Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Comparative Analysis of MobileNetV3 and EfficientNetv2B0 in BISINDO Hand Sign Recognition Using MediaPipe Landmarks
No full textSign language is a vital communication medium for individuals with hearing and speech impairments. In Indonesia, more than 2.6 million people experience hearing disabilities, most of whom rely on Bahasa Isyarat Indonesia BISINDO for daily interaction. However, limited public understanding and the scarcity of professional interpreters continue to hinder inclusive communication. Recent advancements in computer vision and deep learning have enabled camera-based sign language recognition systems that are more affordable and practical compared to sensor-glove solutions. this study presents a comparative analysis between EfficientNetV2-B0 and MobileNetV3-Large in recognizing BISINDO hand sign alphabets using MediaPipe landmarks. The dataset was derived from BISINDO video recordings, from which hand landmarks were extracted using MediaPipe Hands and subsequently converted into two-dimensional skeletal images. In total, 10,309 skeletal images representing BISINDO alphabets A–Z were generated and used for model training and evaluation. Both models were trained under identical configurations using TensorFlow. The results show that MobileNetV3-Large achieved 89.67% test accuracy and an F1-score of 89.76%, while EfficientNetV2-B0 obtains 95.98% test accuracy and an F1-score of 95.93%. These findings highlight the trade-off between the higher classification accuracy of EfficientNetV2-B0 and the superior computational efficiency of MobileNetV3-Large. This research contributes to the development of lightweight, high-performance BISINDO recognition systems for assistive communication applications.Sign language is a vital communication medium for individuals with hearing and speech impairments. In Indonesia, more than 2.6 million people experience hearing disabilities, most of whom rely on Bahasa Isyarat Indonesia BISINDO for daily interaction. However, limited public understanding and the scarcity of professional interpreters continue to hinder inclusive communication. Recent advancements in computer vision and deep learning have enabled camera-based sign language recognition systems that are more affordable and practical compared to sensor-glove solutions. this study presents a comparative analysis between EfficientNetV2-B0 and MobileNetV3-Large in recognizing BISINDO hand sign alphabets using MediaPipe landmarks. The dataset was derived from BISINDO video recordings, from which hand landmarks were extracted using MediaPipe Hands and subsequently converted into two-dimensional skeletal images. In total, 10,309 skeletal images representing BISINDO alphabets A–Z were generated and used for model training and evaluation. Both models were trained under identical configurations using TensorFlow. The results show that MobileNetV3-Large achieved 89.67% test accuracy and an F1-score of 89.76%, while EfficientNetV2-B0 obtains 95.98% test accuracy and an F1-score of 95.93%. These findings highlight the trade-off between the higher classification accuracy of EfficientNetV2-B0 and the superior computational efficiency of MobileNetV3-Large. This research contributes to the development of lightweight, high-performance BISINDO recognition systems for assistive communication applications
Benchmarking Oversampling Strategies to Enhance the Performance of Machine Learning Algorithms in Hypertension Classification
No full textThis study benchmarks the effectiveness of three oversampling techniques, namely SMOTE, Random Oversampling (ROS), and ADASYN, in enhancing machine learning performance for multiclass hypertension classification. Using key physiological features and four optimized algorithms Logistic Regression, Support Vector Machine, Linear Discriminant Analysis, and Artificial Neural Networks, model performance was assessed using accuracy, F1-macro, and ROC AUC metrics. The experimental results indicate that the combination of SMOTE and Linear Discriminant Analysis (LDA) yields the highest overall performance, achieving an accuracy of 0.9773 and an F1-macro score of 0.9848. Logistic Regression demonstrates optimal results when paired with ROS, also reaching an accuracy of 0.9773. Artificial Neural Networks show the most substantial performance improvement under ADASYN, particularly reflected in higher F1-macro values. Although Support Vector Machine is less sensitive to oversampling interventions, it achieves a strong ROC AUC score of 0.9776 when trained using SMOTE. Overall, the findings confirm that oversampling techniques significantly improve classification performance in multilevel hypertension prediction, with SMOTE combined with LDA emerging as the most effective configuration.This study benchmarks the effectiveness of three oversampling techniques, namely SMOTE, Random Oversampling (ROS), and ADASYN, in enhancing machine learning performance for multiclass hypertension classification. Using key physiological features and four optimized algorithms Logistic Regression, Support Vector Machine, Linear Discriminant Analysis, and Artificial Neural Networks, model performance was assessed using accuracy, F1-macro, and ROC AUC metrics. The experimental results indicate that the combination of SMOTE and Linear Discriminant Analysis (LDA) yields the highest overall performance, achieving an accuracy of 0.9773 and an F1-macro score of 0.9848. Logistic Regression demonstrates optimal results when paired with ROS, also reaching an accuracy of 0.9773. Artificial Neural Networks show the most substantial performance improvement under ADASYN, particularly reflected in higher F1-macro values. Although Support Vector Machine is less sensitive to oversampling interventions, it achieves a strong ROC AUC score of 0.9776 when trained using SMOTE. Overall, the findings confirm that oversampling techniques significantly improve classification performance in multilevel hypertension prediction, with SMOTE combined with LDA emerging as the most effective configuration
