Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
YOLOV12 Based on Stationary Vehicle for License Plate Detection
No full textThe use of technology for vehicle license plate recognition in this modern era is increasingly developing in supporting the needs of more effective transportation system management. This research aims to design and implement a vehicle license plate recognition system with the YOLOv12 (You Only Look Once) algorithm. The use of the YOLOv12 algorithm in license plate recognition is due to its superiority in detecting and recognizing objects in real-time with high accuracy. This research method will involve collecting a dataset of vehicle license plates from various viewing angles, lighting conditions, license plate colors, and the shape of the license plate. These datasets are then used to train an adapted YOLOv12 model to detect and recognize characters on license plates. Tests are conducted by measuring the detection accuracy, processing speed, and robustness of the detection system to disturbances such as noise and variations in environmental conditions when detecting license plates. The results of the study shown that this system yielded accuracy rate of 97.5%, recall of 95.4%, precision of 96.7%, and is capable of recognizing characters on vehicle license plates with an accuracy rate of 88%, recall of 87%, and precision of 85.8%. The average processing time is 1 second per image on CPU and 20 seconds per image on GPU. The system\u27s ability to detect vehicle license plates shows that the YOLOv12 algorithm can be used for large-scale vehicle license plate system implementation. The significance of these results lies in their potential application in various fields such as parking management systems, traffic management, and law enforcement, which can improve efficiency and safety
Classification of Foot Wound Severity in Type 2 Diabetes Mellitus Patients Using MobileNetV2-Based Convolutional Neural Network
No full textDiabetic Foot Ulcer (DFU) is a serious complication in Type 2 Diabetes Mellitus patients that may lead to amputation if not properly treated. This study employs the MobileNetV2 architecture based on Convolutional Neural Network (CNN) to classify DFU severity into two categories: severe and non-severe. The dataset consists of 1,000 images, divided into 70% training, 20% validation, and 10% testing. Data preprocessing was performed using normalization, augmentation (rotation, flipping, zooming), and dataset balancing to enhance model generalization. The model was trained for 10 epochs with a batch size of 32, learning rate of 0.001, and Adam optimizer. Experimental results show 98% accuracy on validation data with an average precision, recall, and F1-score of 0.98. On the testing stage, the model achieved 94% accuracy with an average precision, recall, and F1-score of 0.94. The confusion matrix also indicates strong performance in distinguishing both classes. This study demonstrates that MobileNetV2-based CNN with proper preprocessing and hyperparameter settings can serve as an effective supporting method for early DFU severity classification, thereby improving the speed and accuracy of medical decision-making
Performance of Multivariate Missing Data Imputation Methods on Climate Data
No full textClimate data plays an important role in various aspects of life. However, missing data is often found, which can interfere with data processing and reduce the quality of analysis. Therefore, appropriate handling methods are needed to ensure that the analysis results remain valid. This study aims to compare the performance of several imputation methods for missing multivariate data based on the identification of actual missing data patterns, and to determine the appropriate imputation method based on the mechanism of missing data. This study also aims to apply the best method to data with actual missing data patterns to assess its effect on descriptive statistical changes required for further climatological analysis. The methods used include monthly averages, missRanger, k-Nearest Neighbor (k-NN), and Iterative Robust-Model Imputation (IRMI). The missing data information was obtained from Global Surface Summary of the Day (GSOD) data, namely temperature, precipitation, humidity, pressure, and wind speed variables with a daily frequency for 11 years, with a missing data proportion of 11.4%. The missing data patterns were then applied to relatively complete NASA Power data to evaluate the imputation results. The results show that IRMI is less capable of handling extreme missing data conditions, namely 17 completely missing rows. In contrast, k-NN, missRanger, and monthly averages provided better results in both extreme and non-extreme conditions. Of the four methods, monthly averages were chosen because they were able to overcome missing data while maintaining multivariate structure with 58% on sMAPE and 2.64% on relative difference.Climate