Article
Environmental Sciences
Jie Liu, Xin Cao, Pingchuan Zhang, Xueli Xu, Yangyang Liu, Guohua Geng, Fengjun Zhao, Kang Li, Mingquan Zhou
Summary: AMS-Net is an attention-based multi-scale neural network proposed for extracting significant geometric and semantic features of Terracotta Warrior fragments for effective classification. With a multi-scale strategy, it can extract local and global features from different scales in parallel, enhancing accuracy and performance in classification tasks.
Article
Computer Science, Information Systems
Wenmin Yao, Tong Chu, Wenlong Tang, Jingyu Wang, Xin Cao, Fengjun Zhao, Kang Li, Guohua Geng, Mingquan Zhou
Summary: This paper introduces a fracture-surface-based reassembling method named SPPD for Terracotta Warrior fragments, which outperforms conventional methods in real-world experiments and could be a valuable tool for virtual restoration of cultural heritage artifacts.
ISPRS INTERNATIONAL JOURNAL OF GEO-INFORMATION
(2021)
Article
Computer Science, Information Systems
Yao Hu, Guohua Geng, Kang Li, Bao Guo, Pengbo Zhou
Summary: This study focuses on cultural relics restoration and fragment splicing research, proposing the EGG-Net method to automatically calibrate the terracotta warrior dataset. EGG-Net is a self-supervised model that extracts features and segments point cloud data through modular steps, achieving better results than existing methods.
Article
Optics
Jie Liu, Da Song, Guohua Geng, Yu Tian, Mengna Yang, Yangyang Liu, Mingquan Zhou, Kang Li, Xin Cao
Summary: In this paper, a task-driven and learnable down-sampling method named TGPS is proposed to address the issue of excessive redundant data in the dense point clouds of Terracotta Warriors obtained by a 3D scanner. The method utilizes a point-based Transformer unit to embed features and a mapping function to extract input point features for dynamic representation of global features. The contribution of each point to the global feature is estimated using the inner product between the global feature and each point feature, and high-similarity point features are retained based on descending contribution values. A Dynamic Graph Attention Edge Convolution (DGA EConv) is proposed for local feature aggregation, combined with graph convolution operation. Networks for point cloud classification and reconstruction are presented as downstream tasks. Experimental results demonstrate that the method achieves downsampling guided by global features, with TGPS-DGA-Net achieving the best accuracy on both real-world Terracotta Warrior fragments and public datasets.
Article
Computer Science, Software Engineering
Wenxiao Zhang, Huajian Zhou, Zhen Dong, Jun Liu, Qingan Yan, Chunxia Xiao
Summary: Point cloud shape completion is important in 3D vision and robotics applications. Early methods generated global shapes without refining local details. Current methods use local features to preserve observed geometric details, but they ignore long-distance correlation between skeleton and details. In this work, we propose a coarse-to-fine completion framework that leverages neighboring and long-distance cues, and introduces a Skeleton-Detail Transformer and selective attention mechanism. Experimental results show that our network outperforms state-of-the-art methods.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS
(2023)
Review
Environmental Sciences
Su Yang, Shishuo Xu, Wei Huang
Summary: This study conducted a comprehensive systematic survey of recent literature on the use of three-dimensional point cloud data in cultural heritage field, utilizing complex network analysis methods. By constructing various networks, it revealed research hotspots, national collaborations, interdisciplinary patterns, and emerging trends.
Article
Computer Science, Artificial Intelligence
Rajkumar Gothandaraman, Sreekumar Muthuswamy
Summary: This study presents a methodology for detecting symmetry in 3D objects, with Eigenvalues and local surface discontinuity showing better performance in complex models, increasing the accuracy of symmetry estimation.
JOURNAL OF REAL-TIME IMAGE PROCESSING
(2021)
Article
Environmental Sciences
Ming Wei, Ming Zhu, Yaoyuan Zhang, Jiaqi Sun, Jiarong Wang
Summary: The application of 3D scenes has been expanding in recent years, but the reliability of 3D point clouds acquired using sensors is limited, causing difficulties in their utilization. To address this issue, point cloud completion techniques can reconstruct and restore sparse and incomplete point clouds to enhance their realism. In this study, we propose a cyclic global guiding network structure that considers both local details and overall characteristics of the whole cloud for point cloud completion tasks. We introduce fitting planes and layered folding attention modules based on global guidance to strengthen the local effect. Experimental results demonstrate the effectiveness of our method on diverse datasets and its superiority over other networks.
