Hailan Kuang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hailan Kuang has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Xiaolin Ma
Wuhan University of Technology
17 shared publications - Xinhua Liu
Hubei Key Laboratory of Broadband Wireless Communication and Sensor Networks, School of Information Engineering, Wuhan University of Technology, Wuhan, China
16 shared publications - Zehua Dong
Hubei Key Laboratory of Power System Design and Test for Electrical Vehicle and the School of Computer Engineering, Hubei University of Arts and Science, Xiangyang, China
3 shared publications Xinhua Liu
School of Information Engineering, Wuhan University of Technology, Wuhan, China
2 shared publications- Can Ao
Hubei Key Laboratory of Broadband Wireless Communication and Sensor Networks, School of Information Engineering, Wuhan University of Technology, Wuhan, China
1 shared publication - Cen Chen
South China University of Technology
1 shared publication
Latest Publications
- 18 December 2025
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (1) , 16
Multi-view dynamic reconstruction with cross-view smoothing based on surfel
- 18 December 2025
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (1) , 16
Multi-view dynamic reconstruction with cross-view smoothing based on surfel
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (13) , 24509-24516
Fast Reconstruction of Monocular Human Video Based on KAN
Creating 3D digital people from monocular video provides many possibilities for a wide range of users and rich applications. In this paper, we propose a fast, high-quality, and effective method for creating 3D digital humans from monocular videos, achieving fast training (2.5 minutes) and real-time rendering. Specifically, we use 3D Gaussian Splatting (3DGS), based on the introduction of Skinned Multi-Person Linear Model (SMPL) human structure prior, and an optimized KAN (Kolmogorov-Arnold-Net) neural network to build effective posture and linear blend skinning (LBS) weight estimation module to quickly and accurately learn the fine details of the 3D human body. In addition, to achieve fast optimization in the densification and prune stages, we propose a two-stage optimization method. First, the local 3D area that needs to be densified is extracted based on LightGlue, and then KL divergence combined with human body prior is further used to guide Gaussian splitting/cloning and merging operations. We conducted extensive experiments on the ZJU MoCap dataset, and the Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) metrics indicate that we effectively improved rendering quality while ensuring rendering speed.
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (13) , 24509-24516
Fast Reconstruction of Monocular Human Video Based on KAN
Creating 3D digital people from monocular video provides many possibilities for a wide range of users and rich applications. In this paper, we propose a fast, high-quality, and effective method for creating 3D digital humans from monocular videos, achieving fast training (2.5 minutes) and real-time rendering. Specifically, we use 3D Gaussian Splatting (3DGS), based on the introduction of Skinned Multi-Person Linear Model (SMPL) human structure prior, and an optimized KAN (Kolmogorov-Arnold-Net) neural network to build effective posture and linear blend skinning (LBS) weight estimation module to quickly and accurately learn the fine details of the 3D human body. In addition, to achieve fast optimization in the densification and prune stages, we propose a two-stage optimization method. First, the local 3D area that needs to be densified is extracted based on LightGlue, and then KL divergence combined with human body prior is further used to guide Gaussian splitting/cloning and merging operations. We conducted extensive experiments on the ZJU MoCap dataset, and the Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) metrics indicate that we effectively improved rendering quality while ensuring rendering speed.
- 1 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (13) , 22746-22757
An Adaptive and Scalable Framework for Resource-Efficient Deployment of Mixture of Experts in LLM-Based Intelligent IoT Networks
The exponential growth of the Internet of Things (IoT) necessitates the deployment of large-scale models capable of processing the complex and diverse data generated by IoT devices. However, the substantial memory requirements of these models pose significant challenges, especially in scenarios where rapid decision-making and low-latency responses are critical. To address these challenges, we propose three innovative strategies for optimizing large model usage in IoT environments. The first strategy is an adaptive loading scheme, which enables dynamic loading of individual model experts. The second strategy involves an expert-by-expert loading approach, further enhancing the ability to load experts as needed, which optimizes memory usage and accelerates computations. The third strategy employs an interlayer expert reuse mechanism, facilitating the efficient reuse of experts across different layers, thus enhancing response rates without compromising model accuracy. Importantly, these strategies can be directly applied to Mixture of Experts (MoE) large language models without requiring additional training, thereby providing a seamless and efficient solution for leveraging these models in memory-constrained, high-performance IoT environments.
