Tạp chí Khoa học Lạc Hồng- Chuyên san Kỹ thuật

Automated orange quality classification using convolutional neural networks: A deep learning approach for smart agriculture

2025-10-14T10:57:21+07:00

Quality control is the core activity of an agribusiness and food processing industry just to make sure that the customers have access to quality oranges in a reduced wastage system. This study molds a deep learning idea to classify oranges as either good or bad. These images capture critical features such as consistency of color, surface texture, and apparent defects. Brightness adjustments, enhanced contrasts, and even the addition of some noise are among the possible scenes to improve model generalization error performance. The proposed system would give an automated and scalable real-time orange grading system that would gradually reduce the influence of time-based human inspection practices and improve quality. The finding that even a simple CNN without any pre-train models can be used to achieve high accuracy in this classification task indeed, the results provide for deep learning to be effective in fruit sorting, with scope for much else based on larger data sets, as well as real-world deployment potential.

Feasibility and performance study of LLMS on mobile devices for supporting C++ programming learning

2025-10-14T10:57:23+07:00

Learning C++ programming is a complex process that requires mastering both syntax and algorithmic thinking. This study aims to evaluate the feasibility of deploying large language models (LLMs) on mobile devices to support users in learning C++ more effectively. The research involved testing models such as DeepSeek-Coder, Llama, and Gemma, and applying optimization techniques like 4-bit and 8-bit quantization to reduce hardware resource consumption. Experiments measured model accuracy on C++ tasks, memory usage (VRAM, RAM), and inference speed under different optimization levels. Results showed that DeepSeek-Coder-1.3B achieved the highest accuracy among mobile-friendly models, solving around 40% of C++ problems with 3.2GB of VRAM—suitable for smartphones. Meanwhile, DeepSeek-V2-Lite-Instruct (4-bit) reached 64% accuracy but consumed 6GB VRAM, making it more appropriate for laptops. After quantization, the model ran stably on devices such as the Samsung A52S (8GB RAM), requiring approximately 1.9GB of system RAM (excluding OS usage), which ensures acceptable performance on mid-range mobile devices. The findings confirm that deploying LLMs on mobile platforms is feasible and holds significant potential in supporting programming education. In the future, the research team will continue to optimize performance and improve the user interface to enhance the overall learning experience.

Improving accuracy in facial emotion recognition through PCA and Artificial Neural Networks

2025-10-14T10:57:26+07:00

This research addresses the challenge of accurately identifying human emotions through facial expressions using Principal Component Analysis (PCA) combined with Artificial Neural Networks (ANN). The method involves preprocessing facial images, extracting critical features using PCA, and classifying emotional states with ANN. We utilized two standard facial expression datasets JAFFE and FEI for evaluation, focusing on basic emotions: happiness, sadness, surprise, and neutrality. Experimental results demonstrated that the proposed PCA-ANN approach achieved average accuracy rates of 96.3% on JAFFE and 93.8% on FEI datasets, outperforming several traditional methods in terms of computational efficiency and classification accuracy. Despite limitations concerning dataset size and emotion diversity, this research contributes to developing robust systems for real-world applications such as interactive technologies, assistive communication, and security systems. Future directions include expanding emotion recognition capabilities and integrating multimodal data for improved accuracy.

Application of MediaPipe and SVM in sign language recognition to support the hearing impaired

2025-10-14T10:57:28+07:00

Sign language serves as an essential means of communication for the hearing-impaired community. However, access to and adoption of sign language in Vietnam remain limited due to a lack of resources and supporting tools. To address this issue, this study proposes a Vietnamese Sign Language recognition system that combines MediaPipe technology with the Support Vector Machine (SVM) classification algorithm. The training dataset is constructed from hand gesture images, with MediaPipe responsible for detecting and extracting hand landmark features. These features are then classified using the SVM model. Experimental results demonstrate that the system achieves an accuracy rate between 85% and 90%, confirming its potential to support communication for hearing-impaired individuals through sign language recognition technology.

Applying clustering methods to classify customers based on shopping behaviour

2025-10-14T10:57:29+07:00

Customer segmentation is crucial for optimizing marketing strategies. This study applies and compares the effectiveness of three common clustering algorithms: K-Means, Hierarchical Clustering, and Gaussian Mixture Models (GMM) to classify customers based on shopping behavior and demographics (age, gender, total spending). Utilizing three retail datasets (two from Kaggle, one from Sling Academy), the research performs data preprocessing, applies the clustering algorithms, and evaluates their performance using Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index. The results indicate that GMM performs most effectively for segmenting based on total spending and gender, creating distinct clusters. Hierarchical Clustering proves suitable for detailed age-based analysis on specific datasets, while K-Means offers a balanced solution, particularly effective when cluster structures are clear or rapid results are needed. The study recommends selecting appropriate algorithms based on specific business objectives and data characteristics, enabling businesses to develop more effective personalized marketing strategies.

6. Applying machine learning and computer vision for Planogram compliance evaluation in retail environments

2025-10-14T10:57:31+07:00

This paper presents a novel approach for automated planogram compliance assessment in retail environments, with a focus on the Vietnamese market. Addressing the limitations of manual inspection methods—which are time-consuming, error-prone, and difficult to scale—the proposed system integrates recent advances in computer vision and machine learning. Specifically, the method leverages state-of-the-art object detection models, YOLOv11 and YOLOv12, trained on annotated shelf images collected from real retail settings. Detected products are spatially organized using the DBSCAN clustering algorithm, while the Hungarian algorithm is employed to match detected layouts with predefined planograms and compute compliance scores. Experimental results demonstrate high detection accuracy and reliable compliance evaluation, even under complex retail conditions. The combination of advanced YOLO models with spatial reasoning techniques proves effective in handling challenges unique to the Vietnamese retail landscape, such as inconsistent shelf organization and varied packaging. This work contributes a scalable, accurate, and practical solution for enhancing retail execution and operational efficiency.

