Journal on Information Technologies & Communications

Phương pháp diễn giải cho mô hình hỏi đáp hình ảnh tiếng Việt

Nguyễn Thanh Tân — 2025-05-20

Interpretability in Visual Question Answering (VQA) models is a topic of interest because the transparency of the model can help users trust and understand its strengths and limitations. Multimodal models like VQA utilize both image and text data, posing challenges for interpretability techniques that are typically designed for a single type of data. In this work, we propose a combination of the Grad-CAM image explanation method and the Transformer-based text explanation method to analyze the behavior of the model. The results obtained from the Vietnamese VQA dataset lead us to the following observations: i) The visual Transformer extracts global features, while ResNet extracts local features; ii) the model struggles with images containing many distracting elements; iii) the text modality has less impact compared to the image modality; iv) the PhoBERT module shows bias toward certain types of questions.

Một phương pháp giấu tin thuận nghịch có khả năng nhúng cao sử dụng kỹ thuật ảnh kép và hệ tính toán cơ số 13.

Thái Phạm Minh — 2025-05-22

Phương pháp Chen và cộng sự [22] đề xuất một lược đồ RDH mới dựa trên khai thác biến đổi theo hướng (EMD) với hình ảnh kép. Nhờ khai thác các đặc điểm của ma trận EMD, một bít và một chữ số cơ số 5 có thể được giấu vào trong một điểm ảnh gốc để tạo ra một cặp điểm ảnh chứa tin. Dữ liệu sẽ nhúng trên 1 trong 4 hướng, do đó các tác giả nhúng thêm được 1 bít ngoài việc nhúng 1 hệ cơ số 5, nên tỷ lệ nhúng (log₂⁵+1)/2 =1.66 bpp. Để tăng khả năng nhúng, chúng tôi đề xuất một phương pháp nhúng dữ liệu dựa trên hệ cơ số 13. Vì vậy tỷ lệ nhúng có thể xấp xỉ bằng: (log₂¹³)/2 = 1.8502 bpp. Ngoài ra độ biến đổi ảnh trong phương pháp đề xuất nhỏ hơn, nên chất lượng ảnh tốt hơn phương pháp của Chen và cộng sự.

Phương pháp xử lý tín hiệu và tạo ảnh 3D cho các vật thể dưới nước bằng công nghệ sonar chủ động

Nguyễn Văn Đức — 2025-05-25

The article proposes a method of 3D image signal processing for underwater objects by using sonar positioning technology. The proposed method is based on the collected acoustic signal from many receivers. The 3-D imaging algorithm will calculate the pixel matrix value based on the sweep beam aperture, the detected distance from the sonar scanner to the reflected object, the speed of ship movement, etc. After determining the image matrix value, the scanned image is reconstructed on the 2D plane or 3D space. The experimental results were carried out in some deep lake areas such as Hoa Binh, Dong Do lake, and Ha Long bay.

Khmer News Classification in Low-Resource Settings: A comparative Analysis of Embedding Method

Natt Korat — 2025-05-26

Text classification in low-resource languages like Khmer remains challenging due to linguistic complexity, limited annotated data, and noise from real-world applications. This study addresses these challenges by systematically comparing text embedding techniques for Khmer news classification. We evaluate traditional methods (TF-IDF with SVM) against state-of-the-art multilingual transformers (XLM-RoBERTa, LaBSE) using a self-collected dataset of 7,344 Khmer news articles across six categories—political, economic, entertainment, sport, technology, and life. The dataset intentionally retains noise (e.g., mixed-language text, unstructured formatting) to reflect practical scenarios. To address Khmer's lack of word boundaries, we employ word segmentation via khmer-nltk for traditional models, while transformer models leverage their inherent subword tokenization. Experiments reveal that transformer-based embeddings achieve superior performance, with XLM-RoBERTa and LaBSE attaining F1 scores of 94.2% and 94.3%, respectively, outperforming TF-IDF (93.3%). However, the "life" category proves challenging across all models (F1: 85.5–88.1%), likely due to semantic overlap with other categories. Our findings underscore the effectiveness of transformer architectures in capturing contextual nuances for low-resource languages, even with noisy data. This work offers insights for NLP researchers and practitioners, emphasizing the need for domain-specific adaptations and expanded datasets to improve performance in underrepresented languages.

Forecasting Seasonal Trends in Ear Nose Throat Diseases: A Comparative Analysis of Statistical and Machine Learning Models

Phuc Tran Huu Le — 2025-06-23

This study examines patterns of seasonal illnesses at an ENT hospital using statistical methods and
advanced machine learning techniques to improve disease prediction and support healthcare planning. The
data underwent careful cleaning to ensure accuracy, which involved identifying outliers, managing missing
values, and normalizing information. The study looked at how seasonal illnesses develop, using a mix of
techniques. This research used advanced tools like Long Short-Term Memory (LSTM) networks and
Prophet, as well as simpler models such as Holt-Winters and SARIMA. To make the models easier to
understand, there is an application of SHAP (SHapley Additive Explanations) values. Finally, these
statistical measures like Mean Absolute Error (MAE) and confidence periods had been used to check the
accuracy of the forecasts at some stage in the overall performance assessment. Moreover, to discover how
weather influences sickness seasonality, connections between patterns of contamination and environmental
variables like temperature, humidity, and rainfall have been additionally looked at with the aid of the usage
of correlation prediction. In well-known, this blended method shows how conventional and system gaining
knowledge of models can monitor seasonal illness traits. The effects not simplest display disorder styles,
but in addition they assist with allocating resources and making guidelines for better healthcare
management.

Performance Analysis of Deep Learning Models for Software Fault Prediction Using the BugHunter Dataset

Dang Thi Kim Ngan — 2025-06-23

Software fault prediction (SFP) involves the identification of potentially fault-prone modules before the testing
phase in the software development lifecycle. By predicting faults early in the development process, the SFP process enables software developers to focus their efforts on components that may contain faults, thereby enhancing the overall quality and reliability of the software. Machine learning and deep learning techniques have been widely applied to train SFP models. However, these approaches face several challenges, including irrelevant or redundant features, imbalanced datasets, overfitting, and complex model structures. The NASA dataset from the PROMISE repository is the most commonly used dataset for fault prediction. Recently, the BugHunter dataset with its substantially larger number of instances was explored to train the SFP models. In this study, we present the comparative study of three deep learning models, including Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) and four machine learning models as K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), Adaptive Boosting (AdaBoost),
Extreme Gradient Boosting (XGB) to investigate the performance of SFP models on the BugHunter dataset. We employ the Lasso method for feature selection and apply the Synthetic Minority Oversampling Technique (SMOTE) to address the issue of imbalanced data, aiming to enhance the accuracy of the results. The experimental findings reveal that CNN and RNN outperformed other machine learning models, achieving the best overall performance.