data plays an important role in various aspects of life. However, missing data is often found, which can interfere with data processing and reduce the quality of analysis. Therefore, appropriate handling methods are needed to ensure that the analysis results remain valid. This study aims to compare the performance of several imputation methods for missing multivariate data based on the identification of actual missing data patterns, and to determine the appropriate imputation method based on the mechanism of missing data. This study also aims to apply the best method to data with actual missing data patterns to assess its effect on descriptive statistical changes required for further climatological analysis. The methods used include monthly averages, missRanger, k-Nearest Neighbor (k-NN), and Iterative Robust-Model Imputation (IRMI). The missing data information was obtained from Global Surface Summary of the Day (GSOD) data, namely temperature, precipitation, humidity, pressure, and wind speed variables with a daily frequency for 11 years, with a missing data proportion of 11.4%. The missing data patterns were then applied to relatively complete NASA Power data to evaluate the imputation results. The results show that IRMI is less capable of handling extreme missing data conditions, namely 17 completely missing rows. In contrast, k-NN, missRanger, and monthly averages provided better results in both extreme and non-extreme conditions. Of the four methods, monthly averages were chosen because they were able to overcome missing data while maintaining multivariate structure with 58% on sMAPE and 2.64% on relative difference
Comparison of FEM-LSDV Panel Regression with Classical Panel Regression Models in Analyzing Economic Growth in Indonesia
No full textThis study evaluates the performance of multiple panel regression approaches in modeling the determinants of regional economic growth in Indonesia. It specifically compares three classical panel models: the Common Effect Model (CEM), the Random Effect Model (REM), and the Fixed Effect Model (FEM), alongside the Fixed Effect Model with the Least Squares Dummy Variable (FEM LSDV) approach. The analysis is based on panel data covering 34 provinces from 2019 to 2023, using key macroeconomic indicators such as inflation, investment, exports, money supply, open unemployment rate, and participation in the national health insurance program (JKN). The models are assessed using formal statistical tests, including the Chow and Hausman tests, and evaluated through performance metrics such as RMSE, AIC, and R-squared. The results show that the FEM LSDV model offers the best performance, with an R-squared value of 0.7039, RMSE of 0.5442, and an AIC of 365.55. Notably, the model identifies North Maluku Province as contributing positively and significantly to economic growth, while the year 2020 shows a significant negative impact, likely due to the economic disruptions caused by the COVID-19 pandemic. These findings demonstrate the effectiveness of the FEM LSDV approach in capturing both spatial and temporal heterogeneity in regional economic analysis and support its application in policy-oriented research.This study evaluates the performance of multiple panel regression approaches in modeling the determinants of regional economic growth in Indonesia. It specifically compares three classical panel models: the Common Effect Model (CEM), the Random Effect Model (REM), and the Fixed Effect Model (FEM), alongside the Fixed Effect Model with the Least Squares Dummy Variable (FEM LSDV) approach. The analysis is based on panel data covering 34 provinces from 2019 to 2023, using key macroeconomic indicators such as inflation, investment, exports, money supply, open unemployment rate, and participation in the national health insurance program (JKN). The models are assessed using formal statistical tests, including the Chow and Hausman tests, and evaluated through performance metrics such as RMSE, AIC, and R-squared. The results show that the FEM LSDV model offers the best performance, with an R-squared value of 0.7039, RMSE of 0.5442, and an AIC of 365.55. Notably, the model identifies North Maluku Province as contributing positively and significantly to economic growth, while the year 2020 shows a significant negative impact, likely due to the economic disruptions caused by the COVID-19 pandemic. These findings demonstrate the effectiveness of the FEM LSDV approach in capturing both spatial and temporal heterogeneity in regional economic analysis and support its application in policy-oriented research
The Effect of Corporate Governance, Funding Decisions and Working Capital Turnover on Firm Value