Improving the Accuracy of Obesity Classification Using a Stacking Classifier on Imbalanced Data with SMOTE
No full textOverweight continues to be a prevalent public health problem related to lifestyle behavior, eating behaviour and physical activity. The aim of this work is to develop a generalized and robust machine learning model having a high accuracy for categorizing obesity-level. The study applies to the Obesity Dataset with 1610 members and some preprocessing methods such selected data cleaning, categorical attributes transformation, train/test data set split and class imbalance under utilization of SMOTE approach. The modeling process is based on two base learners namely an optimized Random Forest and Gaussian Naïve Bayes that are fused by Stacking Classifier while using Logistic Regression as the meta-model. Experimental results show that the performance of stacking is the best where it obtains an accuracy rate of 86.34%, outperforming each single model. The analysis also reveals enhancements of various classification measures: stacking can indeed model complex non-linear dependencies between instances as well as simple linear ones. In general, the results serve to demonstrate that stacking-based ensemble learning is a strong solution for predicting obesity level and holds promise against early risk detection in preventive health care systems.Overweight continues to be a prevalent public health problem related to lifestyle behavior, eating behaviour and physical activity. The aim of this work is to develop a generalized and robust machine learning model having a high accuracy for categorizing obesity-level. The study applies to the Obesity Dataset with 1610 members and some preprocessing methods such selected data cleaning, categorical attributes transformation, train/test data set split and class imbalance under utilization of SMOTE approach. The modeling process is based on two base learners namely an optimized Random Forest and Gaussian Naïve Bayes that are fused by Stacking Classifier while using Logistic Regression as the meta-model. Experimental results show that the performance of stacking is the best where it obtains an accuracy rate of 86.34%, outperforming each single model. The analysis also reveals enhancements of various classification measures: stacking can indeed model complex non-linear dependencies between instances as well as simple linear ones. In general, the results serve to demonstrate that stacking-based ensemble learning is a strong solution for predicting obesity level and holds promise against early risk detection in preventive health care systems
Performance Evaluation of Face Mask Detection Using Feature Descriptor and Supervised Learning Method
No full textThe use of masks as a measure to prevent the spread of dangerous diseases such as COVID-19 and others has become a social norm. Manual detection is less effective, especially in areas with high mobility. This study develops and evaluates an artificial intelligence (AI)-based face mask detection system using feature description and machine learning models. An optimal and lightweight model can help hospitals implement face mask detection systems in areas prone to disease transmission. Image preprocessing, feature description, supervised learning model studies, and performance evaluation were conducted using accuracy, precision, recall, and F1-score metrics, and a confusion matrix was used to assess the overall model performance. The performance evaluation results show that the combination of the LBP feature description with the random forest model is the best choice, with a relatively high and stable accuracy of around 96.3% with an average value, precision, recall, and F1-score of around 96% using K-Fold Cross-Validation. These findings suggest that this method is helpful in detecting mask use while minimizing error and computation rates. This study contributes to the development of lightweight mask detection systems that can be used in real time
Experimental Comparison of Ground Plane Detection Speed Across Mobile Platforms
