and limited image quality. Automating tooth detection, segmentation, and numbering can enhance clinical efficiency and reduce observer variability. This study developed a you only look once version 8 (YOLOv8)-based deep learning model for automatic detection and segmentation of teeth in panoramic radiographs. A dataset of 302 images containing 9,009 annotated teeth was divided into 70% training, 20% validation, and 10% testing sets. Teeth were labeled using the Federation Dentaire Internationale (FDI) two-digit numbering system with bounding boxes and segmentation masks. Model performance was assessed using precision, recall, F1-score, and mean average precision (mAP) at intersection over union (IoU) thresholds of 0.5 and 0.5–0.95. The model achieved bounding box precision, recall, and F1-score of 0.9075, 0.9352, and 0.9212, respectively, and segmentation scores of 0.9078, 0.9344, and 0.9209. Bounding box mAP@0.5 reached 0.9510 and mask mAP@0.5 was 0.9504, while stricter thresholds lowered performance to 0.7763 and 0.6896. Class-wise analysis showed high accuracy overall but reduced performance on posterior teeth due to anatomical overlap and peripheral image quality. YOLOv8 enables fast, accurate, and robust tooth detection and segmentation, supporting real-time computer-aided diagnosis in orthodontics, prosthodontics, and forensic dentistry.
Volume: 15
Issue: 4
Page: 3591-3602
Publish at: 2026-08-01

Forecasting income inequality in Vietnam’s regions using machine learning

10.11591/ijai.v15.i4.pp3464-3475
T. Thai-Phuong , L. Nguyen-Son
This study applies artificial neural networks (ANN) and support vector regression (SVR) to forecast the Gini coefficient across Vietnam’s seven socio-economic regions using limited data from 2014–2022 (N = 63). Despite the small dataset, robust techniques including repeated cross-validation, temporal backtesting, regularization, and bootstrapping were employed to mitigate overfitting. The results show a modest continued decline in income inequality through 2025, with faster improvements in urban regions (Red River Delta and Southeast) than in rural and highland areas. ANN outperformed SVR with an average cross-validation coefficient of determination (R²) of 0.93. The novelty lies in its granular regional forecasting and region-specific policy recommendations beyond national-level analyses. Key drivers—per capita income and simple housing—support targeted interventions such as housing credits in the Mekong River Delta and agricultural investments in the Central Highlands. The findings provide an evidence-based foundation for regionally tailored strategies to reduce income disparities in Vietnam.
Volume: 15
Issue: 4
Page: 3464-3475
Publish at: 2026-08-01

Android-based tomato leaf disease classification using a lightweight MobileNetV2 convolutional neural network

10.11591/ijai.v15.i4.pp3318-3325
Andi Riansyah , Irfan Eka Mahdy , Mochamad Abdul Basir , Noorminshah A. Iahad
Early screening of tomato leaf diseases is important because foliar symptoms can reduce plant vigor and delay appropriate crop management. This study develops an Android-based tomato leaf disease classification system using MobileNetV2 as a lightweight convolutional neural network (CNN) architecture. The contribution of this work is the integration of model training, independent testing, and on-device Android deployment that supports camera and gallery inputs without relying on server-side computation. The dataset consisted of 1,200 balanced tomato leaf images from five classes: bacterial spot, late blight, target spot, tomato yellow leaf curl virus, and healthy leaf. Images were resized, normalized, augmented for training, and divided into training, validation, and independent testing subsets. The model obtained 94.12% training accuracy, 93.00% validation accuracy, and 89.00% independent test accuracy. The confusion matrix showed that tomato yellow leaf curl virus was classified without error, whereas bacterial spot, late blight, target spot, and healthy leaves produced several misclassifications because of similar lesion and discoloration patterns. The results show that MobileNetV2 is suitable for lightweight mobile disease screening, although larger field datasets, cross-validation, model comparison, and explainability analysis are still needed for broader deployment.
Volume: 15
Issue: 4
Page: 3318-3325
Publish at: 2026-08-01

