Publication Search

Image Quality Assessment has become an important research area in computer vision due to the increasing use of digital imaging in medical, industrial, and remote sensing applications. This study conducted a Systematic Literature Review (SLR) using articles retrieved exclusively from the Scopus database. A total of fifteen Scopus-indexed articles were selected through the PRISMA process consisting of identification, screening, eligibility, and inclusion stages. The findings indicated that convolutional neural networks, transformer architectures, and no-reference approaches dominated recent studies because of their ability to capture complex visual features and improve evaluation accuracy. The application domain strongly influenced method selection, particularly in medical imaging, industrial quality control, and remote sensing. Deep learning methods achieved high performance but required large datasets and high computational resources. Overall, recent research trends shifted from traditional image enhancement toward artificial intelligence-based quality evaluation.

Tomato (Solanum lycopersicum) is a high-value horticultural commodity in Indonesia, yet its cultivation is frequently disrupted by leaf diseases that are difficult to distinguish visually. Diseases such as Bacterial Spot, Early Blight, and Tomato Yellow Leaf Curl Virus often present overlapping visual symptoms, making early and accurate diagnosis a significant challenge for farmers. The manual identification methods currently in use are inefficient and error-prone, ultimately leading to reduced crop yield  and quality. The general objective of this study is to develop software capable of automatically classifying tomato  leaf diseases. Specifically, this research aims to implement the MobileNetV3 Small architecture based on Convolutional Neural  Network (CNN) with ImageNet pre-trained weights to classify 10 types of tomato leaf diseases. The research methodology encompasses dataset collection from Kaggle comprising 10,000 images (1,000 per class), image pre-processing through resizing to 224x224 pixels, and normalization, as well as hyperparameter optimization (optimizer, learning rate, epoch, batch size) via scheduler. Model performance is evaluated using a confusion matrix encompassing accuracy, precision, recall, and F1-score.

Recent Internet of Things (IoT) intrusion detection studies have reported near-perfect benchmark performance for Distributed Denial of Service (DDoS) detection, yet limited attention has been given to understanding how different traffic representations contribute to the detection process under highly imbalanced traffic conditions. This study presents an ablation-driven analysis to investigate the contribution of statistical and temporal representations for large-scale IoT DDoS detection using the CICIoT2023 dataset. Three experimental scenarios are evaluated, including statistical representation, temporal sequence representation, and hybrid statistical–temporal representation. Temporal representations are learned using a one-dimensional Convolutional Neural Network (1D-CNN) with lag-based traffic sequences, while ensemble tree-based classifiers are employed for final classification and representation analysis. In addition, multiple ablation configurations are designed to evaluate the impact of temporal dependency modeling and feature engineering strategies on detection performance. Experimental results show that statistical traffic representations remain highly effective for DDoS detection on CICIoT2023, achieving 99.36% accuracy and 99.31% weighted F1-score in the statistical representation scenario. Feature importance analysis further indicates that engineered statistical features contribute substantially more to the classification process than CNN-based temporal representations. Although temporal modeling captures sequential traffic behavior, its contribution is relatively limited and mainly acts as a complementary representation. Furthermore, the hybrid configuration produces only marginal improvements over the statistical representation alone. These findings highlight the importance of representation-level analysis for understanding the actual contribution of statistical and temporal modeling in modern IoT intrusion detection systems beyond relying solely on benchmark accuracy.

Communication is an essential human right, yet a significant communication gap persists between individuals with sensory disabilities, specifically the deaf and speech-impaired, and the general public. While many technological solutions have been proposed to translate sign language, existing models primarily rely on heavy deep learning architectures such as Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN/LSTM). These models often demand high computational power, leading to latency and limiting real-time application on standard devices. This study proposes a lightweight, fast, and highly responsive sign language translation system specifically designed to recognize static alphabets (A-Z) and single-character air writing. The system utilizes MediaPipe for hand tracking, where feature extraction is intelligently processed by calculating the relative spatial coordinates of fingertips to the wrist, reducing dependency on raw camera coordinates. Classification is performed using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel, prioritizing computational efficiency without sacrificing accuracy. To enhance user experience, the system introduces three key novelties: smart relative feature extraction, an anti-duplication hold system with a 1-second timer to prevent input spamming, and a non-blocking multithreaded audio execution (Daemon Thread) utilizing Google Text-to-Speech (gTTS), ensuring the webcam feed remains fluid during audio playback. Additionally, an alternative air-writing mode is integrated, utilizing geometric heuristics and PyTesseract OCR to read single drawn letters in the air. The results indicate that the proposed system operates swiftly and efficiently, bridging the communication barrier with a hardware-friendly approach.