No full textMarkerless Augmented Reality (AR) technology has become increasingly important in various applications, yet its performance varies significantly across different platforms. This study conducts a comparative experimental analysis of ground plane detection performance between iOS and Android platforms using the Vuforia-based KreasiFurniture application. The research examines detection speed under varying lighting conditions (indoor and outdoor) and camera distances (50 cm, 100 cm, and 150 cm) through systematic testing with five repetitions per condition. Data were analyzed using Three-Way ANOVA with IBM SPSS Statistics 25. Results demonstrate that iOS achieves significantly faster and more consistent detection (mean = 1.402 seconds, SD = 0.143) compared to Android (mean = 1.541 seconds, SD = 0.235), with a statistically significant difference of 0.139 seconds (p = 0.003). The optimal detection distance was found at 100 cm for both platforms (p = 0.018). While lighting conditions showed no significant main effect (p = 0.129), a significant Platform × Light interaction (p = 0.038) was revealed, indicating that iOS maintains stable performance across lighting variations, whereas Android experiences substantial performance degradation in indoor conditions. These findings provide practical recommendations: iOS is preferable for applications requiring consistent indoor performance, 100 cm represents the optimal interaction distance for both platforms, and Android deployments should implement adaptive strategies for variable lighting conditions.Markerless Augmented Reality (AR) technology has become increasingly important in various applications, yet its performance varies significantly across different platforms. This study conducts a comparative experimental analysis of ground plane detection performance between iOS and Android platforms using the Vuforia-based KreasiFurniture application. The research examines detection speed under varying lighting conditions (indoor and outdoor) and camera distances (50 cm, 100 cm, and 150 cm) through systematic testing with five repetitions per condition. Data were analyzed using Three-Way ANOVA with IBM SPSS Statistics 25. Results demonstrate that iOS achieves significantly faster and more consistent detection (mean = 1.402 seconds, SD = 0.143) compared to Android (mean = 1.541 seconds, SD = 0.235), with a statistically significant difference of 0.139 seconds (p = 0.003). The optimal detection distance was found at 100 cm for both platforms (p = 0.018). While lighting conditions showed no significant main effect (p = 0.129), a significant Platform × Light interaction (p = 0.038) was revealed, indicating that iOS maintains stable performance across lighting variations, whereas Android experiences substantial performance degradation in indoor conditions. These findings provide practical recommendations: iOS is preferable for applications requiring consistent indoor performance, 100 cm represents the optimal interaction distance for both platforms, and Android deployments should implement adaptive strategies for variable lighting conditions
Design and Implementation of an IoT-Based Low-Emission Mobile Plastic Melting Machine for Sustainable Paving Block Production in Batam City
No full textPlastic waste accumulation poses a severe environmental burden, particularly in urban and archipelagic regions where centralized treatment infrastructure is limited. While thermal processing offers a pathway for volume reduction and material recovery, inadequate temperature control frequently leads to uncontrolled combustion and the formation of hazardous air pollutants. This study addresses this gap by developing and experimentally validating a low-emission, IoT-enabled mobile plastic melting system designed for decentralized paving block production. The proposed system integrates real-time thermal sensing using a K-type thermocouple and an ESP32-based controller with a compact three-nozzle water spray filtration unit. The control architecture maintains the melting process at approximately 270 °C, thereby preserving polymer viscosity for molding while preventing temperature excursions beyond 300 °C that may initiate combustion and toxic by-product formation. The filtration module operates as a simplified wet scrubber, capturing airborne particulates and simultaneously cooling the exhaust stream. Experimental evaluations confirm that the integrated control–filtration framework achieves stable thermal regulation and substantial suppression of visible exhaust emissions. Under these conditions, molten plastic was consistently transformed into dense paving blocks with smooth surface morphology and without evidence of polymer degradation or charring. The results demonstrate that combining IoT-based thermal governance with low-cost water-spray emission control provides an effective and scalable alternative to open burning for community-level plastic waste recycling. This mobile platform enables environmentally safer conversion of plastic waste into value-added construction materials, offering a practical pathway toward decentralized circular-economy implementation in resource-constrained regions
Association Rule Mining for Truck Body Damage Pattern Analysis Using Apriori and CRISP-DM
