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

30,907 Article Results

Multi-stage hybrid YOLO-driven and MobileNetV2-CNN variants for robust fish freshness classification

10.11591/ijai.v15.i4.pp3660-3671
Raseeda Hamzah , Rosniza Roslan , Amni Munira Khidir , Lala Septem Riza
This study presents a multi-stage hybrid representation learning framework for robust fish freshness classification to address the critical challenge of reliable quality assessment in unpreserved food supply chains. The proposed pipeline operates in two stages: stage-1 employs you only look once (YOLO)v8n as feature-gated detector to validate inputs and eliminate non-fish images, while stage-2 leverages transfer-learned MobileNetV2 variants enhanced with convolutional neural network (CNN) layers, fine-tuning, adaptive learning rate schedulers, and expanded fully connected layers for hierarchical classification into three freshness classes i.e., highly fresh, fresh, and not fresh. The framework has been trained and evaluated on two curated datasets comprising 4,500 fish and non-fish images and 11,111 freshness-labeled including augmented dataset to increase diversity. The experimental results showed significant improvements. The best performance model transfer learning (TL)-MobileNetV2 + CNN + fine-tuning achieved 98.07% training accuracy and 67.72% validation accuracy on 80:20 split, and training accuracy of 97.57%, validation accuracy of 97.21% through 10-fold cross-validation. The comparative benchmarking confirmed that dual-stage design outperformed baseline MobileNetV2 and YOLOv5s models across precision, recall, and F1-score. The findings highlighted significant value of integrating detection-driven validation with transfer-learning classification, and propose new benchmark for intelligent freshness monitoring. For future work, this study aims to explore attention-based models, data integration, and species diversity.
Volume: 15
Issue: 4
Page: 3660-3671
Publish at: 2026-08-01

AI-driven hybrid neural network for electrocardiogram-based authentication and predictive health monitoring

10.11591/ijai.v15.i4.pp3309-3317
Swati Lakshmi Boppana , Velicheti Anantha Lakshmi , Padala SriKavitha , Venkata Ashok Kalaga , Venkateswara Rao Naramala , Suneetha Thalluru , Gunturi S. Raghavendra
The increasing adoption of digital healthcare systems demands secure and reliable patient authentication mechanisms. Traditional methods such as passwords and PINs are vulnerable to security breaches, motivating the use of biometric-based solutions. Among various biometrics, the electrocardiogram (ECG) signal is distinctive, stable, and non-invasive, making it suitable for secure authentication. This paper proposes CardioGuard, an artificial intelligence (AI)–based authentication framework that employs a hybrid deep learning model combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to extract discriminative features from ECG signals and classify users as genuine or impostors. In addition to access control, the system analyzes ECG patterns to support early detection of potential cardiovascular abnormalities. Experimental results demonstrate that CardioGuard achieves improved authentication accuracy and enhanced predictive health insights compared to conventional approaches, highlighting its effectiveness for secure and intelligent healthcare monitoring.
Volume: 15
Issue: 4
Page: 3309-3317
Publish at: 2026-08-01

Advanced personal bankruptcy prediction using tree-based deep learning models

10.11591/ijeecs.v43.i2.pp640-650
Nhat Nguyen Minh , Duy Ngo Hoang Khanh
Due to the unstable economic conditions, worsened by the post-COVID-19 environment and persistent foreign wars in 2024, financial institutions have growing difficulties in accurately predicting customer default probability. This study examines the use of sophisticated tree-based deep learning and deep neural network models for forecasting personal bankruptcy. This research utilises a dataset of roughly 9,800 individuals from Vietnamese financial institutions, spanning from 2012 to 2022, to evaluate the efficacy of models including neural decision tree, deep forest, tabular convolutional neural networks (TBCNN), and neural oblivious decision ensembles (NODE). The results demonstrate that the Deep Forest model far surpasses its competitors, providing nearly flawless predicted accuracy and enhanced interpretability. The findings highlight the efficacy of tree-based deep learning and deep neural network models as effective instruments for financial risk management, especially in volatile and unpredictable economic environments.
Volume: 43
Issue: 2
Page: 640-650
Publish at: 2026-08-01