Predictive maintenance (PdM) plays a crucial role in modern industrial systems by minimizing downtime, reducing maintenance costs, and optimizing asset performance. However, many predictive models operate as “black box” systems, limiting transparency and making it difficult for operators to interpret their outputs. This study aims to integrate Explainable Artificial Intelligence (XAI) techniques with Remaining Useful Life (RUL) prediction models to improve both accuracy and interpretability. Various machine learning and deep learning approaches, including Support Vector Machines (SVM), Random Forest (RF), XGBoost, Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNN), are employed to predict RUL using real-time sensor data from rotating machinery. XAI methods such as SHAP, LIME, and attention mechanisms are applied to provide human-understandable explanations of model predictions. The models are evaluated based on accuracy, Root Mean Square Error (RMSE), and interpretability scores. The results show that XAI-enhanced models outperform traditional approaches in predictive performance while offering greater transparency. These explanations help maintenance engineers better understand the factors influencing predictions, thereby improving decision-making and trust in the system. Nevertheless, the integration of XAI introduces additional computational complexity, which may pose challenges for large-scale industrial implementation. Overall, this study highlights the potential of combining XAI with RUL prediction to develop more reliable, transparent, and effective predictive maintenance solutions.

Attention mechanisms have been widely incorporated into recurrent neural network architectures for financial time series forecasting, with most prior work reporting improvements in price-level error metrics. This study revisits that claim through a controlled empirical comparison of four deep learning architectures on nearly two decades of Telkom Indonesia (TLKM) closing price data from the Indonesia Stock Exchange (IDX). The models evaluated are a three-layer Gated Recurrent Unit (GRU) baseline, a comparable Long Short-Term Memory (LSTM) network, a Bahdanau end-attention GRU (Attn-GRU-V2), and a multi-head self-attention GRU hybrid (Attn-GRU-V3). Each architecture is trained over 30 independent runs with distinct random seeds, and performance is reported as 95% confidence intervals derived from the t-distribution. Statistical comparisons employ the Wilcoxon signed-rank test, a nonparametric paired test appropriate given the confirmed non-normality of residuals. The main finding is a consistent trade-off: the plain GRU achieves the lowest RMSE (94.02 ± 1.22 IDR) across all 30 runs, while Attn-GRU-V2 achieves the highest directional accuracy (45.91 ± 0.09%), surpassing GRU in every independent run. Bahdanau attention weights are nearly uniform across the 30-day lookback window (coefficient of variation: 3.21%), indicating that the mechanism cannot identify selectively informative timesteps in this univariate price series. This finding is consistent with the weak-form Efficient Market Hypothesis for the Indonesian market. An ablation study reveals that a 20-day lookback window maximizes directional accuracy (47.72 ± 0.21%) for the Attn-GRU-V2 model. These results suggest that Bahdanau end-attention consistently and significantly improves directional accuracy relative to a plain GRU baseline, providing an architecturally attributable advantage for direction-based applications, even when absolute price-level error is not reduced. The directional accuracy values remaining below 50% across all models are consistent with a weak-form efficiency characterization of the Indonesian market.