- 1 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (13) , 22746-22757
An Adaptive and Scalable Framework for Resource-Efficient Deployment of Mixture of Experts in LLM-Based Intelligent IoT Networks
The exponential growth of the Internet of Things (IoT) necessitates the deployment of large-scale models capable of processing the complex and diverse data generated by IoT devices. However, the substantial memory requirements of these models pose significant challenges, especially in scenarios where rapid decision-making and low-latency responses are critical. To address these challenges, we propose three innovative strategies for optimizing large model usage in IoT environments. The first strategy is an adaptive loading scheme, which enables dynamic loading of individual model experts. The second strategy involves an expert-by-expert loading approach, further enhancing the ability to load experts as needed, which optimizes memory usage and accelerates computations. The third strategy employs an interlayer expert reuse mechanism, facilitating the efficient reuse of experts across different layers, thus enhancing response rates without compromising model accuracy. Importantly, these strategies can be directly applied to Mixture of Experts (MoE) large language models without requiring additional training, thereby providing a seamless and efficient solution for leveraging these models in memory-constrained, high-performance IoT environments.
- 22 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (17) , 28007-28017
LSDASCFormer: A Transformer-Like Network With Long–Short-Distance Attention and SCConv for Heart Rate Measurement
In recent years, many methods for extracting remote photoplethysmography (rPPG) signals for remote heart rate (HR) estimation based on the transformer approach have predominantly utilized multi-head self-attention mechanisms. However, these methods suffer from high computational complexity, making them unsuitable for deployment on terminal and mobile devices. Additionally, most approaches focus solely on local features, neglecting long-distance features. This paper proposes a transformer-like network structure called LSDASCFormer for heart rate extraction. The fundamental idea is to utilize LSDASCFormer to accurately extract rPPG signals from facial videos, followed by specific post-processing of the rPPG signals to calculate relevant heart rate indicators. LSDASCFormer comprises convolutional blocks, feature fusion blocks (LSDASCFormer block), and rPPG prediction blocks. The introduced feature fusion module effectively extracts both long-distance and short-distance features while reducing computational complexity. Our model demonstrates excellent performance in terms of mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (r) in experiments conducted on the UBFC-rPPG and PURE datasets. On the PURE dataset, our method achieves an MAE of 0.41 bpm, an RMSE of 0.66 bpm, and a Pearson correlation coefficient (R) of 1.00. Similarly, on the UBFC-rPPG dataset, the method yields an MAE of 0.49 bpm, an RMSE of 1.12 bpm, and a Pearson correlation coefficient (R) of 0.99. These results substantiate the effectiveness of our proposed method.
- 22 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (17) , 28007-28017
LSDASCFormer: A Transformer-Like Network With Long–Short-Distance Attention and SCConv for Heart Rate Measurement
In recent years, many methods for extracting remote photoplethysmography (rPPG) signals for remote heart rate (HR) estimation based on the transformer approach have predominantly utilized multi-head self-attention mechanisms. However, these methods suffer from high computational complexity, making them unsuitable for deployment on terminal and mobile devices. Additionally, most approaches focus solely on local features, neglecting long-distance features. This paper proposes a transformer-like network structure called LSDASCFormer for heart rate extraction. The fundamental idea is to utilize LSDASCFormer to accurately extract rPPG signals from facial videos, followed by specific post-processing of the rPPG signals to calculate relevant heart rate indicators. LSDASCFormer comprises convolutional blocks, feature fusion blocks (LSDASCFormer block), and rPPG prediction blocks. The introduced feature fusion module effectively extracts both long-distance and short-distance features while reducing computational complexity. Our model demonstrates excellent performance in terms of mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (r) in experiments conducted on the UBFC-rPPG and PURE datasets. On the PURE dataset, our method achieves an MAE of 0.41 bpm, an RMSE of 0.66 bpm, and a Pearson correlation coefficient (R) of 1.00. Similarly, on the UBFC-rPPG dataset, the method yields an MAE of 0.49 bpm, an RMSE of 1.12 bpm, and a Pearson correlation coefficient (R) of 0.99. These results substantiate the effectiveness of our proposed method.
