HSHao Sun

Hao Sun

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Hao Sun has a total of 98 co-authors. Here are the top co-authors by number of shared publications:

  • Jianwen Chen

    Limb Deformity Correction and Function Reconstruction Department Southern University of Science and Technology Hospital Shenzhen China

    7 shared publications
  • Yue Guo

    Limb Deformity Correction and Function Reconstruction Department Southern University of Science and Technology Hospital Shenzhen China

    7 shared publications
  • Jian Li

    Beijing University of Posts and Telecommunications

    6 shared publications
  • Lunhui Duan

    School of Artificial Intelligence and Data Science and Engineering Research Center of Intelligent Rehabilitation Device and Detection Technology Ministry of Education, Hebei University of Technology Tianjin China

    5 shared publications
  • Mengting Wei

    School of Artificial Intelligence and Data Science, Hebei University of Technology, No. 8 Guangrong Road, Hong Qiao, Tianjin, 300130, China

    4 shared publications
  • Peng Yang

    School of Artificial Intelligence, Hebei University of Technology, Tianjin 300130, China

    4 shared publications

Latest Publications

  • MRF ‐ DPUNet ++: An Improved U‐Net++ Network for Bone Segmentation in CT Images of Tibial and Fibular Fractures

    by Yunqu Sun, Hao Song, Xianzhe Zhang, Jiahui Li, Fuhai Pan, Hao Sun

    Accurate bone segmentation in CT images of tibial and fibular fractures is essential for fracture‐related three‐dimensional reconstruction, preoperative planning, and robot‐assisted reduction. However, the tibia and fibula are elongated bones with complex boundaries, and fracture cases often involve cortical disruption, displacement, fragment overlap, and imaging artifacts. This study aimed to develop an improved U‐Net++ network for automatic bone segmentation in CT images of tibial and fibular fractures. We proposed MRF‐DPUNet++, which incorporates a Depthwise Multi‐Receptive‐Field Block (DMRF Block) into the encoder and a Dual‐Pooling Feature Enhancement Module (DPFEM) before the prediction layer. The DMRF Block uses multi‐branch depthwise dilated convolutions to enhance multi‐scale bone‐structure representation, whereas DPFEM combines average‐pooling and max‐pooling responses to improve feature discrimination immediately before prediction. Experiments were conducted on preoperative CT data from 45 patients with tibial and fibular fractures. MRF‐DPUNet++ achieved a Dice score of 94.86% and an IoU of 93.09% on the tibial and fibular fracture CT segmentation task, outperforming the comparative methods in the main overlap metrics. Ablation experiments showed that both the DMRF Block and DPFEM contributed to improved segmentation performance. The proposed MRF‐DPUNet++ improves bone‐region segmentation accuracy and prediction‐feature discrimination in CT images of tibial and fibular fractures, providing segmentation support for fracture‐region analysis, three‐dimensional reconstruction, and robot‐assisted reduction.

  • Real-Time Dynamic Gesture Recognition Based on Ultrasound: For Human–Robot Interaction

    by Chengming Tian, Qingying He, Fangyu Shen, Hao Sun

    In human-computer interaction based on gesture recognition, visual schemes are susceptible to interference from light and the background, and wearable sensors restrict hand movement. The aforementioned issues have a restrictive effect on the interaction experience. The present paper proposes an ultrasound-based, real-time, dynamic gesture recognition scheme to address these problems. The extraction of dynamic gesture features is achieved through the utilisation of the channel estimation matrix diagram (CEM-D), while algorithms designed for onset and motion detection ensure the accurate acquisition of features. The classification of gestures is accomplished through the utilization of a RepVGG network, which facilitates the construction of an end-to-end recognition system. On a dataset comprising ten kinds of daily dynamic gestures, the method achieves an accuracy of over 95% in environments with noise levels of 30 dB and 70 dB, and takes an average of 0.0483s for a single recognition, which is 0.00969s faster than the traditional gesture recognition method. The findings indicate that the proposed method exhibits exceptional accuracy, anti-interference capabilities and real-time performance, thus providing a highly efficient solution for human-computer interaction involving dynamic gestures.