The increasing sophistication of malware has rendered traditional signature-based detection methods insufficient, necessitating behavior-driven and adaptive analytical frameworks. This study presents a sequential deep learning framework that models system-level API call sequences as structured linguistic representations for behavioral malware detection. Unlike conventional comparative studies, this work systematically evaluates recurrent and attention-based architectures under controlled experimental conditions, with a particular focus on generalization performance and overfitting mitigation. Two neural architectures, a Long Short-Term Memory (LSTM) network and a Transformer-based attention model, are trained on publicly available API call sequence data for binary classification of malicious and benign executables. Beyond standard accuracy metrics, the study further examines model stability, convergence behavior, and the impact of long-range dependency modeling on detection robustness. Experimental results demonstrate that the Transformer architecture achieves superior performance, attaining 95.54% classification accuracy and consistent improvements in precision, recall, and F1-score, indicating a stronger ability to capture complex behavioral dependencies. These findings highlight the effectiveness of attention mechanisms in behavioral malware modeling and provide empirical evidence that NLP-inspired architectures offer a robust and scalable approach for real-world cybersecurity applications.

Automatic facial expression recognition is a significant challenge in human-computer interaction with broad relevance in mental health, security, and behavioral analysis. This study proposes the implementation of Deep Learning using a custom Convolutional Neural Network (CNN) architecture to classify seven basic emotion categories: angry, disgust, fear, happy, sad, surprise, and neutral. Key challenges such as lighting variations and visual feature ambiguity in the FER2013 dataset are addressed through image pre-processing techniques, data augmentation, and the use of Batch Normalization and Dropout layers to prevent overfitting. The research methodology involves a systematic architectural design with three main convolution blocks optimized for computational efficiency. Experimental results show that the proposed model achieved a validation accuracy of 68.2%. Performance analysis based on F1-Score reveals that the "Happy" emotion has the highest detection rate (0.85) due to contrasting facial geometric features, while the "Fear" emotion is the most difficult class to identify (0.41). This study concludes that the use of an optimized standalone CNN architecture provides competitive and efficient performance compared to heavier transfer learning models, making it feasible for implementation on devices with mid-range hardware specifications.

The increasing complexity and scale of modern network traffic driven by IoT and cloud-based infrastructures have made accurate intrusion detection a critical challenge. Conventional network intrusion detection systems (NIDS) and many deep learning–based approaches struggle to reliably detect minority and stealthy attacks due to severe class imbalance and limited discrimination of subtle traffic patterns. To address these limitations, this study proposes a hybrid CNN–RBF–Attention framework for network intrusion detection. The proposed model integrates three complementary components: (i) a convolutional neural network for hierarchical feature extraction from network flow data, (ii) a radial basis function (RBF) network for localized nonlinear classification using prototype-based decision regions, and (iii) an attention mechanism that adaptively weights RBF activations to emphasize discriminative traffic patterns. SMOTE is applied exclusively to the training data to mitigate class imbalance. The framework is evaluated on the widely used CICIDS2017 and CICIDS2018 benchmark datasets in both binary and multiclass settings, using recall, precision, F1-score, confusion matrices, and ROC analysis. Experimental results demonstrate that the proposed hybrid model consistently outperforms standalone CNN and RBF baselines, particularly in terms of recall and F1-score. On the CICIDS2018 dataset, the model achieves 99.81% accuracy and 99.81% F1-score in binary classification, and 99.54% accuracy and 99.54% F1-score in multiclass classification. On CICIDS2017, it achieves 98.12% accuracy and 98.12% F1-score in binary classification, and 98.92% accuracy and 98.92% F1-score in multiclass classification. Confusion matrix and ROC analyses further show strong class separability and reliable performance in low–false-positive-rate regions, which is critical for real-world IDS deployment. These results confirm that combining deep hierarchical feature learning, localized prototype-based classification, and attention-guided refinement yields a robust, operationally reliable intrusion detection framework for highly imbalanced network environments.