Coordinated multi-battery control for single-stage islanded AC microgrids

10.11591/ijeecs.v43.i2.pp384-399
Adhi Kusmantoro , Lukman Harun
Reliable power management is essential for islanded AC microgrids integrating photovoltaic (PV) generation and battery energy storage, particularly under variable solar irradiance and load conditions. This paper proposes a coordinated multi-battery control strategy using a single-stage AC-coupled configuration to ensure uninterrupted power supply while improving system reliability. The proposed topology employs two PV arrays: one directly connected to a battery inverter to supply the AC load and another dedicated to battery charging. Four battery units are coordinated through a fuzzy logic controller (FLC), which sequentially regulates battery discharge based on PV generation and load demand. Simulation studies were conducted under two operating scenarios. In the first scenario, at a solar irradiance of 1000 W/m², the PV system generated approximately 1200 W, and the proposed controller maintained power balance as the load demand increased. In the second scenario, when PV output decreased due to reduced solar irradiance or complete source interruption, the FLC coordinated battery operation at 0.07 s, 0.34 s, 0.64 s, and 0.93 s, ensuring continuous power delivery to the load. The simulation results demonstrate that the proposed coordinated control strategy effectively enhances power continuity, operational stability, and energy management in single-stage islanded AC microgrids.
Volume: 43
Issue: 2
Page: 384-399
Publish at: 2026-08-01

Simulation of a frequency-reconfigurable multiband antenna for 5G and Wi‑Fi 6E/7 applications

10.11591/ijeecs.v43.i2.pp439-449
Ismahane Refsi , Miloud Benchehima , Mohammed Hicham Hachemi , Salah Eddine Brezini
This paper presents a compact frequency-reconfigurable multiband antenna design suitable for 5G and Wi‑Fi 6E/7 applications. Designed on a low-cost FR4 substrate, the antenna occupies a compact footprint of 28×15×1.6 mm, making it one of the smallest designs reported and suitable for integration into space-constrained devices. Its simple geometry enables easy design and integration across platforms. Frequency reconfigurability is achieved using PIN-diode-controlled reactive elements that manipulate current distribution to shift resonant frequencies. The design and optimization were performed using CST Microwave Studio to ensure accurate electromagnetic performance. Simulation results demonstrate that the proposed design supports up to eight distinct operating modes, highlighting its versatile frequency-reconfigurable capability. Among these modes, some achieve excellent performance at specific frequency bands including Wi‑Fi 6E/7 (2.4, 5 and 6 GHz) and 5G sub-6 GHz (3.5 GHz). These findings are confirmed with simulation results showing clearly that the antenna exhibits S_11 below –20 dB across the targeted frequency bands. The antenna also demonstrates satisfactory gain, with values above 1.5 dBi across the desired frequency bands. The antenna reveals a voltage standing wave ratio (VSWR) below 1.5 across the intended frequency bands. These characteristics make the proposed antenna a compact, versatile and efficient solution for current and future wireless communication systems.
Volume: 43
Issue: 2
Page: 439-449
Publish at: 2026-08-01

Affine-invariant feature learning for accurate ulcer detection in wireless capsule endoscopy images

10.11591/ijeecs.v43.i2.pp595-606
S. Bhuvaneswari , M. Sulthan Ibrahim
Ulcers are lesions that develop in the lining of the gastrointestinal (GI) tract, particularly in the stomach and small intestine, and may lead to severe complications such as Crohn’s disease and ulcerative colitis if not detected at an early stage. Conventional endoscopic procedures are often uncomfortable for patients and may provide limited visualization of the entire small intestine. Wireless capsule endoscopy (WCE) has emerged as a non-invasive alternative for comprehensive GI tract examination; however, automated ulcer detection from WCE images remains challenging due to image noise, complex tissue structures, and computational requirements. To address these issues, this paper proposes a Camargo’s Indexive Kuwahara filtering-based affine-invariant sliced regression (CIKF-AISR) framework for accurate and efficient ulcer detection. The proposed framework consists of image acquisition, preprocessing, segmentation, and feature extraction stages. Adaptive CIKF is employed to suppress noise while preserving edge information. Subsequently, Von Neumann locality segmentation combined with the Canberra distance measure is utilized to identify regions of interest (ROIs). Finally, affine-invariant saliency sliced regression extracts discriminative shape, color, and texture features for ulcer detection. Experimental evaluation on the Hyper-Kvasir dataset demonstrates that the proposed method achieves higher ulcer detection accuracy, improved precision, enhanced peak signal-to-noise ratio (PSNR), and lower detection time compared with existing deep CNN and VAE-GAN approaches. These results confirm the effectiveness of the proposed framework for computer-aided GI diagnosis.
Volume: 43
Issue: 2
Page: 595-606
Publish at: 2026-08-01

