Hoai Nam Vu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hoai Nam Vu has a total of 28 co-authors. Here are the top co-authors by number of shared publications:
- Dat Tran-Anh
Thuyloi University
3 shared publications - Manh Dung Nguyen
Posts and Telecommunications Institute of Technology
3 shared publications - Soonghwan Ro
Kongju National University
3 shared publications Bokgil Choi
Department of Electrical Engineering, Kongju National University, Cheonan, South Korea
2 shared publications- Cong Tran
Posts and Telecommunications Institute of Technology
2 shared publications - Cuong Pham
Posts and Telecommunications Institute of Technology, Hanoi, Vietnam
2 shared publications
Latest Publications
- Open Access
- 22 May 2026
- Published by Elsevier in SSRN Electronic Journal
Few-Shot Sign Language Recognition via Dynamic Text and SequentialLabel Pairing
Current multimodal contrastive learning frameworks for Sign Language Recognition (SLR) might often face on challenges such as complex, semantically imbalanced, and high-dimensional video kinematics that are supervised by sparse and static text glosses. Those might limit the learning of fine-grained physical movements, degrading representation quality in low-resource scenarios. To address these challenges, we propose DPSign-FL, a novel few-shot framework that incorporates dense semantic grounding with sequential context modeling. During pre-training, DPSign-FL replaces conventional isolated glosses with detailed textual descriptions of physical actions, injecting fine-grained supervision into the multimodal latent space. For the few-shot fine-tuning phase, we introduce a Label Pairing strategy combined with Cosine-based Prototypical Networks. By synthesizing pseudo-continuous context from isolated signs, this approach enforces sequential boundary awareness and effectively disentangles visually similar kinematic sequences (homophenes). Furthermore, to empirically evaluate our framework on a genuinely low-resource language, we introduce the Text-Based Vietnamese Sign Language (TB-VSL) dataset, the first SLR benchmark featuring explicit physical action descriptions. Extensive experiments demonstrate that DPSign-FL establishes a new state-of-the-art across multiple benchmarks. Notably, the proposed method outperforms the leading baseline on WLASL by an absolute margin of 14.85% (achieving 88.11% in the 5-way 5-shot setting) and yields highly competitive accuracy on the LSA64 and TB-VSL datasets. To facilitate reproducibility, our source code and dataset will be made publicly available upon acceptance.
- 22 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFew-Shot Sign Language Recognition via Dynamic Text and SequentialLabel Pairing
Current multimodal contrastive learning frameworks for Sign Language Recognition (SLR) might often face on challenges such as complex, semantically imbalanced, and high-dimensional video kinematics that are supervised by sparse and static text glosses. Those might limit the learning of fine-grained physical movements, degrading representation quality in low-resource scenarios. To address these challenges, we propose DPSign-FL, a novel few-shot framework that incorporates dense semantic grounding with sequential context modeling. During pre-training, DPSign-FL replaces conventional isolated glosses with detailed textual descriptions of physical actions, injecting fine-grained supervision into the multimodal latent space. For the few-shot fine-tuning phase, we introduce a Label Pairing strategy combined with Cosine-based Prototypical Networks. By synthesizing pseudo-continuous context from isolated signs, this approach enforces sequential boundary awareness and effectively disentangles visually similar kinematic sequences (homophenes). Furthermore, to empirically evaluate our framework on a genuinely low-resource language, we introduce the Text-Based Vietnamese Sign Language (TB-VSL) dataset, the first SLR benchmark featuring explicit physical action descriptions. Extensive experiments demonstrate that DPSign-FL establishes a new state-of-the-art across multiple benchmarks. Notably, the proposed method outperforms the leading baseline on WLASL by an absolute margin of 14.85% (achieving 88.11% in the 5-way 5-shot setting) and yields highly competitive accuracy on the LSA64 and TB-VSL datasets. To facilitate reproducibility, our source code and dataset will be made publicly available upon acceptance.