No full textThis study aims to examine the effect of corporate governance, funding decisions, and working capital turnover on firm value in the consumer non-cyclical sector listed on the Indonesia Stock Exchange (IDX) during the period 2019–2023. The population consists of 125 companies, from which 69 were selected as the final sample using purposive sampling. The data were analyzed using panel data regression in EViews 12. Firm value in this study is measured using the Price-to-Book Value (PBV) ratio. Corporate governance is assessed through the ASEAN Corporate Governance Scorecard (ACGS); funding decisions are proxied by the Debt-to-Asset Ratio (DAR); and working capital turnover is calculated as net sales divided by net working capital. The results show that, in part, only funding decisions have a significant effect on firm value, while corporate governance and working capital turnover do not. The model is statistically valid as shown by the F-test, but the coefficient of determination (R²) is relatively low, at 7.47%, indicating that the explanatory variables account for only a small portion of the variation in firm value
A Roasted Coffee Bean Identification Using ResNet50 Model
No full textIdentification of coffee types after roasting is a major challenge because visual changes make the appearance of coffee beans diverse. Subjective assessment methods are time-consuming, so digital image processing and CNN techniques show potential to solve complex classification problems. This study develops a ResNet50-based CNN model to identify four types of coffee beans (Robusta, Arabica, Excelsa, and Liberica) after roasting and analyzes the effectiveness of pre-processing and augmentation techniques in improving classification performance. The research employed quantitative methodology with three phases: data collection, pre-processing with augmentation, and CNN implementation. The dataset consisted of 2,000 coffee bean images, with 500 images for each class: Arabica, Excelsa, Liberica, and Robusta, ensuring balanced representation across all coffee varieties from a local Indonesian coffee supplier, using smartphone. Preprocessing included normalization and resizing, while augmentation comprised various image transformation techniques. Model performance was evaluated using performance metrics. Results showed an overall accuracy of 94.50%, with Liberica demonstrating exceptional performance (100% precision, 98% recall). Robusta achieved 97% precision and 98% recall, while Arabica showed 86.5% precision with 96% recall. Excelsa achieved 95.6% precision and 86% recall. The model successfully classified 378 out of 400 test samples, with Excelsa representing the primary classification challenge due to visual similarity with other varieties post-roasting. Analysis of misclassifications revealed improved distinction between coffee varieties, with the model demonstrating strong generalization capabilities across all classes. The ResNet50 model successfully identified coffee beans with good accuracy but experienced difficulty distinguishing varieties with similar visual characteristics. Future work should explore improved methods and larger datasets for accuracy
A Two-Stage Braille Recognition System Using YOLOv8 for Detection and CNN for Classification
No full textAutomatic recognition of Braille characters remains a challenge in the field of computer vision, especially due to variations in shape, size, and lighting conditions in images. This research proposes a two-stage system to detect and recognize Braille letters in real time using a deep learning approach. In the first stage, the YOLOv8 model is used to detect the position of Braille characters within an image. The detected regions are then processed in the second stage using a classification model based on the MobileNetV2 CNN architecture. The dataset used consists of 7,016 Braille character images, collected from a combination of the AEyeAlliance dataset and annotated data from Roboflow. To address the class imbalance problem—particularly for letters T to Z which had fewer samples—oversampling and image augmentation techniques were applied that makes the final combined dataset contained approximately 7,616 images. The system was tested on 1,513 images and achieved strong results, with average precision, recall, and F1-score of 0.98, and an overall accuracy of 98%. This two-stage method effectively separates detection and classification tasks, resulting in an efficient and accurate Braille recognition system suitable for real-time applications
Segmentation of Generation Z Spending Habits Using the K-Means Clustering Algorithm: An Empirical Study on Financial Behavior Patterns