Video summarization using deep image captioning models

10.11591/ijeecs.v43.i2.pp547-554
Qudes M. B. Aljelawy , Sarah S. Mohammed , Entessar K. Hanoun
This research presents a novel approach for video summarization by leveraging deep image captioning models. A pretrained image captioning model, namely Salesforce's bootstrapped language image pretraining (BLIP), is used to extract keyframes from a video at regular intervals and produce natural language descriptions. These textual descriptions are then filtered for non-repetition and concatenated into a coherent summary, allowing users to understand the video’s content without viewing it in full. The proposed framework aims to improve video browsing, indexing, and retrieval efficiency, particularly for big datasets. The proposed method achieves a significant reduction in redundancy by 40% compared to raw captioning sequences. Evaluation using semantic consistency checks demonstrates that the BLIP-based framework maintains high descriptive accuracy even in complex scenes, providing a scalable solution for large-scale video indexing.
Volume: 43
Issue: 2
Page: 547-554
Publish at: 2026-08-01

An artificial neural network-based decision support model for early prediction of mathematics learning challenges using the CRISP DM framework

10.11591/ijeecs.v43.i2.pp618-627
Harry Dhika , Surajiyo Surajiyo , Lasia Agustina , Abdul Muchlis
Identifying mathematics learning difficulties remains a challenge for educators due to the subjectivity and inefficiency of conventional methods in capturing psychological factors. To address this limitation, this study proposes an artificial neural network (ANN)-based decision support model developed within the cross-industry standard process for data mining (CRISP-DM) framework. The model integrates 16 academic indicators (quizzes, exams, remedial frequency, online activity) and psychological factors (motivation, anxiety, self-confidence, interest) from 163 student records at SMA Muhammadiyah 16 Jakarta. Synthetic minority over-sampling technique (SMOTE) and focal loss are applied to handle class imbalance and improve reliability. The proposed model achieves 98% accuracy and a 0.97 F1-score in classifying students into three difficulty levels: Easy, moderate, and difficult. These findings demonstrate the model’s effectiveness in capturing complex relationships between cognitive and affective features. Unlike prior studies that rely solely on academic performance, this work contributes a robust, comprehensive data-driven framework that enhances multiclass classification for early educational intervention within the Indonesian context.
Volume: 43
Issue: 2
Page: 618-627
Publish at: 2026-08-01

Evaluating AI-powered ChatGPT for intelligent virtual learning environments through conversational intelligence

10.11591/ijeecs.v43.i2.pp568-575
Isaac Asampana , Henry Akwetey Matey , Ben Ocra , Jones Yeboah Nyame
The rapid advancement of conversational artificial intelligence has transformed virtual learning by enabling intelligent, adaptive, and interactive educational support through natural language communication. Among recent AI technologies, chat generative pre-trained transformer (ChatGPT) has emerged as a powerful conversational intelligence system capable of providing real-time explanations, personalized learning assistance, and contextual academic interactions. This study evaluates the effectiveness of AI-powered ChatGPT in intelligent virtual learning environments through the lens of conversational intelligence. A quantitative survey instrument, developed based on established technology acceptance constructs, was employed to assess users' perceptions of ChatGPT's usefulness, ease of interaction, and educational effectiveness. Using a simple random sampling technique, data were collected from 1,104 participants and analyzed using descriptive statistical methods in SPSS version 25. The results indicate that ChatGPT's conversational capabilities significantly enhance learner engagement, support personalized learning experiences, and improve perceived academic performance through intuitive human-AI interactions. The findings demonstrate that conversational intelligence plays a critical role in improving the effectiveness of intelligent virtual learning environments by facilitating adaptive knowledge delivery and interactive learning support. This study provides valuable insights into the integration of conversational AI technologies for developing next-generation intelligent digital education systems.
Volume: 43
Issue: 2
Page: 568-575
Publish at: 2026-08-01

Confidence-driven adaptive operating-point optimization for photovoltaic systems under partial shading conditions

