Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Comparative Analysis of ResNet-Based Wagner-Scale Classification for Imbalanced DFU Data
No full textDiabetic Foot Ulcers (DFU) are a serious complication of diabetes mellitus and carry a high risk of lower extremity amputation if not treated in a timely manner. The conventional classification process, which relies on visual inspection by clinicians, tends to be subjective and inconsistent. Therefore, this study proposes a multiclass classification model for DFU based on the Wagner Scale (Grades 0–5) using the ResNet-50 architecture with a transfer learning approach as the core machine learning method. The dataset used in this study consists of 1,415 clinical wound images that were annotated and verified by medical professionals. The dataset is highly imbalanced, with 543 images in Grade 0, 110 in Grade 1, 252 in Grade 2, 145 in Grade 3, 293 in Grade 4, and only 72 images in Grade 5. To address this imbalance, random oversampling (ROS) was applied, in addition to standard preprocessing techniques such as normalization and data augmentation to increase training data diversity.Experimental results demonstrate that the proposed model achieves high classification performance based on accuracy, precision, recall, and F1-score. Specifically, the model obtained a precision of 0.96, recall of 0.95, and F1-score of 0.95, indicating consistent and robust classification performance across all Wagner grades. The best configuration (ResNet-50 + ROS) successfully improved the classification performance across minority grades (e.g., Grade 1 and Grade 5). Moreover, the model consistently identifies minority classes and does not exhibit signs of overfitting. Model optimization using the Adam optimizer and data balancing strategies significantly improves the generalization capability of the classifier. These findings indicate that the proposed model is not only effective for automatic DFU classification, but also has great potential to support objective clinical decision making and accelerate diagnosis, particularly in healthcare facilities with limited resources
Impact of XBRL Technology Implementation on Information Asymmetry in Indonesia’s Capital Market
No full textThis research aims to examine whether the publication of financial statements in XBRL format could reduce the level of information asymmetry, measured by the bid-ask spread, in Indonesia’s capital market over eight years of its implementation. Furthermore, this research examines differences in the level of information asymmetry in two observation periods, which are the early stage and the advanced stage of the XBRL implementation. The population in this study are listed companies on the IDX80 index which were sampled using the purposive method. The analytical instrument used is a panel data regression test using a random effect model and the non-parametric Wilcoxon test of statistical differences. Consistent with similar studies, the results show that the publication of financial reports in XBRL format could reduce the level of information asymmetry by providing accurate, integrated, and universally accessible reporting. The difference test further reveals that the level of information asymmetry is lower in the advanced stages compared to the early stages. This suggest that XBRL implementation becomes more effective over time due to the positive developments in institutional readiness and stakeholder facilitation
Identifying Rice Plant Damage Due to Pest Attacks Using Convolutional Neural Networks
No full textRice (Oryza Sativa) is an important crop for meeting global food needs; however, one of the main challenges in its cultivation is the attack of stem borer pests, which can cause significant damage. This study aims to identify the damage caused by these pest attacks using Convolutional Neural Networks (CNN) methods. We developed and trained several CNN architectures, including the proposed architecture, MobileNet, and EfficientNetB0, to detect pest attacks on rice. The dataset used consists of 700 images per class taken directly from the field, where the images depict rice plants that have been peeled or opened to inspect for the presence of pests, specifically stem borer pests. To enhance the quality and diversity of the dataset, we applied a rigorous selection process, ensuring that only high-quality images were used. Additionally, augmentation techniques such as rotation were employed to expand the dataset to 2000 images per class. Labeling was carried out carefully to ensure that each image accurately reflected the condition of the pest attack. The results of the study indicate that the proposed CNN model can identify damage with high accuracy, thereby contributing to efforts to increase rice production through early detection of pest attacks using computer vision technology
Implementation of Generative Language Models (GLM) in Cyber Exercise Secure Coding using Prompt Engineering