Hybrid deep learning model for the task scheduling in cloud computing

10.11591/ijai.v15.i4.pp3683-3691
Kavita Rani , Om Prakash Sangwan , Ritu Garg
Task scheduling plays a crucial role in optimizing performance, reducing costs, and enhancing system reliability by efficiently allocating resources to workloads. Traditional task scheduling methods lack the ability to efficiently manage workloads and resource distribution, leading to potential inefficiencies in performance and energy consumption. To address these limitations, advanced techniques leveraging deep learning and reinforcement learning are explored. This study proposes a deep learning-based model for task scheduling in cloud computing. The model employs a convolutional neural network (CNN) to predict the optimal machines for task allocation. Additionally, Q-learning is integrated with CNN to facilitate load shifting between machines, ensuring efficient utilization of resources. The dataset used in this work consists of task attributes, such as execution time, resource requirements, which were loaded from a CSV file. Comparative analysis with existing models shows that the proposed approach achieves approximately 94% accuracy and consumes less energy than other models, demonstrating its effectiveness in cloud task scheduling.
Volume: 15
Issue: 4
Page: 3683-3691
Publish at: 2026-08-01

Semantic-aligned multimodal human activity recognition using visual and audio data

10.11591/ijece.v16i4.pp2087-2095
Yeeun Park , Junhoo Byun , Siwoo Byun
Human activity recognition (HAR) requires robust performance under heterogeneous sensing conditions for practical deployment. However, single-modality approaches are limited in capturing the rich contextual information inherent in complex human behaviors. This paper presents a semantic-aligned multimodal HAR framework that integrates visual and audio information without assuming instance-level synchronization. To address dataset heterogeneity, samples from the HMDB51 video dataset and the ESC-50 audio dataset are aligned by mapping fine-grained classes into a shared high-level activity label space. For each modality, ResNet-18-based models are trained independently using frame-based visual inputs and 64-bin Mel-spectrogram-based audio representations. During inference, the output logits of the two models are combined through score-level weighted linear fusion. Experimental results show that the proposed multimodal approach consistently outperforms unimodal baselines in terms of accuracy and Macro-F1 score, with particularly notable improvements in activity groups where environmental context plays a significant role. These findings indicate that semantic-aligned score-level fusion can improve recognition robustness even under mismatched dataset conditions.
Volume: 16
Issue: 4
Page: 2087-2095
Publish at: 2026-08-01

A convolutional neural network -based driver monitoring system for drowsiness and distraction detection

10.11591/ijece.v16i4.pp2220-2229
Sara Benkouider , Nasreddine Lagraa
Road accidents caused by driver drowsiness and distraction are a major global concern, as fatigue and inattention significantly slow reaction time and increase accident risk. To address this issue, this paper proposes a vision-based driver monitoring system using facial analysis from an on-board camera. The system detects the face and extracts key regions of interest, including the eyes, mouth, and head, which are analyzed independently using convolutional neural networks. Temporal information is captured by aggregating the convolutional neural network (CNN) outputs over a fixed time window. Drowsiness is estimated by fusing eye and mouth features with a multilayer perceptron, while distraction is detected based on head movements. An important advantage of the proposed approach is its robustness to partial input loss, allowing the system to remain functional even when some facial regions are missing or occluded, such as when wearing sunglasses or face masks. Experimental results show high detection accuracy, reaching 97.3% for drowsiness and 98% for distraction under ideal conditions, with only limited performance degradation in challenging scenarios. These results confirm the suitability of the proposed system for real-time driver monitoring applications.
Volume: 16
Issue: 4
Page: 2220-2229
Publish at: 2026-08-01

Enhancement of YOLOv8 for object detection in adverse weather conditions using generative adversarial network