10.11591/ijeecs.v43.i2.pp672-682
Sarah Kawther Sedjar , Mourad Benmessaoud
Partial shading conditions (PSC) generate highly nonlinear multi-peak photo voltaic (PV) characteristics, complicating reliable global maximum power point tracking (GMPP). Although numerous intelligent optimization techniques exist, most rely on extensive exploration mechanisms that limit their applicability in embedded real-time controllers. This paper introduces a confidence-driven adaptive maximum power point tracking (MPPT) framework in which the optimization search space is dynamically regulated according to the reliability of a predictive operating-region estimator. Unlike standard artificial neural network (ANN)-assisted MPPT strategies that offer only power prediction, the proposed approach exploits a confidence index to continuously contract or expand the exploration domain, minimizing search effort while preserving global tracking capability. The framework was developed using real measurements from the PV DAQ database and validated through a nonlinear two-diode thermal-electrical PV model incorporating irradiance mismatch and temperature-dependent effects. Static and dynamic PSC scenarios were investigated to evaluate convergence behavior and computational performance. Experimental results demonstrate that the proposed confidence-governed strategy achieves an average tracking efficiency of 90.89%, reduces convergence effort through adaptive search-space contraction, and matches real measurements with an R2 value of 0.8664, offering a low-complexity solution for real-time embedded PV energy management.
Volume: 43
Issue: 2
Page: 672-682
Publish at: 2026-08-01

An overview of SSP structures in dragon graphs and domination and independent domination fuzzy SSP graphs

10.11591/ijeecs.v43.i2.pp485-494
R. Mary Jeya Jothi
Social network modeling is one of the many applications of dominant sets in graph theory. Social networks are made up of personalities or sets of individuals linked by relationships; graph theory provides a useful framework for examining and simulating these structures. We have already covered the idea of domination and how to calculate the domination numeral for various fuzzy graphs. Unlike anything found in the literature to date, we presented the idea of domination of fuzzy SSP graphs in this study. It is also spoken about the several super strongly perfect (SSP) structural parameters in fuzzy SSP graphs. The study appears to contribute to the ground of theory of graphs by extending the notion of domination to fuzzy SSP graphs. The paper also covers how to acquire specific dominant sets, coloring numbers, and cliques (maximal) in the context of fuzzy SSP graphs. In these kinds of scenarios, researchers can employ the following broad techniques to identify dominating sets.
Volume: 43
Issue: 2
Page: 485-494
Publish at: 2026-08-01

Low-cost real-time campus assistive navigation device for the visually impaired: The University of Ilorin case study

10.11591/ijeecs.v43.i2.pp555-567
Mahmud Hafeez Owolabi , Idajili John Ojochegbe , Ayinla Shehu Lukman , Jimoh-Mahmud Aishat Oladayo , Yusuf Abdulrahman Olalekan , Olaogun Jerry Oluwajomiloju
Navigating dynamic environments remains a significant challenge for visually impaired individuals due to limited spatial awareness, which restricts their mobility, independence, and safety. Traditional aids such as white canes and guide dogs provide physical support but lack contextual feedback. While recent advances in artificial intelligence (AI) and computer vision (CV) have enabled real-time sensing, many assistive systems remain costly, complex, and dependent on continuous network connectivity. This study introduces a low‑cost, portable campus assistive navigation (CAN) device that delivers real‑time perception and offline guidance. The system employs a custom dataset, an optimized YOLOv11 model, and a Pi Camera for continuous object detection, supported by an HC‑SR04 ultrasonic sensor for obstacle avoidance and auditory alerts. Training and validation confirmed robust convergence, with precision ~0.95, recall near 1.0, and mAP@0.5 exceeding 0.9, while mAP@0.5:0.95 remained above 0.8, demonstrating reliable detection and generalization under strict thresholds. Field tests further reported confidence scores of 0.74–0.84, 98% accuracy in distance measurement, and GPS localization within ±1.5 m. Real‑time auditory and haptic feedback via Bluetooth headphones enhanced mobility and safety. The CAN device offers a scalable, affordable solution for autonomous navigation in campus environments.
Volume: 43
Issue: 2
Page: 555-567
Publish at: 2026-08-01

Evaluation of hybrid parallelism for scalable training of DenseNet-121 in diabetic retinopathy classification