Article
Environmental Sciences
Ruidong Hao, Zhonghui Wei, Xu He, Kaifeng Zhu, Jun Wang, Jiawei He, Lei Zhang
Summary: This paper presents a novel network, MAPGNet, for adaptive point cloud growth, which generates a sparse skeletal point cloud using skeletal features and then adaptively grows local point clouds in the spherical neighborhood to complement the details. Experimental results demonstrate that MAPGNet has advantages in point cloud completion.
Article
Environmental Sciences
Jiabo Xu, Zeyun Wan, Jingbo Wei
Summary: This paper presents a new end-to-end neural network for point cloud completion. The proposed method selects regular voxel centers as reference points and designs the encoder and decoder with Patchify, transformers, and multilayer perceptrons. An implicit classifier is incorporated in the decoder to mark valid voxels for diffusion. The effectiveness of the proposed model is validated through experiments on various datasets, showing improved accuracy and detail in predicting point coordinates with uniform distributions.
Article
Environmental Sciences
Ming Wei, Jiaqi Sun, Yaoyuan Zhang, Ming Zhu, Haitao Nie, Huiying Liu, Jiarong Wang
Summary: The paper proposes a network model for point cloud completion task, which improves the accuracy of reconstruction and the quality of detail recovery. Experimental results demonstrate that our network outperforms other networks on ShapeNet and Complete3D datasets.
Article
Computer Science, Artificial Intelligence
Ivan Sipiran, Alexis Mendoza, Alexander Apaza, Cristian Lopez
Summary: The Josefina Ramos de Cox museum in Lima, Peru, decided to digitize archaeological pieces to support research and education. However, the 3D scanning process caused imperfections in the objects' surface. This paper proposes a data-driven method to repair the digital objects' surface.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2022)
Article
Environmental Sciences
Pedro Martin-Lerones, David Olmedo, Ana Lopez-Vidal, Jaime Gomez-Garcia-Bermejo, Eduardo Zalama
Summary: This paper discusses the process of processing 3D point cloud data for use in BIM software to improve efficiency and accuracy in heritage asset management. Specific H-BIM specifications have been developed for heritage projects, and non-proprietary formats like IFC are used for information exchange.
Article
Engineering, Multidisciplinary
Michalina Wojtkowska, Michal Kedzierski, Paulina Delis
Summary: This study discusses the use of artificial neural networks and point clouds to calculate displacements of cultural heritage structures. The model trained on a laboratory dataset was able to determine displacements of the building facade with a relative accuracy of 3% and a success rate of 85%. Deformations derived from digital surface models generated from point clouds had a relative accuracy of 7%, while values determined by image-based close-range photogrammetry methods were 35%. An innovative aspect is the use of neural networks to determine deformations based on sub-models generated from the point cloud, along with a supervised-trained high accuracy predictive model. The practical significance lies in creating an end-to-end solution that can automatically detect and estimate the value of deformation, providing a major advantage over other methods.
Review
Construction & Building Technology
Victoria Andrea Cotella
Summary: Interest in semantic segmentation of 3D point clouds using ML and DL has grown due to their key role in scene insight across various applications. However, there is a research gap regarding the interface between point cloud segmentation and the HBIM workflow. This study aims to perform a systematic review of the current bibliography to advance innovative strategies in the field of BIM and AI.
AUTOMATION IN CONSTRUCTION
(2023)
Article
Optics
Xueli Xu, Guohua Geng, Xin Cao, Kang Li, Mingquan Zhou
Summary: This study proposes a transformer-based end-to-end network (TDNet) for point cloud denoising. The encoder utilizes the structure of a transformer in natural language processing to extract features and transform the point cloud. The decoder learns the latent manifold of each sampled point, resulting in a clean point cloud. An adaptive sampling approach is introduced to reconstruct the surface. Extensive experiments demonstrate the superiority of the proposed network.