10.11591/ijece.v16i4.pp2230-2246
Talifhani Calvin Tshipota , Chunling Tu , Mukatshung Claude Nawej , Sempe Thom Leholo
Detecting objects in bad weather like rain, fog, snow, or low light is still difficult because visibility drops, noise increases, and contrast gets worse, all of which hurt detection accuracy. Most current methods either improve detector designs or use image preprocessing on their own. They usually focus on just one type of weather and do not use a common way to evaluate results. This paper introduces a YOLOv8 framework improved with a generative adversarial network (GAN) for image enhancement before detection. Instead of just making images look better, the GAN is trained to help the object detector work better in tough conditions, so it can extract features more effectively when images are degraded. The model was tested on datasets with different weather conditions using standard metrics like Precision, Recall, F1-score, and mean average precision (mAP). Results show that this method consistently improves performance, with up to a 6.5% increase in mAP@0.5 over YOLOv8-STE and 9.2% over IA-YOLO, especially in foggy and low-light situations. These results show that adding GAN-based preprocessing to YOLOv8 makes detection more reliable and still keeps the process fast. This framework offers a practical and scalable solution for real-world uses like self-driving cars, smart transportation, and surveillance.
Volume: 16
Issue: 4
Page: 2230-2246
Publish at: 2026-08-01

Teaching analysis of sub-synchronous resonance of thermal power plants using virtual laboratory

10.11591/ijece.v16i4.pp1677-1687
Sugiarto Kadiman , Ratna Kartikasari
The development of a MATLAB/Simulink-based virtual laboratory for studying sub-synchronous resonance (SSR) offers a robust educational platform for analyzing complex power system interactions. Based on the IEEE first benchmark model, this virtual environment provides a safe, efficient, and comprehensive tool for engineering students to study the dangerous interactions between series-compensated lines and turbo-generator shaft systems. The simulation features a 920 MVA, 60 Hz turbo-generator connected to an infinite bus through a series-compensated line. Students can analyze torsional interaction, which result from energy exchange between the electrical network and the mechanical shaft. Higher degrees of series compensation increase the risk of SSR, as the electrical resonant frequency matches the complement of one of the mechanical shaft torsional modes.
Volume: 16
Issue: 4
Page: 1677-1687
Publish at: 2026-08-01

From climate time series to planting windows in chili (Capsicum frutescens): a SARIMA–SVM–XGBoost framework with balanced-accuracy thresholding

10.11591/ijece.v16i4.pp2210-2219
Efrans Christian , Nova Noor Kamala Sari , Ressa Priskila , Septian Geges
This study proposes a spatio-temporal decision-support framework that integrates Seasonal Autoregressive Integrated Moving Average (SARIMA), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) to derive adaptive planting windows for chili (Capsicum frutescens) at the sub-district level. The framework addresses key challenges in climate-sensitive agriculture, including spatial data leakage and class imbalance, by employing Leave-One-Group-Out (LOGO) cross-validation and Balanced Accuracy–based threshold optimization. The proposed system transforms heterogeneous environmental data into actionable recommendations by combining climate forecasting, land suitability assessment, and yield prediction within a unified pipeline. Experimental results indicate that the framework effectively captures seasonal climate dynamics and produces consistent planting recommendations aligned with agronomic conditions, enabling multiple planting cycles per year. The primary contribution of this work lies in a transparent and generalizable integration of statistical and machine learning models into a practical decision-support framework. The proposed approach bridges predictive modeling and real-world agricultural decision-making and can be extended to other crops and regions for climate-adaptive agricultural planning.
Volume: 16
Issue: 4
Page: 2210-2219
Publish at: 2026-08-01

Physics-based modeling of cobalt-doped nickel-zinc on-chip ferrite inductors

10.11591/ijece.v16i4.pp1805-1816
Bambang Mulyo Raharjo , Dicky Rezky Munazat , Sudirman Rohadi
The miniaturization of integrated voltage regulators (IVRs) for multi-core processors is fundamentally bottlenecked by the high-frequency magnetic losses of conventional inductor cores. This study presents a rigorous computational framework to optimize Cobalt-doped Nickel-Zinc ferrite (Ni_(1-x) Zn_0.4 Co_x Fe_2 O_4) for 10 MHz on-chip power delivery. Utilizing Landau-Lifshitz-Gilbert (LLG) relaxation dynamics and Maxwell-Wagner interfacial polarization, the complex electromagnetic dispersion was modeled and quantitatively validated against recent empirical literature. Furthermore, high-temperature power loss density was bounded using the trust region reflective (TRF) numerical curve fitting algorithm to validate an anisotropy-compensated "thermal valley" at 80 °C. A multi-objective sensitivity analysis identified a high-efficiency Cobalt "sweet spot" at a concentration of x=0.04. This specific formulation optimally stiffens domain walls, safely shifting the resonance frequency to 40 MHz and maximizing the quality factor (Q) at the 10 MHz operational target. When applied to a simulated 3.5 V to 1.0 V DC-DC buck converter, the optimized x=0.04 core demonstrated its adequacy for on-chip applications by maintaining >90% power efficiency under a rigorous 2.0 A load. These predictive results mathematically prove that precision Cobalt doping is a highly viable strategy for suppressing parasitic losses in next-generation 3D-IC power delivery networks.
Volume: 16
Issue: 4
Page: 1805-1816
Publish at: 2026-08-01