No full textThis study investigates damage patterns in truck body components by applying the Apriori association rule mining algorithm within the CRISP-DM framework. The analysis is based on 281 historical repair records from CV Lestari’s fleet throughout 2024. The dataset encompasses 14 attributes, including vehicle types, route categories, body materials, and load conditions. To ensure the robustness of the generated rules, parameter tuning was conducted using a grid search approach, resulting in minimum support and confidence thresholds of 15% and 60%, respectively. A total of 50 association rules were derived, with several rules demonstrating high confidence values and lift values above 1.1, indicating meaningful and non-random correlations. Notably, structural frame damage is strongly associated with mountainous routes and heavy loads, while door and hinge damage tends to occur in aluminum box-bodied trucks operating under medium loads. These patterns were aligned with practical insights from field technicians and further contextualized through technical recommendations, such as reinforcing high-stress points and adjusting inspection schedules for high-risk configurations. The findings support the formulation of predictive maintenance strategies, enabling companies to transition from reactive repairs to proactive, data-driven decision-making. By integrating rule-based insights into maintenance planning, the study contributes to reducing unexpected failures, optimizing inspection frequency, and enhancing overall fleet reliability
Sentiment Analysis of the Free Nutritious Meal Program (MBG) on Social Media X (Twitter) Using K-Nearest Neighbor and Artificial Neural Network
No full textThe Free Nutritious Meal Program (Makan Bergizi Gratis/MBG) is a national policy initiated by the Indonesian government to improve public nutritional status, particularly among children and vulnerable groups. Since its implementation, the program has generated extensive public discussion on social media, reflecting diverse opinions, support, and criticism. This study aims to analyze public sentiment toward the MBG program on social media X (Twitter) using machine learning-based text classification methods. A total of 9,038 Indonesian-language tweets were collected and processed through text preprocessing, semi-automatic sentiment labeling with manual validation, and feature extraction using the Term Frequency–Inverse Document Frequency (TF–IDF) method. Sentiments were classified into three categories: positive, neutral, and negative. The performance of K-Nearest Neighbor (KNN), Artificial Neural Network (ANN), and ANN with class balancing using Synthetic Minority Over-Sampling Technique (ANN + SMOTE) was evaluated using accuracy, precision, recall, and F1-score metrics supported by confusion matrix analysis. The results indicate that the ANN + SMOTE model achieved the highest performance with an accuracy of 93.58%, outperforming ANN (92.59%) and KNN (86.28%). The sentiment distribution indicates that public opinion toward the MBG program is predominantly neutral (52.1%), followed by positive (40.0%) and negative (7.9%) sentiments. These findings suggest that while the MBG program is generally well received, negative sentiments provide important feedback related to program implementation and governance.The Free Nutritious Meal Program (Makan Bergizi Gratis/MBG) is a national policy initiated by the Indonesian government to improve public nutritional status, particularly among children and vulnerable groups. Since its implementation, the program has generated extensive public discussion on social media, reflecting diverse opinions, support, and criticism. This study aims to analyze public sentiment toward the MBG program on social media X (Twitter) using machine learning-based text classification methods. A total of 9,038 Indonesian-language tweets were collected and processed through text preprocessing, semi-automatic sentiment labeling with manual validation, and feature extraction using the Term Frequency–Inverse Document Frequency (TF–IDF) method. Sentiments were classified into three categories: positive, neutral, and negative. The performance of K-Nearest Neighbor (KNN), Artificial Neural Network (ANN), and ANN with class balancing using Synthetic Minority Over-Sampling Technique (ANN + SMOTE) was evaluated using accuracy, precision, recall, and F1-score metrics supported by confusion matrix analysis. The results indicate that the ANN + SMOTE model achieved the highest performance with an accuracy of 93.58%, outperforming ANN (92.59%) and KNN (86.28%). The sentiment distribution indicates that public opinion toward the MBG program is predominantly neutral (52.1%), followed by positive (40.0%) and negative (7.9%) sentiments. These findings suggest that while the MBG program is generally well received, negative sentiments provide important feedback related to program implementation and governance
Analysis of the Impact of Violent Content on Social Media on Adolescent Cyberpsychology Using Support Vector Machine and Random Forest
No full textAdolescent exposure to violent content on social media has emerged as a critical issue due to its potential impact on mental health and cyberpsychological well-being. This study aims to classify multiple cyberpsychological impacts experienced by adolescents as a result of exposure to violent content on social media using a multi-label machine learning approach. A quantitative method was employed using self-reported data collected from 550 Indonesian adolescents aged 12–18 years through an online questionnaire. Psychological impacts were measured using adapted instruments from the Depression Anxiety Stress Scales (DASS-21) and cyberpsychology scales, then transformed into multi-label targets. Support Vector Machine (SVM) and Random Forest algorithms were implemented using a One-vs-Rest strategy. Model performance was evaluated using Hamming Loss, precision, recall, and Macro F1-score. The results indicate that SVM outperformed Random Forest with a Hamming Loss of 