Muhammad Alfhi Saputra

Muhammad Alfhi Saputra

Lecturer & Doctoral Student

Publications

You can also find my articles on my Google Scholar profile.

Journal Articles

Recognizing Indonesian sign language (Bisindo) gesture in complex backgrounds

Published in Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), 2024

Sign language, particularly Indonesian sign language (Bisindo), is vital for deaf individuals, but learning it is challenging. This study aims to develop an automated Bisindo recognition system suitable for diverse backgrounds. Previous research focused on greenscreen backgrounds and struggled with natural or complex backgrounds. To address this problem, the study proposes using Faster region-based convolutional neural networks (RCNN) and YOLOv5 for hand and face detection, MobileNetV2 for feature extraction, and long short-term memory (LSTM) for classification. The system is also designed to focus on computational efficiency. YOLOv5 model achieves the best result with a sentence accuracy (SAcc) of 49.29% and a word error rate (WER) of 16.42%, with a computational time of 0.0188 seconds, surpassing the baseline model. Additionally, the system achieved a SacreBLEU score of 67.77%, demonstrating its effectiveness in Bisindo recognition across various backgrounds. This research improves accessibility for deaf individuals by advancing automated sign language recognition technology.

Recommended citation: Saputra, M. A., & Rakun, E. (2024). Recognizing Indonesian sign language (Bisindo) gesture in complex backgrounds. Indonesian Journal of Electrical Engineering and Computer Science, 36(3), 1583-1593. https://doi.org/10.11591/ijeecs.v36.i3.pp1583-1593
Download Paper | Download Slides

Peringkas Teks Otomatis Bahasa Indonesia secara Abstraktif Menggunakan Metode Long Short-Term Memory

Published in eProceedings of Engineering, 2021

Salah satu topik dalam bidang Natural Language Processing (NLP) yang cukup menantang adalah peringkas teks otomatis. Dalam praktiknya peringkas teks otomatis terbagi menjadi dua pendekatan, yaitu ekstraktif dan abstraktif. Pendekatan abstraktif dinilai lebih baik karena cara kerjanya mendekati cara kerja manusia ketika meringkas teks atau yang disebut parafrase. Metode yang digunakan pada penelitian ini adalah Long Short-Term Memory (LSTM) yang mana metode tersebut telah sukses melakukan peringkasan dalam Bahasa Inggris. Dataset yang digunakan adalah kumpulan artikel berita media daring Bahasa Indonesia. Hasil terbaik yang didapatkan pada pengujian dengan metode LSTM menggunakan metode evaluasi ROUGE-1 adalah 0.13846. Kata kunci: peringkas teks otomatis, abstraktif, Bahasa Indonesia, long short-term memory, ROUGE Abstract One topic about natural language processing that is quite challenging is automatic text summarization. Automatic-text-summarization is practically divided into two kinds of approach, namely extractive and abstractive. Abstractive-approach is considered better since it resembles how humans work in terms of text summarizing or paraphrasing. A method used in this study is Long Short-Term Memory (LSTM) which has succeeded to summarize texts in English. Datasets that have been used are a number of online news articles in Bahasa Indonesia. The best result gained using LSTM based on the ROUGE-1 evaluation is 0.13846.

Recommended citation: Saputra, M. A., & Al Maki, W. F. (2021). Peringkas Teks Otomatis Bahasa Indonesia secara Abstraktif Menggunakan Metode Long Short-Term Memory. eProceedings of Engineering, 8(2).
Download Paper | Download Slides

Conference Papers

Evaluating the Adversarial Vulnerability of YOLOv11 for Indonesian Sign Language (BISINDO) Recognition

Published in 2025 8th International Conference on Informatics and Computational Sciences (ICICoS), 2025

Indonesian Sign Language (BISINDO) is a common form of communication used by the deaf community in Indonesia. Various deep learning models have been successfully utilized to recognize BISINDO gestures, including object detection models such as YOLO. However, these models remain vulnerable to adversarial attacks. This study aims to evaluate the robustness of the YOLOv11 model against three types of adversarial attacks, namely the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Basic Iterative Method (BIM). We utilize the BISINDO sign image dataset, which consists of 26 alphabet gesture classes (A to Z). The YOLOv11 model achieved a baseline performance of 0.945 mAP on clean data. The model was tested on adversarial images, and the results show a significant mAP drop of 49% under PGD and BIM attacks. In contrast, FGSM only reduces mAP by about 13%. This study provides an initial foundation for further research on model robustness and strength against adversarial attacks in the context of sign language recognition. Furthermore, the findings highlight the importance of developing robust BISINDO recognition systems to ensure reliable communication accessibility in real-world applications.