Open-pit mining activities in Indonesia contribute significantly to the national economy but require stringent monitoring to mitigate environmental degradation. Conventional monitoring methods relying on terrestrial surveys are often constrained by vast coverage areas, high operational costs, and limited field accessibility. This study aims to develop an artificial intelligence model capable of automatically detecting and mapping mining areas to enhance surveillance efficiency. The applied method is Deep Semantic Segmentation utilizing the U-Net Convolutional Neural Network (CNN) architecture. The model was trained using Sentinel-2 satellite imagery, focusing exclusively on Red, Green, and Blue (RGB) spectral channels to replicate human visual perception. Experimental results demonstrate that the proposed model performs reliable segmentation of mining areas, achieving an Accuracy of 93.58% and a Global Intersection over Union (IoU) of 0.8067. These findings indicate that the U-Net architecture can effectively extract spatial features of mines even when utilizing standard visual data. This research contributes to the development of an efficient, cost-effective, and scalable digital monitoring prototype to support innovation in sustainable environmental governance.

The increasing sedentary lifestyle in the digital era has the potential to cause various health problems due to lack of physical activity. One approach that can be taken to encourage physical activity is through the use of digital games with body movement-based control mechanisms. This study aims to develop a body gesture-based game character control system using a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model. CNN is used to extract spatial features from each video frame, while LSTM serves to model the temporal relationship between frames so that movement patterns can be recognized sequentially. The research method used refers to the Machine Learning Lifecycle stages, starting from data collection, preprocessing, model development, to implementation in the endless runner game genre. Testing results show that the CNN–LSTM model is capable of classifying body gestures and generating outputs that can be used as commands to control game characters. The implementation of this system enables more natural and interactive game interactions without conventional input devices, and has the potential to encourage players to lead a more active lifestyle.

Financial distress is a condition of declining financial health of a company that can develop gradually and lead to business failure if not detected early. With the increasing complexity of the business environment and the limitations of conventional statistical methods, Artificial Intelligence/AI is increasingly being adopted in the development of early warning systems (EWS) to predict financial distress. This study aims to examine the development of AI-based EWS research, identify the most widely used algorithms, and evaluate the effectiveness of AI models compared to conventional methods in predicting financial distress. The method used is a comprehensive systematic literature review of 15 relevant scientific articles. The results show that the paradigm has shifted from statistical models to machine learning and deep learning. Random Forest and Artificial Neural Network are the most widely used algorithms and have better predictive performance. This study offers a conceptual synthesis of the progress, effectiveness, and challenges of applying AI in predicting financial distress and opens opportunities for further research on the development of contextual and interpretative EWS.

This study is motivated by the increasing need for accurate modeling and classification of one-dimensional signal data in intelligent systems. The rapid development of deep learning has led to the adoption of more adaptive and complex neural network architectures capable of capturing both temporal dependencies and local patterns in sequential data. This research aims to analyze and compare the performance of several deep learning models, namely Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and a hybrid Convolutional Neural Network–GRU (CNN–GRU) model for signal data classification. The research method employs a quantitative experimental approach involving data preprocessing, windowing, model training, and performance evaluation. Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. The results indicate that the hybrid CNN–GRU model outperforms the other models, particularly in capturing local features and long-term temporal dependencies within signal data. These findings suggest that the integration of convolutional layers and recurrent mechanisms enhances feature representation and learning stability. This study is expected to contribute both theoretically and practically to the development of deep learning models for signal processing and time-series-based intelligent applications.

Human–Robot Interaction (HRI) systems increasingly rely on data-driven approaches to interpret multimodal sensory inputs and support natural interaction. However, purely neural-based HRI models often suffer from limited interpretability and insufficient context-aware decision-making, which can reduce user trust and adaptability in dynamic interaction scenarios. To address these limitations, this study proposes a hybrid neural–symbolic HRI framework that integrates multimodal neural perception with explicit symbolic reasoning for adaptive and interpretable robot behavior. The proposed system combines deep neural networks for processing visual, speech, and gesture inputs with a rule-based symbolic reasoning layer that models interaction context, user states, and behavioral constraints. A loosely coupled integration strategy enables neural outputs to be transformed into symbolic representations, allowing logical inference to guide action selection while preserving perceptual accuracy. The framework was evaluated through controlled HRI experiments comparing a neural-only baseline with the proposed hybrid configuration across multiple interaction scenarios. Experimental results demonstrate that the hybrid neural–symbolic system significantly improves interaction accuracy, contextual responsiveness, and user satisfaction, while achieving substantial gains in interpretability. These findings indicate that symbolic reasoning effectively complements neural perception by enhancing transparency and context-aware adaptation without compromising performance. The study concludes that hybrid neural–symbolic architectures provide a promising foundation for developing trustworthy, adaptive, and human-centered HRI systems.