- 6 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (19) , 30275-30286
CMRPPGFormer: 3-D Spatio-Temporal Convolutional Modulation Transformer Network for Remote Heart Rate Estimation
Most of the existing remote heart rate (HR) estimation methods based on deep learning networks use convolutional neural networks which excel at local feature extraction. This paper proposes a HR estimation method based on a vision transformer network named as CMRPPGFormer. The basic idea of the proposed method is to extract precise remote photoplethysmography (rPPG) signals from facial videos by the CMRPPGFormer and then calculate relevant indicators of HR by specific post-processing on rPPG signals. The CMRPPGFormer consists of the convolutional stem block, the feature fusion block and the rPPG predictor block. The feature fusion block, the key of the CMRPPGFormer, can effectively extract and fuse local and global features of rPPG signals, in which a three dimensional (3D) spatio-temporal convolutional modulation transformer (3D-ST-CMTransformer) model is proposed. The 3D-ST-CMTransformer extracts global features of rPPG signals by its 3D convolution modulation attention layer and feed-forward network effectively. Experiment results on two public datasets, PURE and UBFC-rPPG, validate the better performance of the proposed method compared to the state-of-the-art baseline methods. Particularly, CMRPPGFormer obtains 0.39 bpm MAE, 0.64 bpm RMSE, 0.99 r value of Pearson’s correlation coefficient and 0.59 bpm MAE, 1.77 bpm RMSE, 0.99 r value of Pearson’s correlation coefficient on the PURE and UBFC-rPPG datasets, respectively. Compared to the baseline methods, it achieves the lowest MAE and RMSE, and the highest r.
- 6 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (19) , 30275-30286
CMRPPGFormer: 3-D Spatio-Temporal Convolutional Modulation Transformer Network for Remote Heart Rate Estimation
Most of the existing remote heart rate (HR) estimation methods based on deep learning networks use convolutional neural networks which excel at local feature extraction. This paper proposes a HR estimation method based on a vision transformer network named as CMRPPGFormer. The basic idea of the proposed method is to extract precise remote photoplethysmography (rPPG) signals from facial videos by the CMRPPGFormer and then calculate relevant indicators of HR by specific post-processing on rPPG signals. The CMRPPGFormer consists of the convolutional stem block, the feature fusion block and the rPPG predictor block. The feature fusion block, the key of the CMRPPGFormer, can effectively extract and fuse local and global features of rPPG signals, in which a three dimensional (3D) spatio-temporal convolutional modulation transformer (3D-ST-CMTransformer) model is proposed. The 3D-ST-CMTransformer extracts global features of rPPG signals by its 3D convolution modulation attention layer and feed-forward network effectively. Experiment results on two public datasets, PURE and UBFC-rPPG, validate the better performance of the proposed method compared to the state-of-the-art baseline methods. Particularly, CMRPPGFormer obtains 0.39 bpm MAE, 0.64 bpm RMSE, 0.99 r value of Pearson’s correlation coefficient and 0.59 bpm MAE, 1.77 bpm RMSE, 0.99 r value of Pearson’s correlation coefficient on the PURE and UBFC-rPPG datasets, respectively. Compared to the baseline methods, it achieves the lowest MAE and RMSE, and the highest r.
- 31 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (13) , 15199-15209
Shuffle-rPPGNet: Efficient Network With Global Context for Remote Heart Rate Variability Measurement
In the field of remote physiological indicator measurement, there are relatively few studies on measuring heart rate variability (HRV). Moreover, some methods using deep neural networks are computationally intensive due to high network complexity. In this paper, a HRV measurement method is proposed based on an efficient three dimensional convolutional neural network (3DCNN) model defined as Shuffle-rPPGNet. The basic idea of the proposed method is to extract more precise remote photoplethysmography (rPPG) signals from facial videos by this model and then calculate relevant indicators of HRV through the post-processed rPPG signals. Shuffle-rPPGNet uses a lightweight network structure named ShuffleNetV2 as a baseline to construct an efficient 3DCNN with additional parts including 3D Global Context, 3D Channel Shuffle and Up-sample blocks, etc. The proposed method has been evaluated on two public datasets, i.e., UBFC-rPPG and PURE datasets. The results show that the proposed method has better performance in HRV measurement compared to the state-of-the-art methods. Especially, the mean absolute errors of the average values of the normal-to-normal interval sequence reach 2.35 and 6.62 respectively on these two public datasets.