Article
Optics
Jie Liu, Yu Tian, Guohua Geng, Haolin Wang, Da Song, Kang Li, Mingquan Zhou, Xin Cao
Summary: In this paper, a novel unsupervised representation learning network, UMA-Net, is proposed for downstream 3D object classification. Experimental results show that the model achieves comparable performance in 3D object classification tasks, narrowing the gap between unsupervised and supervised learning approaches.
JOURNAL OF THE OPTICAL SOCIETY OF AMERICA A-OPTICS IMAGE SCIENCE AND VISION
(2022)
Article
Optics
Haolin Wang, Zhonghao Wang, Jingle Wang, Kang Li, Guohua Geng, Fei Kang, Xin Cao
Summary: This paper proposes an improved U-net network, ICA-Unet, for automatic and precise segmentation of brown adipose tissue (BAT). By introducing depth-wise over-parameterized convolutional layers, channel attention blocks, and image information entropy blocks, the method achieves excellent segmentation results on the PET/CT images of 368 patients.
JOURNAL OF INNOVATIVE OPTICAL HEALTH SCIENCES
(2022)
Article
Pharmacology & Pharmacy
Weicheng Huang, Jingyi Wang, Haolin Wang, Yuxiang Zhang, Fengjun Zhao, Kang Li, Linzhi Su, Fei Kang, Xin Cao
Summary: This study compared the performance of radiomics and deep learning in predicting EGFR mutation status in lung cancer patients based on PET/CT images. A hybrid model combining deep learning features and a clinically significant characteristic achieved better diagnostic performance than radiomics and deep learning models alone, providing more personalized treatment options for NSCLC patients.
FRONTIERS IN PHARMACOLOGY
(2022)
Article
Pharmacology & Pharmacy
Jingyi Wang, Xing Lv, Weicheng Huang, Zhiyong Quan, Guiyu Li, Shuo Wu, Yirong Wang, Zhaojuan Xie, Yuhao Yan, Xiang Li, Wenhui Ma, Weidong Yang, Xin Cao, Fei Kang, Jing Wang
Summary: This study assessed the significance of mutation mutual exclusion information in the optimization of radiomics algorithms for predicting gene mutations and developed a composite model combining mutation information to improve prediction accuracy.
FRONTIERS IN PHARMACOLOGY
(2022)
Article
Multidisciplinary Sciences
Xueli Xu, Kang Li, Yifei Ma, Guohua Geng, Jingyu Wang, Mingquan Zhou, Xin Cao
Summary: This study proposes a point cloud simplification framework that uses a virtual camera to obtain multi-angle images, extracts feature lines using deep neural networks, automatically extracts feature points of the point cloud based on a mapping relationship, and ultimately obtains a simplified point cloud. Experimental results demonstrate the superiority of this method in retaining geometric features and achieving a high simplification rate.
SCIENTIFIC REPORTS
(2022)
Article
Physics, Applied
Mengfei Du, Yi Chen, Weitong Li, Linzhi Su, Huangjian Yi, Fengjun Zhao, Kang Li, Lin Wang, Xin Cao
Summary: Cerenkov luminescence tomography (CLT) is a non-invasive technique for three-dimensional detection of radiopharmaceuticals. The proposed multi-stage cascade neural network improves the performance of CLT reconstruction by introducing an attention mechanism and a special constraint. Numerical simulations and in vivo experiments demonstrate that this method achieves superior accuracy and shape recovery capability compared to existing methods.
JOURNAL OF APPLIED PHYSICS
(2022)
Article
Optics
Yu Tian, Da Song, Mengna Yang, Jie Liu, Guohua Geng, Mingquan Zhou, Kang Li, Xin Cao
Summary: In this paper, we propose ULD-Net, an unsupervised learning approach for point cloud analysis. We introduce a dense similarity learning method that achieves consistency across global-local views. Our ULD-Net outperforms context-based unsupervised methods and achieves comparable performances to supervised models in shape classification and segmentation tasks.