No full textGeneration Z, born between 1997 and 2012, exhibits unique consumption behaviors shaped by digital technology, modern lifestyles, and evolving financial decision-making patterns. This study segments their financial behavior using the K-Means clustering algorithm applied to the “Generation Z Money Spending” dataset from Kaggle. In addition to K-Means, alternative clustering algorithms—K-Medoids and Hierarchical Clustering—are evaluated to compare their effectiveness in identifying behavioral patterns. The dataset consists of 1,700 individuals with 15 numerical spending attributes, including rent, food, entertainment, education, savings, and investments. All data were normalized using Min-Max Scaling prior to clustering. The analysis identifies six distinct clusters, ranging from highly consumption-oriented groups (with higher spending on entertainment and online shopping) to financially conscious groups prioritizing savings and investments. A quantitative approach was used, incorporating exploratory data analysis, correlation testing, and the Elbow Method to determine the optimal number of clusters. The optimal cluster count of six is supported by a Davies-Bouldin Index (DBI) score of 2.412, indicating acceptable but improvable cluster separation. Each cluster displays unique characteristics: Cluster 0 (average age 20.6) focuses on savings and investments with moderate essential spending; Cluster 1 (average age 23.6) prioritizes education and higher rent expenses; Cluster 2 (average age 20.3) is digitally oriented, spending more on online shopping and entertainment; Cluster 3 (average age 25.2) demonstrates financial stability with balanced expenditures; Cluster 4 (average age 24.9) emphasizes savings and investments with moderate living costs; and Cluster 5 (average age 24.96) combines strong saving habits with balanced essential and leisure spending. Model performance was assessed using the Davies-Bouldin Index, Silhouette Score, and Calinski-Harabasz Index to ensure comprehensive evaluation of cluster quality. The findings highlight the diverse spending behaviors of Generation Z, offering valuable insights for businesses, policymakers, and financial service providers to develop targeted strategies aligned with each segment’s characteristics.Generation Z, born between 1997 and 2012, exhibits unique consumption behaviors shaped by digital technology, modern lifestyles, and evolving financial decision-making patterns. This study segments their financial behavior using the K-Means clustering algorithm applied to the “Generation Z Money Spending” dataset from Kaggle. In addition to K-Means, alternative clustering algorithms—K-Medoids and Hierarchical Clustering—are evaluated to compare their effectiveness in identifying behavioral patterns. The dataset consists of 1,700 individuals with 15 numerical spending attributes, including rent, food, entertainment, education, savings, and investments. All data were normalized using Min-Max Scaling prior to clustering. The analysis identifies six distinct clusters, ranging from highly consumption-oriented groups (with higher spending on entertainment and online shopping) to financially conscious groups prioritizing savings and investments. A quantitative approach was used, incorporating exploratory data analysis, correlation testing, and the Elbow Method to determine the optimal number of clusters. The optimal cluster count of six is supported by a Davies-Bouldin Index (DBI) score of 2.412, indicating acceptable but improvable cluster separation. Each cluster displays unique characteristics: Cluster 0 (average age 20.6) focuses on savings and investments with moderate essential spending; Cluster 1 (average age 23.6) prioritizes education and higher rent expenses; Cluster 2 (average age 20.3) is digitally oriented, spending more on online shopping and entertainment; Cluster 3 (average age 25.2) demonstrates financial stability with balanced expenditures; Cluster 4 (average age 24.9) emphasizes savings and investments with moderate living costs; and Cluster 5 (average age 24.96) combines strong saving habits with balanced essential and leisure spending. Model performance was assessed using the Davies-Bouldin Index, Silhouette Score, and Calinski-Harabasz Index to ensure comprehensive evaluation of cluster quality. The findings highlight the diverse spending behaviors of Generation Z, offering valuable insights for businesses, policymakers, and financial service providers to develop targeted strategies aligned with each segment’s characteristics
Vision Transformer for Pneumonia Classification with Grad-CAM Explainability
No full textPneumonia is still one of the main causes of death around the world, especially in kids and older people. To lower the death rate, early and accurate diagnosis is very important. Chest X-ray (CXR) imaging is widely used for this purpose, but manual reading of CXR images can be time-consuming and may lead to differences in interpretation between observers. To address this problem, this study presents a pneumonia classification model based on the Vision Transformer (ViT) architecture combined with Gradient-weighted Class Activation Mapping (Grad-CAM) to make the model’s decisions more interpretable. The model was trained on a publicly available CXR dataset with 5,863 images that were split into Normal and Pneumonia classes, using a 70:15:15 split for training, validation, and