Recommended citation: Saputra, M. A., Heryadi, Y., Sonata, I., Wulandhari, L. A., & Girsang, A. S. (2025, October). Evaluating the Adversarial Vulnerability of YOLOv11 for Indonesian Sign Language (BISINDO) Recognition. In 2025 8th International Conference on Informatics and Computational Sciences (ICICoS) (pp. 241-246). IEEE.
Download Paper | Download Slides

Oil Palm Condition Monitoring via UAV Imagery Using YOLOv11 Enhanced for Class Imbalance

Published in 2025 17th International Conference on Information Technology and Electrical Engineering (ICITEE), 2025

Effective monitoring of oil palm tree conditions is crucial for advancing precision agriculture. Previous studies have primarily focused on detecting and counting oil palm trees, but such approaches are insufficient without the ability to assess tree health. This study proposes a modified YOLOv11 model for oil palm condition detection using the MOPAD dataset, which includes five condition classes. The modification enhances the loss function by introducing class weighting based on data distribution, applying weighted Focal Loss to mitigate the dominance of easy negatives, employing SmoothL1 Loss for more stable bounding box regression, and approximating Distribution Focal Loss (DFL) for distance distribution prediction. Experimental results demonstrate that the proposed model outperforms baseline methods-YOLOv8, Deformable DETR, and standard YOLOv11-achieving superior performance with a precision of 99.80%, recall of 99.47%, and F1-score of 99.64%. Visual analysis of inference outputs further confirms that the modified YOLOv11 delivers more consistent condition detection and closer alignment with ground truth, validating the effectiveness of the loss rebalancing strategy in addressing class imbalance for UAV-based oil palm monitoring.

Recommended citation: Saputra, M. A., Soeparno, H., Arifin, Y., & Budiharto, W. (2025, October). Oil Palm Condition Monitoring via UAV Imagery Using YOLOv11 Enhanced for Class Imbalance. In 2025 17th International Conference on Information Technology and Electrical Engineering (ICITEE) (pp. 1-6). IEEE.
Download Paper | Download Slides

Online Gambling Promotion Detection in Indonesian YouTube Comments Using Semi-Supervised IndoBERT Classification

Published in The 10th International Conference on Computer Science and Computational Intelligence (ICCSCI), 2025

The widespread promotion of online gambling through user-generated content on social media platforms has raised significant concerns, especially in regions with high internet penetration such as Indonesia. This research proposes a classification-based approach to detect online gambling promotional content in Indonesian-language YouTube comments using a semi-supervised learning strategy. A total of 10,101 comments were collected, of which 10% were manually labeled as either Promotion or Non-Promotion. Initial training using only the labeled data employed IndoBERT, a pre-trained language model for Bahasa Indonesia, which achieved an accuracy of 0.953, precision of 0.945, recall of 0.977, and F1-score of 0.961. Although the baseline performance was strong, the IndoBERT model was slightly outperformed by the SVM classifier. To address the limitations of labeled data, pseudo-labeling was applied to the remaining unlabeled data, which was then combined with the labeled subset to retrain the IndoBERT model. The evaluation results showed a significant improvement in classification performance after semi-supervised training, achieving an accuracy of 0.986, precision of 0.991, recall of 0.980, and F1-score of 0.985, surpassing all previously tested models. These findings highlight the effectiveness of IndoBERT in understanding informal online language and the advantage of semi-supervised learning in resource-constrained annotation scenarios.

Recommended citation: Santika, S. P., Saputra, M. A., Zahra, A., & Suhartono, D. (2025). Online Gambling Promotion Detection in Indonesian YouTube Comments Using Semi-Supervised IndoBERT Classification. Procedia Computer Science, 269, 1269-1278.
Download Paper | Download Slides

Deep Learning and Explainable AI for Accurate and Interpretable Software Defect Prediction

Published in The 2024 International Conference on Data Engineering and Enterprise System (ICDEES 2024), 2024

Software defect prediction is an essential aspect of software development process as defects significantly impact software reliability and usability. Several methods have been studied to predict the defects in software, but most of the time, the prediction result lack of transparancy about which features actually contributes to the prediction results. In this study, we adopt the Convolutional Neural Network (CNN) model for defect prediction, then apply the Local Interpretable Model-Agnostic Explanations (LIME) to gain interpretability on the prediction result. In order to address the data imbalance issue, we apply the Synthetic Minority Over-sampling Technique (SMOTE). Experiments on the NASA Promise repository datasets (CM1, JM1, KC1, KC2, and PC1) shows that the model achieve accuracy ranging from 81% to 92% across the dataset. Furthermore, through the LIME analysis, some metrics such as Lines of Code (LoC) and Effort (e) give more substantial influence in the defect predictions.

Recommended citation: Saputra, M. A., Lumban Gaol, F., Soeparno, H., & Arifin, Y. (2024). Deep learning and explainable AI for accurate and interpretable software defect prediction. In Proceedings of The 2024 International Conference on Data Engineering and Enterprise System (The ICDEES 2024). Yogyakarta, Indonesia, November 29-30, 2024.
Download Paper | Download Slides