The increasing adoption of collaborative robots in modern manufacturing environments requires reliable perception systems that can ensure both safety and operational efficiency during human–robot collaboration. This study proposes a CNN-based real-time computer vision system for object and human detection in shared robotic workspaces. The research focuses on developing and evaluating a single-stage deep learning detection model optimized for real-time performance while maintaining high detection accuracy. The proposed methodology includes dataset preparation, model training using transfer learning, real-time system implementation, and comprehensive performance evaluation. Experimental results demonstrate that the developed system achieves high detection accuracy, as reflected by strong precision, recall, and mean Average Precision (mAP) values, while maintaining low inference latency suitable for real-time operation. The system consistently operates above real-time frame-rate thresholds, ensuring timely perception updates required for safety-related decision-making in collaborative robotic environments. Graphical and quantitative analyses further confirm the stability of inference performance under dynamic interaction scenarios involving human movement and multiple objects. Compared with existing approaches, the proposed system provides a balanced trade-off between accuracy and computational efficiency, making it practical for deployment in safety-aware human–robot collaboration scenarios. Overall, the findings indicate that CNN-based real-time object detection systems can effectively support perception and situational awareness in collaborative robotics, contributing to safer and more efficient industrial automation.

This study explores the integration of deep learning based approaches in real time video content analysis for intelligent human computer interaction (HCI) in multimedia systems. Traditional video analysis techniques, such as rule-based methods and offline processing, struggle with real time performance and adaptability to complex video data. In contrast, the deep learning model used in this research, particularly Convolutional Neural Networks (CNNs), provides high accuracy in object detection, feature extraction, and real time processing. The integration of CNNs with interactive visualization modules enables dynamic adjustments to video content based on user interactions, ensuring a seamless and engaging user experience. The system was benchmarked in terms of its processing speed, accuracy, and responsiveness, showing significant improvements over traditional approaches in real time video analysis. Moreover, the study demonstrates that combining deep learning with real time visualization enhances the efficiency of interactive multimedia applications, making it suitable for dynamic environments such as surveillance, security monitoring, and interactive media. Despite the system's strong performance, challenges such as computational demands in high-resolution video processing were identified, highlighting the need for further optimization. Future work will focus on optimizing the system for different hardware platforms, incorporating multimodal inputs, and refining deep learning models to address computational bottlenecks. This research contributes to advancing HCI by providing insights into the integration of deep learning for real time video content analysis, which is pivotal for enhancing the interactivity and adaptability of intelligent multimedia systems.

Explainable Artificial Intelligence (XAI) has become a critical area of research within artificial intelligence, focusing on improving the transparency and interpretability of machine learning (ML) models, often referred to as "black-box" models. The need for XAI techniques arises from the inherent complexity of ML models, which can make their decision-making processes difficult for users to understand. This study investigates various XAI techniques, including LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations), to assess their impact on model interpretability without significantly compromising predictive performance. A comparative experimental design was used, applying these XAI methods to different ML models, including deep neural networks and ensemble methods, within large-scale enterprise data analytics systems. The results indicate that XAI methods significantly enhance model transparency and decision traceability, allowing users to understand the influence of individual features on predictions. While a slight reduction in predictive accuracy was observed, especially with simpler models, the trade-off between interpretability and performance was deemed acceptable, particularly in fields requiring transparency, such as healthcare, finance, and autonomous systems. The use of XAI in enterprise data systems has practical implications for fostering trust and enabling informed decision-making among stakeholders. Furthermore, the study discusses the challenges and limitations of applying XAI techniques, such as complexity, scalability, and model-specific limitations. Future research is suggested to focus on developing more scalable and efficient XAI methods, enhancing their applicability across various model types, and addressing the challenges of real-time applications. This will be crucial in ensuring the widespread adoption of XAI in critical domains, promoting the ethical use of AI while maintaining predictive accuracy.