23.16% and a Macro F1-score of 0.42, particularly in predicting dominant labels such as anxiety and decreased self-confidence. However, both models showed limited performance in predicting minority labels such as depression and academic decline due to data imbalance. These findings highlight the importance of handling imbalanced data in cyberpsychology-based machine learning research and demonstrate the potential of multi-label classification in representing the complexity of psychological impacts of digital violence on adolescents.Adolescent exposure to violent content on social media has emerged as a critical issue due to its potential impact on mental health and cyberpsychological well-being. This study aims to classify multiple cyberpsychological impacts experienced by adolescents as a result of exposure to violent content on social media using a multi-label machine learning approach. A quantitative method was employed using self-reported data collected from 550 Indonesian adolescents aged 12–18 years through an online questionnaire. Psychological impacts were measured using adapted instruments from the Depression Anxiety Stress Scales (DASS-21) and cyberpsychology scales, then transformed into multi-label targets. Support Vector Machine (SVM) and Random Forest algorithms were implemented using a One-vs-Rest strategy. Model performance was evaluated using Hamming Loss, precision, recall, and Macro F1-score. The results indicate that SVM outperformed Random Forest with a Hamming Loss of 23.16% and a Macro F1-score of 0.42, particularly in predicting dominant labels such as anxiety and decreased self-confidence. However, both models showed limited performance in predicting minority labels such as depression and academic decline due to data imbalance. These findings highlight the importance of handling imbalanced data in cyberpsychology-based machine learning research and demonstrate the potential of multi-label classification in representing the complexity of psychological impacts of digital violence on adolescents
Evaluation of YOLOv8 and Faster R-CNN for Image-Based Food Detection
No full textDifficulties in manually tracking nutrition lead to the need for automatic food detection systems. However, Indonesian food presents tough challenges to recognize because similar-looking foods and different serving styles make it hard. This study looks at two deep learning models that follow different approaches: YOLOv8, which is known for being fast and efficient, and Faster R-CNN, which is known for being very accurate. The goal is to find the best model for use on mobile devices. This research uses a strict and standardized way to test the models to make sure the comparison is fair. A public dataset with 1,325 images from Roboflow was used. To deal with uneven class distribution, the images were split using Stratified Random Sampling. Before training, the images were resized using letterbox method to keep their original shape and normalized for pixel values. Both models were trained for the same number of epochs (100) and used the same optimizer (SGD) to ensure fair comparisons. The results show that YOLOv8 performs better in all areas. It achieved 88.6% mAP@50 accuracy and 62.0% mAP@50-95 precision. Faster R-CNN got 85.5% and 55.6% respectively. YOLOv8 also excels in sensitivity or Recall, reaching 87.7% compared to 61.7% for Faster R-CNN. The F1-Score, which balances accuracy and sensitivity, is 84.0% for YOLOv8 and 72% for Faster R-CNN. In terms of speed and size, YOLOv8 is much better. It runs in 13.5 ms and is 21.5 MB in size. That makes it 7.7 times faster and 7.3 times smaller than Faster R-CNN. Based on these results, YOLOv8 is the best choice for developing mobile-based nutrition tracking systems.Difficulties in manually tracking nutrition lead to the need for automatic food detection systems. However, Indonesian food presents tough challenges to recognize because similar-looking foods and different serving styles make it hard. This study looks at two deep learning models that follow different approaches: YOLOv8, which is known for being fast and efficient, and Faster R-CNN, which is known for being very accurate. The goal is to find the best model for use on mobile devices. This research uses a strict and standardized way to test the models to make sure the comparison is fair. A public dataset with 1,325 images from Roboflow was used. To deal with uneven class distribution, the images were split using Stratified Random Sampling. Before training, the images were resized using letterbox method to keep their original shape and normalized for pixel values. Both models were trained for the same number of epochs (100) and used the same optimizer (SGD) to ensure fair comparisons. The results show that YOLOv8 performs better in all areas. It achieved 88.6% mAP@50 accuracy and 62.0% mAP@50-95 precision. Faster R-CNN got 85.5% and 55.6% respectively. YOLOv8 also excels in sensitivity or Recall, reaching 87.7% compared to 61.7% for Faster R-CNN. The F1-Score, which balances accuracy and sensitivity, is 84.0% for YOLOv8 and 72% for Faster R-CNN. In terms of speed and size, YOLOv8 is much better. It runs in 13.5 ms and is 21.5 MB in size. That makes it 7.7 times faster and 7.3 times smaller than Faster R-CNN. Based on these results, YOLOv8 is the best choice for developing mobile-based nutrition tracking systems