No full textWith the advancement of technology, the need for secure software is becoming increasingly urgent due to the rise in vulnerabilities in applications. In 2022, the National Cyber and Encryption Agency (BSSN) recorded 2,348 cases of web defacement, with one of the main causes being the lack of attention to secure coding practices during software development. This study explores the utilization of Generative Language Models (GLMs), such as ChatGPT, in secure coding training to enhance developers' skills. GLMs were implemented in a cybersecurity platform designed specifically for secure coding training, also serving as learning assistants that users can interact with during the cyber exercise. The study results show that the cyber exercise using GLMs significantly improved users' secure coding skills, as evidenced by comparing pre-test and post-test scores, indicating an increase in knowledge and proficiency in secure coding practices
Enhanced Heart Disease Diagnosis Using Machine Learning Algorithms: A Comparison of Feature Selection
No full textHeart disease or cardiovascular disease is one of the leading causes of death in the world. Based on WHO data, in 2019, as many as 17.9 million people died from cardiovascular disease. If early prevention is not carried out immediately, of course, the victims will increase every year. Therefore, with the increasingly rapid development of technology, especially in the health sector, it is hoped that it can help medical personnel in treating patients suffering from various diseases, especially heart disease. So in this study, it will be more focused on the selection of relevant features or attributes to increase the accuracy value of the Machine Learning algorithm. The algorithms used include Random Forest and SVM. Meanwhile, for feature selection, several feature selection techniques are used, including information gain (IG), Chi-square (Chi2) and correlation feature selection (CFS). The use of these three techniques aims to obtain the main features so that they can minimize irrelevant features that can slow down the machine process. Based on the results of the experiment with a comparison of 70:30, it shows that CFS-SVM is superior by using nine features, which obtain the highest accuracy of 92.19%, while CFS-RF obtains the best value with eight features of 91.88%. By using feature selection and hyperparameter techniques, SVM obtained an increase of 10.88%, and RF obtained an increase of 9.47%. Based on the performance of the model using the selected relevant features, it shows that the proposed CFS-SVM shows good and efficient performance in diagnosing heart disease
Customer Satisfaction Evaluation in Online Food Delivery Services: A Systematic Literature Review
No full textThe rapid growth of online food delivery services has heightened the need for effective customer satisfaction measurement. This systematic literature review examines 476 papers, selecting 15 key studies to identify prevailing evaluation approaches. Findings reveal that sentiment analysis and PLS-SEM are the most frequently used analytical methods, each appearing in six studies. Satisfaction measurement relies on sentiment polarity scores in five studies and SERVQUAL frameworks in three studies. Data collection primarily involves surveys in seven studies and user-generated content in six studies, but limited demographic diversity reduces generalizability. Three key future research directions emerge. Advanced analytical techniques appear in 5 of 11 future works in the analysis methods domain. Expanding evaluation metrics is mentioned in 6 of 12 proposals in the evaluation domain. Exploring demographic context is highlighted in 10 of 25 recommendations in the dataset’s domain, with dataset development receiving twice the attention of methodological advancements. These results provide researchers with a structured framework for customer satisfaction evaluation while guiding food delivery platforms in refining service quality. By systematically mapping current methodologies and future priorities, this study bridges gaps between academia and industry, ensuring more effective customer satisfaction assessments
Sentiment Analysis of ChatGPT on Indonesian Text using Hybrid CNN and Bi-LSTM
No full textThis study explores sentiment analysis on Indonesian text using a hybrid deep learning approach that combines Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM). Due to the complex linguistic structure of the Indonesian language, sentiment classification remains challenging, necessitating advanced methods to capture both local patterns and sequential dependencies. The primary objective of this research is to improve sentiment classification accuracy by leveraging a hybrid model that integrates CNN for feature extraction and Bi-LSTM for contextual understanding. The dataset consists of 800 manually labeled samples collected from social media platforms, preprocessed using case folding, stop word removal, and lemmatization. Word embeddings are generated using the Word2Vec CBOW model, and the classification model is trained using a hybrid architecture. The best performance was achieved with 32 Bi-LSTM units, a dropout rate 0.5, and L2 regularization, which was evaluated using Stratified K-Fold cross-validation. Experimental results demonstrate that the hybrid model outperforms conventional deep learning approaches, achieving 95.24% accuracy, 95.09% precision, 95.15% recall, and 95.99% F1 score. These findings highlight the effectiveness of hybrid architectures in sentiment analysis for low-resource languages. Future work may explore larger datasets or transfer learning to enhance generalizability