This research aims to develop a multiclass flower image classification system using the Convolutional Neural Network (CNN) algorithm with the EfficientNet architecture. The main problem addressed is the difficulty of manual identification of flower species that share high visual similarity. The research stages include collecting 17,299 flower images across 19 classes, performing data preprocessing such as image resizing, pixel normalization, and augmentation, followed by model training using the EfficientNet transfer learning approach. The model was trained for 10 epochs with an 80:20 training-validation data split. The evaluation results show that the model achieved a validation accuracy of 98.05% with a loss value of 0.0968, and an average precision, recall, and F1-score of 0.98. The trained model was then implemented into a web-based application built using the Next.js framework, enabling users to upload flower images and obtain real-time classification results via the Hugging Face API. The system successfully identified flower species with a confidence level of 99.87%. These findings demonstrate that combining a modern CNN architecture with transfer learning provides efficient and highly accurate flower classification performance, which can be effectively implemented for educational and digital conservation purposes.

Communication is a fundamental aspect of human life. However, individuals with hearing and speech impairments often face barriers in communicating with the general public. The Indonesian Sign System (SIBI) serves as a communication solution for the deaf and speech-impaired community in Indonesia, yet public understanding of SIBI remains limited. To address this issue, this study aims to develop an automatic translation model from SIBI sign language into Indonesian text by utilizing Deep Learning technology, specifically the Convolutional Neural Network (CNN) algorithm. CNN was chosen for its ability to effectively recognize visual patterns, making it suitable for processing hand gesture images in sign language. This research involved collecting and classifying a dataset of hand images based on the alphabet or words in SIBI, which were then used to train the CNN model. The designed CNN model was built to accurately classify hand signs and translate them into Indonesian text. The results of this study have the potential to serve as a supportive solution for inclusive communication between the deaf community and the wider public, and can be further developed for contextual sentence translation. Keywords: Indonesian Sign System (SIBI), CNN, Deep Learning, Automatic Translation, Inclusive Communication

Deteksi nyeri secara objektif merupakan tantangan penting dalam dunia medis, terutama bagi pasien yang tidak mampu menyampaikan rasa sakitnya secara verbal, seperti bayi, lansia, atau penderita gangguan komunikasi. Teknologi non- invasif berbasis sensor menjadi solusi potensial untuk mengatasi keterbatasan metode subjektif. Penelitian ini bertujuan meninjau secara sistematis literatur terkini mengenai penerapan Radar MIMO dan algoritma kecerdasan buatan dalam deteksi nyeri non-invasif. Metode yang digunakan adalah Systematic Literature Review (SLR) dengan pedoman PRISMA 2020, melalui penelusuran basis data IEEE Xplore, ScienceDirect, PubMed, Google Scholar, dan SpringerLink untuk periode 2021– 2025. Dari hasil seleksi diperoleh 17 artikel inklusi yang mencakup penggunaan Radar MIMO, UNBC-McMaster, BioVid, Medical Imaging (CT/MRI), Radar SISO, serta studi review, survey, bibliometrik, dan teoretis. Dari sisi algoritma, CNN dan SVM menjadi pendekatan paling dominan, diikuti Neural Network dan metode lain, dengan tren yang mengarah pada penggunaan multimodal untuk meningkatkan akurasi. Hasil penilaian kualitas dengan GRADE menunjukkan mayoritas studi berkualitas sedang, dengan keterbatasan utama pada ukuran sampel kecil, pelabelan nyeri yang belum konsisten, bias populasi, serta kurangnya validasi klinis nyata. Kesimpulannya, Radar MIMO dan algoritma deep learning memiliki potensi besar untuk deteksi nyeri non-invasif. Namun, penelitian lanjutan perlu difokuskan pada pembangunan dataset yang lebih inklusif, standarisasi pelabelan nyeri, serta pengujian dalam konteks klinis, dengan memperhatikan aspek etika dan privasi agar teknologi ini dapat diimplementasikan secara luas dalam layanan kesehatan.