- 31 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 23 (13) , 15199-15209
Shuffle-rPPGNet: Efficient Network With Global Context for Remote Heart Rate Variability Measurement
In the field of remote physiological indicator measurement, there are relatively few studies on measuring heart rate variability (HRV). Moreover, some methods using deep neural networks are computationally intensive due to high network complexity. In this paper, a HRV measurement method is proposed based on an efficient three dimensional convolutional neural network (3DCNN) model defined as Shuffle-rPPGNet. The basic idea of the proposed method is to extract more precise remote photoplethysmography (rPPG) signals from facial videos by this model and then calculate relevant indicators of HRV through the post-processed rPPG signals. Shuffle-rPPGNet uses a lightweight network structure named ShuffleNetV2 as a baseline to construct an efficient 3DCNN with additional parts including 3D Global Context, 3D Channel Shuffle and Up-sample blocks, etc. The proposed method has been evaluated on two public datasets, i.e., UBFC-rPPG and PURE datasets. The results show that the proposed method has better performance in HRV measurement compared to the state-of-the-art methods. Especially, the mean absolute errors of the average values of the normal-to-normal interval sequence reach 2.35 and 6.62 respectively on these two public datasets.
- Open Access
- 26 April 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (9) , 4364
A Chinese Grammatical Error Correction Method Based on Iterative Training and Sequence Tagging
Chinese grammatical error correction (GEC) is under continuous development and improvement, and this is a challenging task in the field of natural language processing due to the high complexity and flexibility of Chinese grammar. Nowadays, the iterative sequence tagging approach is widely applied to Chinese GEC tasks because it has a faster inference speed than sequence generation approaches. However, the training phase of the iterative sequence tagging approach uses sentences for only one round, while the inference phase is an iterative process. This makes the model focus only on the current sentence’s current error correction results rather than considering the results after multiple rounds of correction. In order to address this problem of mismatch between the training and inference processes, we propose a Chinese GEC method based on iterative training and sequence tagging (CGEC-IT). First, in the iterative training phase, we dynamically generate the target tags for each round by using the final target sentences and the input sentences of the current round. The final loss is the average of each round’s loss. Next, by adding conditional random fields for sequence labeling, we ensure that the model pays more attention to the overall labeling results. In addition, we use the focal loss to solve the problem of category imbalance caused by the fact that most words in text error correction do not need error correction. Furthermore, the experiments on NLPCC 2018 Task 2 show that our method outperforms prior work by up to 2% on the F0.5 score, which verifies the efficiency of iterative training on the Chinese GEC model.
- 26 April 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (9) , 4364
Open AccessA Chinese Grammatical Error Correction Method Based on Iterative Training and Sequence Tagging
Chinese grammatical error correction (GEC) is under continuous development and improvement, and this is a challenging task in the field of natural language processing due to the high complexity and flexibility of Chinese grammar. Nowadays, the iterative sequence tagging approach is widely applied to Chinese GEC tasks because it has a faster inference speed than sequence generation approaches. However, the training phase of the iterative sequence tagging approach uses sentences for only one round, while the inference phase is an iterative process. This makes the model focus only on the current sentence’s current error correction results rather than considering the results after multiple rounds of correction. In order to address this problem of mismatch between the training and inference processes, we propose a Chinese GEC method based on iterative training and sequence tagging (CGEC-IT). First, in the iterative training phase, we dynamically generate the target tags for each round by using the final target sentences and the input sentences of the current round. The final loss is the average of each round’s loss. Next, by adding conditional random fields for sequence labeling, we ensure that the model pays more attention to the overall labeling results. In addition, we use the focal loss to solve the problem of category imbalance caused by the fact that most words in text error correction do not need error correction. Furthermore, the experiments on NLPCC 2018 Task 2 show that our method outperforms prior work by up to 2% on the F0.5 score, which verifies the efficiency of iterative training on the Chinese GEC model.