- 19 March 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 204, 55-63
Lightweight moment-residual-coherent patterns for image recognition
- 19 March 2026
- Published by Elsevier in Pattern Recognition Letters
- Vol. 204, 55-63
Lightweight moment-residual-coherent patterns for image recognition
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Wireless Communications Letters
- Vol. 15 (21622337) , 1588-1592
Diffusion Model-Enhanced Environment Reconstruction in ISAC
Recently, environment reconstruction (ER) in integrated sensing and communication (ISAC) systems has emerged as a promising approach for achieving high-resolution environmental perception. However, the initial results obtained from ISAC systems are coarse and often unsatisfactory due to the high sparsity of the point clouds and significant noise variance. To address this problem, we propose a noise–sparsity-aware diffusion model (NSADM) post-processing framework. Leveraging the powerful data recovery capabilities of diffusion models, the proposed scheme exploits spatial features and the additive nature of noise to enhance point cloud density and denoise the initial input. Simulation results demonstrate that the proposed method significantly outperforms existing model-based and deep learning-based approaches in terms of Chamfer distance and root mean square error.
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Wireless Communications Letters
- Vol. 15 (21622337) , 1588-1592
Diffusion Model-Enhanced Environment Reconstruction in ISAC
Recently, environment reconstruction (ER) in integrated sensing and communication (ISAC) systems has emerged as a promising approach for achieving high-resolution environmental perception. However, the initial results obtained from ISAC systems are coarse and often unsatisfactory due to the high sparsity of the point clouds and significant noise variance. To address this problem, we propose a noise–sparsity-aware diffusion model (NSADM) post-processing framework. Leveraging the powerful data recovery capabilities of diffusion models, the proposed scheme exploits spatial features and the additive nature of noise to enhance point cloud density and denoise the initial input. Simulation results demonstrate that the proposed method significantly outperforms existing model-based and deep learning-based approaches in terms of Chamfer distance and root mean square error.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 107735-107746
Agri-VLG: Robust Identification of Counterfeit Agricultural Products Using Few-Shot Annotated Data and Open-Source Knowledge
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 107735-107746
Open AccessAgri-VLG: Robust Identification of Counterfeit Agricultural Products Using Few-Shot Annotated Data and Open-Source Knowledge
- Open Access
- 17 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 195987-195995
AgriDetectVL: Emphasizes the Agriculture-Focused Application Combined With Visual–Language Integration
Counterfeit-product monitoring in agriculture demands models that exploit temporal context, accept operator feedback, and run under tight compute budgets.We introduce AgriDetectVL, an interactive, resource-efficient vision–language model that fuses time-series imagery with human inputs. AgriDetectVL couples an efficient visual backbone with a lightweight Sequence Prompt Transformer that summarizes recent observations and feedback into compact prompts. Class names and domain phrases are encoded as text prototypes, and images are mapped into a shared, L2-normalized space; decisions are made by temperature-scaled cosine scoring, enabling single-pass, low-latency inference and straightforward zero/few-shot extension. Evaluated on TLU-Fruit (fine-grained varieties) and TLU-States (state/ripeness), AgriDetectVL consistently surpasses strong Convolutional Neural Network (CNN), transformer, and Vision-Language Model (VLM) baselines across F1-Score (F1), accuracy, Area Under the ROC Curve (AUC), and Matthews Correlation Coefficient (MCC), while meeting edge-device constraints. Ablations confirm that sequence-aware prompting and prototype guidance are the primary sources of gain. In longitudinal tests, human-in-the-loop operation reduces manual corrections over time, indicating practical readiness for field deployment. Code is available at: https://github.com/NguyenAnhDucIT/AgriDetectVL.
- 17 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 195987-195995
Open AccessAgriDetectVL: Emphasizes the Agriculture-Focused Application Combined With Visual–Language Integration
Counterfeit-product monitoring in agriculture demands models that exploit temporal context, accept operator feedback, and run under tight compute budgets.We introduce AgriDetectVL, an interactive, resource-efficient vision–language model that fuses time-series imagery with human inputs. AgriDetectVL couples an efficient visual backbone with a lightweight Sequence Prompt Transformer that summarizes recent observations and feedback into compact prompts. Class names and domain phrases are encoded as text prototypes, and images are mapped into a shared, L2-normalized space; decisions are made by temperature-scaled cosine scoring, enabling single-pass, low-latency inference and straightforward zero/few-shot extension. Evaluated on TLU-Fruit (fine-grained varieties) and TLU-States (state/ripeness), AgriDetectVL consistently surpasses strong Convolutional Neural Network (CNN), transformer, and Vision-Language Model (VLM) baselines across F1-Score (F1), accuracy, Area Under the ROC Curve (AUC), and Matthews Correlation Coefficient (MCC), while meeting edge-device constraints. Ablations confirm that sequence-aware prompting and prototype guidance are the primary sources of gain. In longitudinal tests, human-in-the-loop operation reduces manual corrections over time, indicating practical readiness for field deployment. Code is available at: https://github.com/NguyenAnhDucIT/AgriDetectVL.