  • IcoTag3D: Enhanced 6-DoF Pose Estimation for Robotic Arms Using TriangleTag Markers on an Icosahedron

    by Qingying He, Xiao Li, Chengming Tian, Fangyu Shen, Yuanyuan Liu, Hao Sun

    High-precision pose estimation using fiducial markers has many applications in medical device tracking, virtual reality alignment, navigation, and more. However, the accuracy of pose estimation and detection capabilities are often constrained by the shape and scale of the fiducial marker plane. In this article, we propose a triangular planar fiducial marker affixed to a positive icosahedron for pose estimation. This design expands the angular observation range, increases the marker scale, and consequently enhances estimation accuracy and recognition distance. The 2-D coordinates of the feature points from the markers are detected and extracted from the environment. Subsequently, the 3-D coordinates of these feature points are obtained using the triangulation method. This process results in the formation of 2-D–3-D point pairs. High-quality interior points are then filtered using the random sample consensus (RANSAC) method. The initial position is determined through the efficient perspective-n-point (EPnP) method, followed by the application of Levenberg–Marquardt (LM) optimization. We evaluated the performance of IcoTag3D through both simulations and physical experiments. The results from the simulation experiments indicate that IcoTag3D exhibits significantly lower maximum rotation angle error, reprojection error, and translation error at the submillimeter level. In addition, it demonstrates an improved recognition distance compared with the method of attaching ArUco markers to icosahedra. Physical experiments have further confirmed the feasibility of IcoTag3D.


  • An Automatic Measurement Method of the Tibial Deformity Angle on X‐Ray Films Based on Deep Learning Keypoint Detection Network

    by Ning Zhao, Cheng Chang, Yuanyuan Liu, Xiao Li, Zicheng Song, Yue Guo, Jianwen Chen, Hao Sun

    In the clinical application of the parallel external fixator, medical practitioners are required to quantify deformity parameters to develop corrective strategies. However, manual measurement of deformity angles is a complex and time‐consuming process that is susceptible to subjective factors, resulting in nonreproducible results. Accordingly, this study proposes an automatic measurement method based on deep learning, comprising three stages: tibial segment localization, tibial contour point detection, and deformity angle calculation. First, the Faster R‐CNN object detection model, combined with ResNet50 and FPN as the backbone, was employed to achieve accurate localization of tibial segments under both occluded and nonoccluded conditions. Subsequently, a relative position constraint loss function was added, and ResNet101 was used as the backbone, resulting in an improved RTMPose keypoint detection model that achieved precise detection of tibial contour points. Ultimately, the bone axes of each tibial segment were determined based on the coordinates of the contour points, and the deformity angles were calculated. The enhanced keypoint detection model, Con_RTMPose, elevated the Percentage of Correct Keypoints (PCK) from 63.94% of the initial model to 87.17%, markedly augmenting keypoint localization precision. Compared to manual measurements conducted by medical professionals, the proposed methodology demonstrates an average error of 0.52°, a maximum error of 1.15°, and a standard deviation of 0.07, thereby satisfying the requisite accuracy standards for orthopedic assessments. The measurement time is approximately 12 s, whereas manual measurement requires about 15 min, greatly reducing the time required. Additionally, the stability of the models was verified through K‐fold cross‐validation experiments. The proposed method meets the accuracy requirements for orthopedic applications, provides objective and reproducible results, significantly reduces the workload of medical professionals, and greatly improves efficiency.


  • Open Access

    Deep learning-based semantic segmentation of human features in bath scrubbing robots

    by Chao Zhuang, Tianyi Ma, Bokai Xuan, Cheng Chang, Baichuan An, Minghuan Yin, Hao Sun


  • Research on Abnormal Lung Sound Recognition and Diagnosis Based on Improved CNN and Transfomer

