Hue University Journal of Science: Techniques and Technology

Design of a high-speed data transmission link using free-space optics communication technology for metropolitan information networks

Quang Phuoc Vuong — 2025-12-31

This paper describes the design and simulation of a high-speed data transmission link based on free-space optical communication (FSO) technology, motivated by the need for quick, cost-effective and high bandwidth solutions for metropolitan information networks. The proposed design helps to deal with unique characteristics of metropolitan information networks which are high node density with challenging physical infrastructures, high bandwidth connectivity demands at short to medium transmission distances, high atmospheric turbulence due to air pollution, smog, and building motion caused by transportation and construction activities. In the paper, key system parameters including receiver aperture diameter, transmitter beam divergence, electrical filter cut-off frequency, transmitted power and various environmental conditions are considered simultaneously to provide a design solution with a higher adaptability to channel instability due to high atmospheric turbulence and infrastructure constraints of urban area. The Q-factor is used as a quality criterion to evaluate through out the simulation and analyze of the system performance. For the data rate as high as 20 Gb/s of this study, the proposed FSO system design is ideal for scenarios that require short to medium-range high-speed links (up to 4 km) and under moderate weather conditions (visibility ≥ 2 km).

Evaluation of the effectiveness of modern object detection models on spatial image data

Dung Nguyen — 2025-12-31

Object detection in aerial imagery, especially from unmanned aerial vehicles (UAVs), presents numerous challenges due to varying altitudes, occlusions, and diverse object scales—particularly the detection of small objects. This paper provides a comparative evaluation of three advanced object detection models: YOLOv11, RT-DETR, and RF-DETR, using the VisDrone2019 dataset, which includes complex urban and suburban scenes captured from UAVs. We analyze the models based on key performance metrics such as mean average precision (mAP), inference speed, model size, and computational complexity. Experimental results show that YOLOv11 achieves the highest processing speed, making it especially suitable for real-time applications due to its fast inference and strong edge-device performance. RF-DETR, on the other hand, achieves the best accuracy, with the fastest mAP@0.5 and mAP@[0.5:0.95] scores of 46.9% and 26.6%, respectively, demonstrating effectiveness in complex scenarios with high object density and occlusions. RT-DETR offers a balanced trade-off between speed and accuracy, making it a practical choice for applications requiring both responsiveness and reliable detection quality. These findings clarify the strengths and limitations of each model and provide practical guidance for selecting suitable object detection models in UAV-based surveillance and tracking tasks.

An integrated approach combining neural networks and genetic algorithms for multisource time series forecasting

Ngo Van Son — 2025-12-31

This paper proposes a data-driven framework that leverages multisource information to significantly enhance time series forecasting within the tourism domain. Specifically, we integrate a Multilayer Perceptron (MLP) neural network with the NSGA-II algorithm for efficient hyperparameter optimization. In addition to historical data on monthly international tourist arrivals to Vietnam, two exogenous data sources are incorporated: (i) sentiment indices extracted from tourist reviews on TripAdvisor, and (ii) search trend data from Google Trends. Critical hyperparameters, including input window size, learning rate, number of epochs, early stopping, and normalization methods, are simultaneously optimized using NSGA-II. Experimental results demonstrate that the proposed NSGA-II-MLP model, utilizing multisource data, consistently outperforms the standard MLP and exhibits robust performance across both the pre-COVID-19 and volatile pandemic periods. The results underscore the effectiveness of combining multisource data with evolutionaroptimization for tourism demand forecasting.

Fine-tuning deep learning models on microscopic images of liver and intestine cells of shrimps using k-fold cross-validation

Thi Thu Thao Khong — 2025-12-31

We employ the fine-tuning technique to train compact and efficient deep convolutional neural networks—specifically MobileNet_V2, MobileNet_V3_Small, and MobileNetV3-Large – to classify the nutritional status of farmed shrimp. The classification is based on microscopic images of liver and intestinal cells, enabling rapid and scalable assessment of shrimp health through image-based diagnostics. The experiment was conducted on a dataset comprising 854 cellular images, and used k-fold cross-validation to split the dataset into the training and test sets. The pre-trained MobileNet_V3_Large was fine-tuned on our cellular image dataset using 10-fold cross-validation, achieving the highest mean classification accuracy of 90.89%. This study demonstrates the potential of applying deep learning techniques to the monitoring and nutritional management of farmed shrimp, aiming to enhance productivity in aquaculture operations.