- Open Access
- 17 February 2022
- Published by MDPI AG in Electronics
- Vol. 11 (4) , 627
A Stereo Calibration Method of Multi-Camera Based on Circular Calibration Board
In the application of 3D reconstruction of multi-cameras, it is necessary to calibrate the camera used separately, and at the same time carry out multi-stereo calibration, and the calibration accuracy directly affects the effect of the 3D reconstruction of the system. Many researchers focus on the optimization of the calibration algorithm and the improvement of calibration accuracy after obtaining the calibration plate pattern coordinates, ignoring the impact of calibration on the accuracy of the calibration board pattern coordinate extraction. Therefore, this paper proposes a multi-camera stereo calibration method based on circular calibration plate focusing on the extraction of pattern features during the calibration process. This method preforms the acquisition of the subpixel edge acquisition based on Franklin matrix and circular feature extraction of the circular calibration plate pattern collected by the camera, and then combines the Zhang’s calibration method to calibrate the camera. Experimental results show that compared with the traditional calibration method, the method has better calibration effect and calibration accuracy, and the average reprojection error of the multi-camera is reduced by more than 0.006 pixels.
- 17 February 2022
- Published by MDPI AG in Electronics
- Vol. 11 (4) , 627
Open AccessA Stereo Calibration Method of Multi-Camera Based on Circular Calibration Board
In the application of 3D reconstruction of multi-cameras, it is necessary to calibrate the camera used separately, and at the same time carry out multi-stereo calibration, and the calibration accuracy directly affects the effect of the 3D reconstruction of the system. Many researchers focus on the optimization of the calibration algorithm and the improvement of calibration accuracy after obtaining the calibration plate pattern coordinates, ignoring the impact of calibration on the accuracy of the calibration board pattern coordinate extraction. Therefore, this paper proposes a multi-camera stereo calibration method based on circular calibration plate focusing on the extraction of pattern features during the calibration process. This method preforms the acquisition of the subpixel edge acquisition based on Franklin matrix and circular feature extraction of the circular calibration plate pattern collected by the camera, and then combines the Zhang’s calibration method to calibrate the camera. Experimental results show that compared with the traditional calibration method, the method has better calibration effect and calibration accuracy, and the average reprojection error of the multi-camera is reduced by more than 0.006 pixels.
- Open Access
- 2 February 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (3) , 1599
A Keyword Detection and Context Filtering Method for Document Level Relation Extraction
Relation extraction (RE) is the core link of downstream tasks, such as information retrieval, question answering systems, and knowledge graphs. Most of the current mainstream RE technologies focus on the sentence-level corpus, which has great limitations in practical applications. Moreover, the previously proposed models based on graph neural networks or transformers try to obtain context features from the global text, ignoring the importance of local features. In practice, the relation between entity pairs can usually be inferred just through a few keywords. This paper proposes a keyword detection and context filtering method based on the Self-Attention mechanism for document-level RE. In addition, a Self-Attention Memory (SAM) module in ConvLSTM is introduced to process the document context and capture keyword features. By searching for word embeddings with high cross-attention of entity pairs, we update and record critical local features to enhance the performance of the final classification model. The experimental results on three benchmark datasets (DocRED, CDR, and GBA) show that our model achieves advanced performance within open and specialized domain relationship extraction tasks, with up to 0.87% F1 value improvement compared to the state-of-the-art methods. We have also designed experiments to demonstrate that our model can achieve superior results by its stronger contextual filtering capability compared to other methods.
- 2 February 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (3) , 1599
Open AccessA Keyword Detection and Context Filtering Method for Document Level Relation Extraction
Relation extraction (RE) is the core link of downstream tasks, such as information retrieval, question answering systems, and knowledge graphs. Most of the current mainstream RE technologies focus on the sentence-level corpus, which has great limitations in practical applications. Moreover, the previously proposed models based on graph neural networks or transformers try to obtain context features from the global text, ignoring the importance of local features. In practice, the relation between entity pairs can usually be inferred just through a few keywords. This paper proposes a keyword detection and context filtering method based on the Self-Attention mechanism for document-level RE. In addition, a Self-Attention Memory (SAM) module in ConvLSTM is introduced to process the document context and capture keyword features. By searching for word embeddings with high cross-attention of entity pairs, we update and record critical local features to enhance the performance of the final classification model. The experimental results on three benchmark datasets (DocRED, CDR, and GBA) show that our model achieves advanced performance within open and specialized domain relationship extraction tasks, with up to 0.87% F1 value improvement compared to the state-of-the-art methods. We have also designed experiments to demonstrate that our model can achieve superior results by its stronger contextual filtering capability compared to other methods.