Hui Sun

Hui Sun


Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 August 2021 in Sensors
Verified
Journal Article
Open access
PDF file
YZYanni Zhang
Yanni Zhang
Yiming LiuYiming Liu
Yiming Liu
QLQiang Li
Qiang Li
JWJianzhong Wang
Jianzhong Wang
MQMiao Qi
Miao Qi
JKJun Kong
Jun Kong
Recently, deep learning-based image deblurring and deraining have been well developed. However, most of these methods fail to distill the useful features. What is more, exploiting the detailed image features in a deep learning framework always requires a mass of parameters, which inevitably makes the network suffer from a high computational burden. We propose a lightweight fusion distillation network (LFDN) for image deblurring and deraining to solve the above problems. The proposed LFDN is designed as an encoder–decoder architecture. In the encoding stage, the image feature is reduced to various small-scale spaces for multi-scale information extraction and fusion without much information loss. Then, a feature distillation normalization block is designed at the beginning of the decoding stage, which enables the network to distill and screen valuable channel information of feature maps continuously. Besides, an information fusion strategy between distillation modules and feature channels is also carried out by the attention mechanism. By fusing different information in the proposed approach, our network can achieve state-of-the-art image deblurring and deraining results with a smaller number of parameters and outperform the existing methods in model complexity.

23 January 2023 in Applied Sciences
Verified
Journal Article
Open access
PDF file
RTRui Tang
Rui Tang
HSHui Sun
Hui Sun
DLDi Liu
Di Liu
HXHui Xu
Hui Xu
MQMiao Qi
Miao Qi
JKJun Kong
Jun Kong
Object detection has drawn the attention of many researchers due to its wide application in computer vision-related applications. In this paper, a novel model is proposed for object detection. Firstly, a new neck is designed for the proposed detection model, including an efficient SPPNet (Spatial Pyramid Pooling Network), a modified NLNet (Non Local Network) and a lightweight adaptive feature fusion module. Secondly, the detection head with double residual branch structure is presented to reduce the delay of a decoupled head and improve the detection ability. Finally, these improvements are embedded in YOLOX as plug-and-play modules for forming a high-performance detector, EYOLOX (EfficientYOLOX). Extensive experiments demonstrate that the EYOLOX achieves significant improvements, which increases YOLOX-s from 40.5% to 42.2% AP on the MS COCO dataset with a single GPU. Moreover, the performance of the detection of EYOLOX also outperforms YOLOv6 and some SOTA methods with the same number of parameters and GFLOPs. In particular, EYOLOX has only been trained on the COCO-2017 dataset without using any other datasets, and only the pre-training weights of the backbone part are loaded.

31 May 2022 in Mathematics
Verified
Journal Article
Open access
PDF file
SWSuhua Wang
Suhua Wang
HJHongjie Ji
Hongjie Ji
HSHui Sun
Hui Sun
YWYuling Wang
Yuling Wang
MLMengzhu Lu
Mengzhu Lu
HSHui Sun
Hui Sun
The goal of the recommender system is to learn the user’s preferences from the entity (user–item) historical interaction data, so as to predict the user’s ratings on new items or recommend new item sequences to users. There are two major challenges: (1) Datasets are usually sparse. The item side is often accompanied by some auxiliary information, such as attributes or context; it can help to slightly improve its representation. However, the user side is usually presented in the form of ID due to personal privacy. (2) Due to the influences of confounding factors, such as the popularity of items, users’ ratings on items often have bias that cannot be recognized by the traditional recommendation methods. In order to solve these two problems, in this paper, (1) we explore the use of a graph model to fuse the interactions between users and common rating items, that is, incorporating the “neighbor” information into the target user to enrich user representations; (2) the do() operator is used to deduce the causality after removing the influences of confounding factors, rather than the correlation of the data surface fitted by traditional machine learning. We propose the EGCI model, i.e., enhanced graph learning for recommendation via causal inference. The model embeds user relationships and item attributes into the latent semantic space to obtain high-quality user and item representations. In addition, the mixed bias implied in the rating process is calibrated by considering the popularity of items. Experimental results on three real-world datasets show that EGCI is significantly better than the baselines.