Automated Indonesian Plate Recognition: YOLOv8 Detection and TensorFlow-CNN Character Classification
Full text linkThe precise identification and reading of Indonesian vehicle number plates are important in many areas, including the enforcement of law, collection of charges, management of parking areas, and safety measures. This study integrates the implementation of the YOLOv8 object detection algorithm with three OCR methods: EasyOCR, TesseractOCR, and TensorFlow. YOLOv8 is capable of identifying license plates from images and videos at a high speed and reliability under different conditions and therefore is used in this study to perform plate detection in images and videos. After licenses are detected, OCR techniques are performed to segment and read the letters. Both EasyOCR and TesseractOCR performed moderately well on static images achieving accuracy rates of 70% and 68% respectively, but both suffered significantly lower performance in video scenarios. Of the 100 video frames, EasyOCR was able to correctly identify characters in 61 frames and TesseractOCR in 58 frames, while the TensorFlow-based model outperformed the other two with 75 correct recognitions. Furthermore, easy OCR and static images as input while the TensorFlow-based models completed them with 100% accuracy. This observation can be explained by its design, which utilizes a CNN with ReLU activation and Softmax outputs, trained on 10,261 annotated characters and was enhanced with five different data augmentation techniques. The model shows strong performance in its ability to handle dynamic conditions such as motion blur, changing light conditions, and rotation of the plate angle. The results underscore the drawbacks of one-size-fits-all OCR applications in real-world use cases and stress the need for bespoke model training, as well as hierarchical contouring, in the context of automatic license plate recognition (ALPR). This study provides additional insights into ALPR systems by delivering a robust, scalable, and real-time tool for plate and character recognition, which is essential for intelligent transportation systems
Expertise Retrieval Using Adjusted TF-IDF and Keyword Mapping to ACM Classification Terms
Full text linkIn an era of collaboration, knowing someone's expertise is becoming increasingly necessary. Recognizing individuals' proficiency can be challenging because it requires considerable manual time. This study explores the expertise of lecturers from the Computer Science Department, Universitas Indonesia (Fasilkom UI), based on scientific publications. The data were obtained from the Sinta journal website’s scrapping process, which includes Scopus, Garuda, and Google Scholar data sources. The approach used was keyword extraction using the adjusted TF-IDF. The resulting keywords were then mapped to the ACM classification class using cosine similarity calculations with various embedding models, including BERT, BERT multilingual, FastText, XLM Roberta, and SBERT. The experimental results highlighted that combining the adjusted TF-IDF with mapping to the ACM classes using SBERT is a promising approach for gaining the best expertise. The use of abstract data has proved to be better than that of full-text data. Using title-abstract-EN data achieved a score of 0.49 for both the P@1 and NDCG@1 metrics, whereas the title-abstract-ENID data attained a score of 0.75 for both metrics P@1 and NDCG@1
Enhancing Tomato Leaf Disease Detection via Optimized VGG16 and Transfer Learning Techniques
Full text linkIdentification of tomato leaf disease remains difficult because standard approaches are frequently incorrect in identifying distinct signs. Convolutional Neural Networks (CNNs) perform well in image classification and pattern identification, although they are prone to overfitting. Thus, max pooling was employed to reduce dimensionality while retaining crucial information. This paper offers an improved CNN through hyperparameter tuning and compares it to Transfer Learning models such as InceptionV3, NASNetMobile, and VGG16, which were chosen for their efficiency and accuracy. The dataset comprises 7,178 photos classified as Healthy, Leaf Late Blight, Septoria Leaf Spot, and Yellow Leaf Curl Virus, collected from Kaggle.. The dataset is separated into three sections: training, validation, and testing, with a ratio of 70:15:15. The results of this study revealed that the proposed method achieved the highest accuracy of 98.24%. In the application of transfer learning, the inceptionV3 model achieved an accuracy of 96.94%, whereas NASNetMobile obtained 97.50%, and VGG16 showed an accuracy of 96.76%. The evaluation is based on accuracy, precision, recall, F1-score and Inference time to determine the optimum model for accuracy and computing efficiency. This project uses the proposed method and Transfer Learning Techniques to categorize illness images on tomato leaves. These findings will drive further research to improve tehe performance of the proposed method for foliar disease classification and comparable applications