- Open Access
- 28 July 2025
- Published by Elsevier in ICT Express
- Vol. 12 (2) , 289-294
Transformer guided Multimodal VQA model for Fruit recognitions
- 28 July 2025
- Published by Elsevier in ICT Express
- Vol. 12 (2) , 289-294
Open AccessTransformer guided Multimodal VQA model for Fruit recognitions
- 19 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (22) , 27828-27840
Sign Language Recognition With Self-Learning Fusion Model
Sign Language Recognition (SLR) is the task of recognizing human actions that represent the language, which is not only helpful for deaf-mute people but also a means for human-computer interaction. Although data from wearable sensors have been proven useful for this task, it is still difficult to collect such data for training deep fusion models. In this study, our contributions are two-fold: 1) we collect and release a dataset for SLR consisting of both video and sensor data obtained from wearable devices, and 2) we propose the first self-learning fusion model for SLR, termed STSLR, that utilizes a portion of annotated data to simulate sensor embedding vectors. By virtue of the simulated sensor features, the video features from video-only data are enhanced to allow the fusion model to recognize the annotated actions more effectively. We empirically demonstrate the superiority of STSLR over competitive benchmarks on our newly released dataset and well-known publicly available ones.
- 19 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (22) , 27828-27840
Sign Language Recognition With Self-Learning Fusion Model
Sign Language Recognition (SLR) is the task of recognizing human actions that represent the language, which is not only helpful for deaf-mute people but also a means for human-computer interaction. Although data from wearable sensors have been proven useful for this task, it is still difficult to collect such data for training deep fusion models. In this study, our contributions are two-fold: 1) we collect and release a dataset for SLR consisting of both video and sensor data obtained from wearable devices, and 2) we propose the first self-learning fusion model for SLR, termed STSLR, that utilizes a portion of annotated data to simulate sensor embedding vectors. By virtue of the simulated sensor features, the video features from video-only data are enhanced to allow the fusion model to recognize the annotated actions more effectively. We empirically demonstrate the superiority of STSLR over competitive benchmarks on our newly released dataset and well-known publicly available ones.
- Open Access
- 9 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 120200-120212
Integrative Few-Shot Classification and Segmentation for Landslide Detection
There has been an ongoing demand for monitoring landslides due to the heavy economic losses and casualties caused by such natural disasters. In this paper, we introduce a swift landslide detection system that can detect and segment landslides occurring on roads. To tackle the challenges of data collection, we propose an automatic annotation procedure to create a new landslide dataset consisting of 2963 images, termed the LandslidePTIT dataset. Additionally, we construct a novel deep learning architecture that can perform both classification and segmentation tasks well from a few annotated images of landslides. Specifically, the model consists of four main modules that are delicately designed to solve the few-shot segmentation problem using landslide images, namely hypercorrelation construction, attentive squeeze block, a cross-feature layer, and broadcast and squeeze layer. Experimental results exhibit the superiority of the proposed methods in comparison with competitive baselines, in terms of both quantitative and qualitative manners.
- 9 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 120200-120212
Open AccessIntegrative Few-Shot Classification and Segmentation for Landslide Detection
There has been an ongoing demand for monitoring landslides due to the heavy economic losses and casualties caused by such natural disasters. In this paper, we introduce a swift landslide detection system that can detect and segment landslides occurring on roads. To tackle the challenges of data collection, we propose an automatic annotation procedure to create a new landslide dataset consisting of 2963 images, termed the LandslidePTIT dataset. Additionally, we construct a novel deep learning architecture that can perform both classification and segmentation tasks well from a few annotated images of landslides. Specifically, the model consists of four main modules that are delicately designed to solve the few-shot segmentation problem using landslide images, namely hypercorrelation construction, attentive squeeze block, a cross-feature layer, and broadcast and squeeze layer. Experimental results exhibit the superiority of the proposed methods in comparison with competitive baselines, in terms of both quantitative and qualitative manners.