10.11591/ijeecs.v43.i2.pp662-671
Indar Sugiarto , Djoni Haryadi Setiabudi , Darrell Cornelius Rivaldo , Taweesak Kijkanjanarat , Resmana Lim
Training large and complex deep learning models is often constrained by GPU memory limitations and prolonged training times. While several parallelism strategies have been proposed, this study specifically evaluates hybrid parallelism—a combination of data parallelism and pipeline parallelism—to address both challenges simultaneously. Using a case study on diabetic retinopathy (DR) classification with the DenseNet-121 architecture, we analyze the trade-off between computational efficiency and memory scalability. Results show that although hybrid parallelism does not yet provide speedup compared to a single-GPU setup—due to communication overhead and pipeline fragmentation—it enables training of large models that exceed the memory capacity of a single GPU. The trained model achieved a validation accuracy of 0.737, a quadratic weighted kappa (QWK) of 0.861, and a weighted F1-score of 0.749. In contrast, pure data parallelism showed a potential speedup of up to 1.9× in scenarios where the model still fits within a single GPU. These findings highlight the critical role of hybrid parallelism in overcoming the memory wall in large-scale model training, though optimization to reduce overhead remains a key challenge.
Volume: 43
Issue: 2
Page: 662-671
Publish at: 2026-08-01

Autonomous trenching robot with intelligent obstacle detection and path optimization for precision cable installation

10.11591/ijeecs.v43.i2.pp425-438
Muhammad Omar , Hamza Ali Nisar , Muhammad Usman , Husnain Siddique , Suffian Zaman , Saad Saleem Khan , Justyna Robinson
Trenching for underground cable and pipeline installation is typically labor intensive, time-consuming, and potentially hazardous, particularly in environments with buried obstacles. This paper presents a low-cost autonomous trenching robot with intelligent obstacle detection and path optimization to improve excavation efficiency, safety, and accuracy. The proposed system integrates ultrasonic and infrared sensors with an embedded controller for real-time obstacle detection and autonomous navigation. A path optimization algorithm automatically adjusts the trenching route whenever an obstacle is detected, allowing continuous operation while reducing unnecessary movement and energy consumption. The robot employs a tracked mobile platform and an automated trenching mechanism capable of maintaining consistent trench depth and width under different terrain conditions. Experimental results demonstrate that the proposed system accurately detects obstacles, successfully replans its path in real time, and performs reliable autonomous trenching with minimal human intervention. Compared with conventional manual trenching methods, the developed robot improves operational efficiency, enhances excavation accuracy, and reduces safety risks for workers. The proposed system provides a practical and scalable solution for underground cable and pipeline installation and has strong potential for future applications in intelligent construction, infrastructure development, and autonomous civil engineering.
Volume: 43
Issue: 2
Page: 425-438
Publish at: 2026-08-01

Probability density-based quantized spiking neural network for efficient intrusion detection in mobile ad hoc networks

10.11591/ijeecs.v43.i2.pp460-471
Amruth Veerabhadraiah , Devaraj Verma Chitragar
Mobile ad hoc networks (MANETs) enable infrastructure-free wireless communication through decentralized and self-organizing network architectures, making them well suited for dynamic environments such as disaster recovery, military operations, and intelligent transportation systems. However, their distributed topology and continuously changing network structure expose them to sophisticated routing attacks, particularly blackhole attacks (BHA) and wormhole attacks (WHA), which significantly degrade network reliability and challenge conventional intrusion detection techniques. Existing deep learning approaches often struggle to accurately distinguish malicious from legitimate traffic because of high-dimensional feature spaces and limited computational resources in MANET environments. To address these challenges, this paper proposes a probability density-based quantized spiking neural network (PDF-QSNN) framework for efficient intrusion detection in MANETs. The proposed framework first employs adaptive moth flame optimization (AMFO) to identify the most informative features, thereby reducing data redundancy and computational complexity. Subsequently, probability density-based feature encoding generates a discriminative probabilistic representation of network behavior, while the quantized spiking neural network performs lightweight and energy-efficient intrusion classification. Experimental evaluations using simulated BHA and WHA datasets demonstrate that the proposed framework achieves classification accuracies of 92.86% and 91.56%, respectively, outperforming conventional convolutional neural networks (CNN) and stacked recurrent long short-term memory (SRLSTM) models. These results demonstrate the effectiveness of the proposed framework for accurate and computationally efficient intrusion detection in resource constrained MANET environments.
Volume: 43
Issue: 2
Page: 460-471
Publish at: 2026-08-01
1 2 3
...
2060 2061
Show 17 of 2061

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