    by Yanli Geng, Zelei Li, Hao Sun

    Lung auscultation is one of the effective means to diagnose lung diseases. This study aims at the lung sounds of patients, and studies the diagnosis and recognition method of abnormal lung sounds in patients with lung diseases by combining convolutional neural network CNN (Convolutional Neural Network) and Transformer. In this study, a realtime lung sound acquisition system was designed, and the lung sound information was uploaded to the PC terminal by using the wireless WIFI module. Under the guidance of professional physicians, the collected lung sounds of patients were labeled to establish a feature data set (normal, light, common, and heavy). Secondly, based on the Butterworth filter Gaussian filter and the nonnegative matrix factorization method, the lung sound denoising was performed and the MFCC feature information of the lung sound was extracted. Finally, the lung sound auxiliary diagnosis model was designed to analyze the characteristics of lung sounds, and the patient's condition was diagnosed and evaluated. This paper proposes an SCBAM-CNN-Transformer classification model. After optimizing CNN by improving the channel attention mechanism CBAM, the improved CNN model is combined with Transformer. It combines the advantages of CNN spatial feature representation and Transformer sequence coding, forming a strong global feature perception and local feature extraction ability. The algorithm proposed in this paper has good robustness and generalization ability, and the recognition accuracy is 94.8 %. The experimental results show that the proposed algorithm can more effectively identify abnormal lung sounds and determine the infection level than other deep learning models.


  • Pain Expression Recognition Method Based on Fusion Attention Mechanism Capsule Network

    by Yanli Geng, Li Yin, Shijie Lu, Sun Hao

    Facial expression serves as a crucial indicator for pain recognition. However, existing approaches often face challenges in accurately extracting features from complex spatial information in facial expressions. To address this, this study introduces a pain recognition method based on a fusion attention mechanism capsule network. The main contributions of this paper are outlined as follows. Firstly, a comprehensive pain expression database is established by using video image sequences. Subsequently, the capsule attention mechanism is introduced after the primary capsule layer of the network, facilitating more effective feature extraction and improving the accuracy of information transmission within the capsule network. Additionally, the fully connected decoder in the capsule network is replaced by the deconvolution decoder, resulting in a reduction in the overall parameter count. Experimental results on the pain expression database demonstrate the efficacy of the proposed capsule network, achieving an impressive accuracy of 98.21%. Importantly, the proposed approach outperforms other deep learning models in terms of accuracy for pain recognition.

  • An automatic measurement method for ankle key angles based on point cloud segmentation network

    by Cheng Chang, Hao Sun, Xin Guo, Zhenhui Sun, Mengkun Wu, Minghuan Yin, Baichuan An, Chao Zhuang

    In clinical practice, anterior distal tibialangle (ADTA) and lateral distal tibialangle (LDTA) are important evaluation indexes, which can reflect the degree of deformity and determine the type of deformity to a certain extent. The measurement process of these parameters involves the localization of multiple key regions. To simplify the process of measuring ADTA and LDTA, an automatic measurement method is proposed. First, this study constructs an ankle 3D point cloud dataset by 3D reconstruction of ankle electron computed tomography (CT) data, and then, an ankle point cloud segmentation network with a Self‐Attention mechanism and kernel point convolution combined with each other is constructed, and the ankle point cloud data with labels are input to this network to predict the ankle bone feature points. Finally, the segmented feature points are fitted, and the tibial backbone axis and distal tibial articular surface can be drawn, and based on the fitting results, ADTA and LDTA are obtained. Compared with manual measurements by physicians, the proposed method in this study achieved an average error of 0.46° with a passing rate of 97.7% for ADTA measurement. For LDTA measurement, the average error was 0.69° with a passing rate of 98%. The ankle joint key angle measurement system developed in this study meets the accuracy requirements of orthopedics and significantly reduces the workload for doctors.


  • Open Access

    Sound Source Localization Indoors Based on Two-Level Reference Points Matching

    by Shuopeng Wang, Peng Yang, Hao Sun

    A dense sample point layout is the conventional approach to ensure the positioning accuracy for fingerprint-based sound source localization (SSL) indoors. However, mass reference point (RPs) matching of online phases may greatly reduce positioning efficiency. In response to this compelling problem, a two-level matching strategy is adopted to shrink the adjacent RPs searching scope. In the first-level matching process, two different methods are adopted to shrink the search scope of the online phase in a simple scene and a complex scene. According to the global range of high similarity between adjacent samples in a simple scene, a greedy search method is adopted for fast searching of the sub-database that contains the adjacent RPs. Simultaneously, in accordance with the specific local areas’ range of high similarity between adjacent samples in a complex scene, the clustering method is used for database partitioning, and the RPs search scope can be compressed by sub-database matching. Experimental results show that the two-level RPs matching strategy can effectively improve the RPs matching efficiency for the two different typical indoor scenes on the premise of ensuring the positioning accuracy.