- Open Access
- 22 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 146667-146679
Multistage Real-Time Fire Detection Using Convolutional Neural Networks and Long Short-Term Memory Networks
Fire is one of the most commonly occurring disasters and is the main cause of catastrophic personal injury and devastating property damage. An early detection system is necessary to prevent fires from spreading out of control. In this paper, we propose a multistage fire detection method using convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. In the first stage, fire candidates are detected by using their salient features, such as their color, flickering frequency, and brightness. In the second stage, a pretrained CNN model is used to extract the 2D features of flames that are the input for the LSTM network. In the last stage, a softmax classifier is utilized to determine whether the flames represent a true fire or a nonfire moving object. The experimental results show that our proposed method can achieve competitive performance compared with other state-of-the-art methods and is suitable for real-world applications.
- 22 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 146667-146679
Open AccessMultistage Real-Time Fire Detection Using Convolutional Neural Networks and Long Short-Term Memory Networks
Fire is one of the most commonly occurring disasters and is the main cause of catastrophic personal injury and devastating property damage. An early detection system is necessary to prevent fires from spreading out of control. In this paper, we propose a multistage fire detection method using convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. In the first stage, fire candidates are detected by using their salient features, such as their color, flickering frequency, and brightness. In the second stage, a pretrained CNN model is used to extract the 2D features of flames that are the input for the LSTM network. In the last stage, a softmax classifier is utilized to determine whether the flames represent a true fire or a nonfire moving object. The experimental results show that our proposed method can achieve competitive performance compared with other state-of-the-art methods and is suitable for real-world applications.
- Open Access
- 21 July 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 8, 132625-132641
Detecting and Tracking Sinkholes Using Multi-Level Convolutional Neural Networks and Data Association
Sinkholes can cause severe property damage and threaten public safety. Therefore, the early prediction and detection of sinkholes are important measures for protecting both citizenry and infrastructure. Although many studies have made significant progress on sinkhole detection, challenges remain, including long-term data collection and the discovery of lightweight machine learning models that can be deployed to analyze sinkhole images. In this paper, we propose a method that takes advantage of the recent success of deep learning models to detect and track sinkholes via video streaming. Our system consists of two main stages: sinkhole detection with a cascaded convolutional neural network and sinkhole tracking with a data association algorithm. The experimental results show that a sinkhole can be tracked in real time using the dataset [1]. Furthermore, we implement the system on a Jetson TX2 embedded board (weighing 85 grams), which can operate at 13.2 FPS (frames per second). With an average IoU (intersection over union) score of 88% for sinkhole tracking and an accuracy of 97,6% for sinkhole detection on a 45-minute dataset, this study demonstrates the feasibility of sinkhole detection and tracking using IR images and their suitability for practical applications.
- 21 July 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 8, 132625-132641
Open AccessDetecting and Tracking Sinkholes Using Multi-Level Convolutional Neural Networks and Data Association
Sinkholes can cause severe property damage and threaten public safety. Therefore, the early prediction and detection of sinkholes are important measures for protecting both citizenry and infrastructure. Although many studies have made significant progress on sinkhole detection, challenges remain, including long-term data collection and the discovery of lightweight machine learning models that can be deployed to analyze sinkhole images. In this paper, we propose a method that takes advantage of the recent success of deep learning models to detect and track sinkholes via video streaming. Our system consists of two main stages: sinkhole detection with a cascaded convolutional neural network and sinkhole tracking with a data association algorithm. The experimental results show that a sinkhole can be tracked in real time using the dataset [1]. Furthermore, we implement the system on a Jetson TX2 embedded board (weighing 85 grams), which can operate at 13.2 FPS (frames per second). With an average IoU (intersection over union) score of 88% for sinkhole tracking and an accuracy of 97,6% for sinkhole detection on a 45-minute dataset, this study demonstrates the feasibility of sinkhole detection and tracking using IR images and their suitability for practical applications.