JOURNAL OF THE OPTICAL SOCIETY OF AMERICA A-OPTICS IMAGE SCIENCE AND VISION
(2022)
Article
Optics
Yi Chen, Weitong Li, Mengfei Du, Linzhi Su, Huangjian Yi, Fengjun Zhao, Kang Li, Lin Wang, Xin Cao
Summary: This study proposes a new method for CLT reconstruction, which improves spatial location accuracy and shape recovery ability through non-negative iterative three operator splitting strategy and elastic net regularization. Experimental results demonstrate superior performance in terms of location accuracy, shape recovery capability, and robustness.
Article
Optics
Weitong Li, Mengfei Du, Yi Chen, Haolin Wang, Linzhi Su, Huangjian Yi, Fengjun Zhao, Kang Li, Lin Wang, Xin Cao
Summary: Cerenkov Luminescence Tomography (CLT) is a novel imaging modality that can display the three-dimensional distribution of radioactive probes. Traditional model-based methods face challenges in obtaining accurate reconstruction results due to severe ill-posed inverse problem. Deep learning-based methods have emerged as a solution to improve the performance by directly learning the mapping relation between surface photon intensity and radioactive source distribution.
JOURNAL OF INNOVATIVE OPTICAL HEALTH SCIENCES
(2023)
Article
Optics
Yi Chen, Mengfei Du, Gege Zhang, Jun Zhang, Kang Li, Linzhi Su, Fengjun Zhao, Huangjian Yi, And Xin Cao
Summary: Cone-beam X-ray luminescence computed tomography (CB-XLCT) is a promising dual-modal imaging technology for early tumor detection. However, the low absorption and high scattering of light in tissues pose challenges in CB-XLCT reconstruction. A proposed strategy using dictionary learning and group structure (DLGS) effectively addresses these challenges and achieves superior CB-XLCT reconstruction performance. Experimental results demonstrate the method's accuracy, target shape, robustness, dual-source resolution, and in vivo applicability.
Article
Optics
Jie Liu, Da Song, Guohua Geng, Yu Tian, Mengna Yang, Yangyang Liu, Mingquan Zhou, Kang Li, Xin Cao
Summary: In this paper, a task-driven and learnable down-sampling method named TGPS is proposed to address the issue of excessive redundant data in the dense point clouds of Terracotta Warriors obtained by a 3D scanner. The method utilizes a point-based Transformer unit to embed features and a mapping function to extract input point features for dynamic representation of global features. The contribution of each point to the global feature is estimated using the inner product between the global feature and each point feature, and high-similarity point features are retained based on descending contribution values. A Dynamic Graph Attention Edge Convolution (DGA EConv) is proposed for local feature aggregation, combined with graph convolution operation. Networks for point cloud classification and reconstruction are presented as downstream tasks. Experimental results demonstrate that the method achieves downsampling guided by global features, with TGPS-DGA-Net achieving the best accuracy on both real-world Terracotta Warrior fragments and public datasets.
Article
Optics
Yi Chen, Mengfei Du, Jun Zhang, Gege Zhang, Linzhi Su, Kang Li, Fengjun Zhao, HuangJian Yi, Lin Wang, Xin Cao
Summary: Fluorescence molecular tomography (FMT) is an optical imaging technology for visualizing the three-dimensional distribution of fluorescently labelled probes in vivo. Improving FMT reconstruction is challenging due to light scattering and ill-posed inverse problems. This study proposes a generalized conditional gradient method with adaptive regularization parameters (GCGM-ARP) to enhance FMT reconstruction performance. The GCGM-ARP method combines elastic-net (EN) regularization and adaptive adjustment of regularization parameters to achieve more accurate and robust reconstructions. Experimental results demonstrate that the GCGM-ARP method outperforms other reconstruction methods, exhibiting superior performance in source localization, dual-source resolution, morphology recovery, and robustness.
Article
Optics
Minghua Zhao, Yahui Xiao, Jiaqi Zhang, Xin Cao, Lin Wang
Summary: Optical molecular tomography is a promising pre-clinical molecular imaging technique that provides 3D information about tumor distribution. This study introduces a new method that improves reconstruction results by establishing the mapping relationship between internal source distribution and surface photon density.
JOURNAL OF INNOVATIVE OPTICAL HEALTH SCIENCES
(2023)