Research and development of a method to select the segmentation algorithm for the problem of detecting irregularities in time series data

2025-10-14T10:57:33+07:00

Research and Enhancement of Segmentation Methods for Anomaly Detection in Time Series Data. This paper presents a study and improvement of segmentation techniques applied to the problem of anomaly detection in time series data. The work outlines the process of collecting and constructing essential datasets, including both periodic and non-periodic time series, along with a detailed description of the SWAT 2019 dataset. Subsequently, the paper delves into the segmentation process, focusing on the extraction of anomalous segments and the evaluation of experimental results. Finally, the study discusses the selection of the maximum allowable error (max_error), a critical parameter for optimizing the segmentation process and improving the performance of anomaly detection.

Imbalanced data classification using random forest with Ward clustering

2025-10-14T10:57:34+07:00

This study introduces a Modified Balanced Random Forest algorithm to improve classification performance on imbalanced datasets. The proposed method enhances the Balanced Random Forest by applying a clustering based under sampling strategy during each bootstrap iteration. Four clustering methods were evaluated including K Means, Spectral Clustering, Agglomerative Clustering, and Ward Hierarchical Clustering. Among these, the Ward Hierarchical Clustering technique achieved the best performance. Experimental results show that the proposed method outperforms standard Random Forest and Balanced Random Forest, reaching a true positive rate of 93.42 percent, a true negative rate of 93.60 percent, and an area under the curve accuracy of 93.51 percent, while also reducing processing time. These results confirm the effectiveness of the proposed approach for imbalanced data classification.

Artificial Intelligence: A catalyst for sustainable education and enhanced learning quality

2025-10-14T10:57:35+07:00

Artificial Intelligence (AI) is increasingly playing a central role in reshaping industries, economies, and social life, directly impacting the foundations of sustainable development as well as the education system. The application of AI across various sectors has led to significant advancements in work efficiency and decision-making processes. However, alongside these benefits, there are concerns regarding the ethical, social, and environmental consequences that this technology may bring. This paper analyzes how AI can be leveraged as a tool to support the achievement of the United Nations' Sustainable Development Goals (SDGs), with a focus on the education sector. Through three case studies, the paper examines the dual role of AI – both as a driver of innovation and a source of new challenges – and offers insights into policy-making, trade direction, and workforce training. The conclusion suggests that, with proper guidance, AI can become a key factor in advancing global sustainable development and significantly contribute to the reform of education, preparing learners to adapt to a continuously changing world.

A bibliometric study for mapping the Metaverse of Global research patterns and Country level differences

2025-10-14T10:57:36+07:00

Metaverse research has gained global attention as a multidisciplinary field that brings together technology, media, and human interaction. This study presents a bibliometric analysis of Metaverse-related publications from 2012 to 2021, using data from the Scopus database and visualization tools such as VOSviewer to identify research trends, keyword associations, and patterns of international collaboration. The findings show that the field is growing rapidly, with the United States, China, and Germany leading in publication output. Virtual reality is identified as the most frequently studied topic. This study offers a concise overview of current research and provides direction for future academic and technological progress in the Metaverse.

11. Application of physics-informed neural networks in simulating heat transfer and mass diffusion

2025-10-14T10:57:38+07:00

This paper presents a novel approach to simulating classical physical phenomena-specifically heat transfer and mass diffusion-using Physics-Informed Neural Networks (PINNs), a class of deep neural networks that incorporate physical constraints. Unlike conventional machine learning models, PINNs allow the integration of empirical data with partial differential equations (PDEs) governing the underlying physical systems. This results in models capable of making accurate predictions even in the presence of incomplete or noisy data. The study constructs and trains PINN models for two canonical problems: heat conduction in a one-dimensional (1D) rod and concentration diffusion in a closed medium. Simulation results demonstrate that the PINNs achieve significantly lower prediction errors compared to standard neural networks without physical constraints, while also exhibiting strong generalization capabilities and numerical stability. This method offers a promising new direction for simulating physical processes, particularly in scenarios where real-world data are limited-making it well-suited for applications in education, engineering, and scientific research.

12. Automated detection of concrete spalling in post-earthquake structures using deep learning

2025-10-14T10:57:39+07:00

Post-earthquake structural assessment is critical in determining the extent of damage and guiding emergency response efforts. Spalling, characterized by the detachment of concrete layers, serves as a key indicator of seismic damage and can significantly impact structural integrity. This study develops an automated classification model utilizing deep learning to distinguish between spalling and non-spalling cases in concrete structures. The proposed method employs transfer learning with ResNet50 and EfficientNet-B3 to optimize accuracy and inference efficiency. The dataset, collected from real-world post-earthquake reconnaissance, consists of high-resolution images categorized into spalling and non-spalling classes. Key preprocessing techniques, including pixel normalization, data augmentation, and class balancing, were applied to improve model robustness and mitigate class imbalance issues. Performance evaluation showed that ResNet50 outperforms EfficientNet-B3 in overall accuracy (77% vs. 71%), while EfficientNet-B3 achieved higher recall (90% vs. 85%), making it more sensitive to detecting spalling cases. The study highlights the challenges posed by dataset variability and proposes future enhancements such as advanced augmentation, multi-modal data integration, and self-supervised learning. The findings contribute to the advancement of AI-driven structural health monitoring, offering an efficient tool for rapid post-disaster damage assessment.