1 January 2023 in Mathematical Biosciences and Engineering
Verified
Journal Article
Open access
PDF file
CZCaixia Zheng
Caixia Zheng
HLHuican Li
Huican Li
YGYingying Ge
Yingying Ge
YHYanlin He
Yanlin He
YYYugen Yi
Yugen Yi
JKJun Kong
Jun Kong
Retinal vessel segmentation is very important for diagnosing and treating certain eye diseases. Recently, many deep learning-based retinal vessel segmentation methods have been proposed; however, there are still many shortcomings (e.g., they cannot obtain satisfactory results when dealing with cross-domain data or segmenting small blood vessels). To alleviate these problems and avoid overly complex models, we propose a novel network based on a multi-scale feature and style transfer (MSFST-NET) for retinal vessel segmentation. Specifically, we first construct a lightweight segmentation module named MSF-Net, which introduces the selective kernel (SK) module to increase the multi-scale feature extraction ability of the model to achieve improved small blood vessel segmentation. Then, to alleviate the problem of model performance degradation when segmenting cross-domain datasets, we propose a style transfer module and a pseudo-label learning strategy. The style transfer module is used to reduce the style difference between the source domain image and the target domain image to improve the segmentation performance for the target domain image. The pseudo-label learning strategy is designed to be combined with the style transfer module to further boost the generalization ability of the model. Moreover, we trained and tested our proposed MSFST-NET in experiments on the DRIVE and CHASE_DB1 datasets. The experimental results demonstrate that MSFST-NET can effectively improve the generalization ability of the model on cross-domain datasets and achieve improved retinal vessel segmentation results than other state-of-the-art methods.

1 January 2022 in Electronic Research Archive
Verified
Journal Article
Open access
PDF file
HXHui Xu
Hui Xu
JKJun Kong
Jun Kong
MLMengyao Liang
Mengyao Liang
HSHui Sun
Hui Sun
MQMiao Qi
Miao Qi
Human behavior recognition has always been a hot spot for research in computer vision. In this paper, we propose a novel video behavior recognition method based on Actional-Structural Graph Convolution and a Temporal Extension Module under the framework of a Spatio-Temporal Graph Convolution Neural Network, which can optimize the spatial and temporal features simultaneously. The basic network framework of our method consists of three parts: spatial graph convolution module, temporal extension module and attention mechanism module. In the spatial dimension, the action graph convolution is utilized to obtain abundant spatial features by capturing the correlations of distant joint features, and the structural graph convolution expands the existing skeleton graph to acquire the spatial features of adjacent joints. In the time dimension, the sampling range of the temporal graph is expanded for extracting the same and adjacent joints of adjacent frames. Furthermore, attention mechanisms are introduced to improve the performance of our method. In order to verify the effectiveness and accuracy of our method, a large number of experiments were carried out on two standard behavior recognition datasets: NTU-RGB+D and Kinetics. Comparative experiment results show that our proposed method can achieve better performance.

4 June 2025
Verified
Book Chapter
Open access
PDF file
SZShanshan Zhao
Shanshan Zhao
HSHui Sun
Hui Sun
XWXiuzhuo Wei
Xiuzhuo Wei
YSYao Shi
Yao Shi
The rapid proliferation of multimodal healthcare data, including electronic health records (EHRs), wearable device data, and medical imaging, presents significant opportunities for improving elderly chronic disease management. However, challenges such as data heterogeneity, high communication costs, and model robustness have hindered the effective utilization of these data. This study proposes an optimized federated learning (FL) framework designed to address these issues, enabling secure and efficient integration of multimodal big data across distributed healthcare institutions. The framework incorporates weighted aggregation to mitigate the impact of client-side data heterogeneity, gradient compression to reduce communication overhead, and federated adversarial training to improve model robustness under noisy and adversarial conditions. Experimental results on multimodal healthcare datasets demonstrate that the proposed framework outperforms traditional centralized and baseline FL approaches, achieving a 5% improvement in F1-score, a 40% reduction in communication costs, and enhanced model resilience to adversarial inputs. These findings highlight the framework’s potential to support early disease detection, personalized treatment, and chronic disease management, particularly in privacy-sensitive and resource-constrained environments. The study underscores the practical value of federated learning in real-world healthcare applications and provides a foundation for future research on advanced privacy-preserving techniques and large-scale deployment in clinical settings.