testing. The ViT model achieves an accuracy of 96.41% on the test set and a high recall for pneumonia cases, while class weighted loss helps to maintain more balanced predictions between the two classes. The Area Under the Curve (AUC) of 0.975 indicates strong discrimination between pneumonia-positive and normal samples. Grad-CAM visualizations, supported by a randomization test and occlusion analysis, provide an initial qualitative view of the lung regions that influence the model’s predictions and often overlap with radiologically plausible areas. However, the heatmaps have not been formally evaluated by radiologists, and the correspondence between highlighted regions and pneumonia consolidation patterns has not yet been quantitatively validated. Therefore, the proposed ViT Grad-CAM framework should be regarded as an exploratory step toward explainable pneumonia classification on chest X-rays rather than a system that is ready for clinical deployment
Bi-LSTM with Explainable AI for Session Duration-Based Customer Lifetime Value Proxy on Multi-Category E-Commerce Platforms
No full textThe rapid growth of multi-category e-commerce platforms has increased the importance of behavioral data for predicting Customer Lifetime Value (CLV). However, monetary-based CLV estimation is often infeasible due to incomplete or unavailable transaction records. This study adopts session duration as a short-term behavioral proxy for CLV and proposes a Bidirectional Long Short-Term Memory (Bi-LSTM) model enhanced with a Temporal Attention mechanism to improve predictive accuracy. The publicly available REES46 dataset, consisting of 1,6 million events and 276.000 unique sessions, is used with preprocessing steps including label encoding, temporal feature construction, and outlier-aware sampling to address the highly right-skewed distribution of session durations. Four baseline models Decision Tree, Random Forest, Extreme Gradient Boosting (XGBoost), and conventional Long Short-Term Memory (LSTM) are implemented for comparative evaluation. The baseline LSTM achieves MAE = 0,0080 and RMSE = 0,0322. The proposed Bi-LSTM v3 model, equipped with Temporal Attention and structured sampling, demonstrates substantial performance improvement, achieving MAE = 0,0043 (≈368 seconds) and RMSE = 0,0172 (≈1466 seconds), representing an accuracy gain of approximately 45–50% over the baseline. Explainability analysis using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) confirms that the time_diff feature is the dominant contributor at both global and local levels, aligning with the behavior of the attention mechanism. Additionally, the integration of Explainable Artificial Intelligence (XAI) provides transparent insights into model decision patterns. These findings show that combining Bi-LSTM, Temporal Attention, and XAI yields an accurate and interpretable framework for session duration prediction, supporting the use of session duration as a feasible CLV proxy in multi-category e-commerce environments.The rapid growth of multi-category e-commerce platforms has increased the importance of behavioral data for predicting Customer Lifetime Value (CLV). However, monetary-based CLV estimation is often infeasible due to incomplete or unavailable transaction records. This study adopts session duration as a short-term behavioral proxy for CLV and proposes a Bidirectional Long Short-Term Memory (Bi-LSTM) model enhanced with a Temporal Attention mechanism to improve predictive accuracy. The publicly available REES46 dataset, consisting of 1,6 million events and 276.000 unique sessions, is used with preprocessing steps including label encoding, temporal feature construction, and outlier-aware sampling to address the highly right-skewed distribution of session durations. Four baseline models Decision Tree, Random Forest, Extreme Gradient Boosting (XGBoost), and conventional Long Short-Term Memory (LSTM) are implemented for comparative evaluation. The baseline LSTM achieves MAE = 0,0080 and RMSE = 0,0322. The proposed Bi-LSTM v3 model, equipped with Temporal Attention and structured sampling, demonstrates substantial performance improvement, achieving MAE = 0,0043 (≈368 seconds) and RMSE = 0,0172 (≈1466 seconds), representing an accuracy gain of approximately 45–50% over the baseline. Explainability analysis using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) confirms that the time_diff feature is the dominant contributor at both global and local levels, aligning with the behavior of the attention mechanism. Additionally, the integration of Explainable Artificial Intelligence (XAI) provides transparent insights into model decision patterns. These findings show that combining Bi-LSTM, Temporal Attention, and XAI yields an accurate and interpretable framework for session duration prediction, supporting the use of session duration as a feasible CLV proxy in multi-category e-commerce environments