Performance analysis of 5G NR network planning at 2300 MHz in Indonesia: A comparative study of LoS and NLoS scenarios

10.11591/ijece.v16i4.pp1944-1954
Putri Rahmawati , Lia Hafiza , Muhammad Adam Nugraha , Syifa Maliah Rachmawati
This study presents a systematic comparative analysis of 5G new radio (NR) network planning at 2300 MHz for Bandung City, Indonesia, evaluating line-of-sight (LoS) and non-line-of-sight (NLoS) propagation scenarios to determine infrastructure requirements and performance characteristics for urban deployment. Employing the 3GPP TR 38.901 Urban Macro propagation model, link-budget analysis, and Atoll-based simulation for a 167.31 km² urban area, the study evaluates coverage performance under projected 2025 deployment conditions. The results reveal significant differences between propagation scenarios. LoS conditions require 45 gNodeBs for uplink coverage, achieving a synchronization signal reference signal received power (SS-RSRP) of -94.63 dBm and a synchronization signal signal-to-interference-plus-noise ratio (SS-SINR) of 10.87 dB, both categorized as “Good.” In contrast, the NLoS scenario requires substantially denser deployment with 635 gNodeBs, resulting in improved SS-RSRP performance of -71.98 dBm (“Excellent”) and SS-SINR of 12.32 dB (“Good”), along with more uniform coverage distribution. The findings indicate that improved KPI performance and coverage uniformity in NLoS environments can be achieved through substantially increased infrastructure density, highlighting the trade-off between network quality, deployment complexity, and infrastructure cost in urban 5G NR planning.
Volume: 16
Issue: 4
Page: 1944-1954
Publish at: 2026-08-01

Comparative performance analysis of lightweight face identification algorithm

10.11591/ijece.v16i4.pp2042-2060
Wuyun Wang , Suchada Sitjongsataporn
With the wide application of face recognition in resource-constrained scenarios like mobile and embedded devices, lightweight algorithms have become a research focus, but existing studies lack multi-dimensional, scenario-based performance comparisons. This paper studies the performance evaluation and application adaptation of lightweight face recognition algorithms, innovatively builds a scenario-based evaluation system, verifies the performance improvement of combining traditional algorithms with MobileNet, and constructs an efficient, stable and low-cost system. It elaborates on face recognition principles, including key links of face detection, feature extraction and matching, introduces traditional algorithms such as Eigenfaces, Fisherfaces and LBPH, and focuses on MobileNet’s characteristics: reducing computation and parameters via depthwise separable convolution, and adjustable width and resolution. Four comparative experiments verify the "traditional algorithms + MobileNet" hybrid strategy. Results show the combination achieves 98.1% accuracy, 4.3 percentage points higher than single MobileNet; LBPH + MobileNet balances performance and resource consumption best, with 110MB memory, 40% CPU usage and 315ms processing time. The hybrid strategy improves accuracy and efficiency in different scenarios, aiming to provide a scientific basis for the engineering application and subsequent optimization of lightweight face recognition algorithms, and supporting algorithm selection and performance improvement in resource-constrained scenarios.
Volume: 16
Issue: 4
Page: 2042-2060
Publish at: 2026-08-01
1 2 3
...
2091 2092
Show 44 of 2092

Discover Our Library

Embark on a journey through our expansive collection of articles and let curiosity lead your path to innovation.

Explore Now
Library 3D Ilustration