基于知识图谱驱动的图神经网络推荐模型

知识图谱(KG)蕴含丰富的结构与关联信息,不仅可以缓解推荐系统中数据稀疏、冷启动等问题,还可以更准确地进行个性化推荐,因此提出一种基于知识图谱驱动的端到端图神经网络推荐模型KGLN.首先使用单层神经网络框架对图中单个节点进行特征融合,并加入影响因子来改变不同邻居实体的聚合权重;然后通过迭代的方式将单层扩展到多层,使实体...     

基于层次意图解耦的图卷积神经网络推荐模型

当前意图推荐研究提取出的用户意图趋向扁平化,忽略了意图间的层次关系。针对以上问题,提出了一种基于层次意图解耦的图卷积神经网络推荐模型(HIDR),将用户—项目交互图划分为多个动态交互子图,以刻画从细粒度到粗粒度的用户意图层次图。首先,在每个意图交互子图中根据节点高阶连接性自适应地聚合来自高阶邻域的信息,解耦提取用户细粒度意图表示;然后,依据低层次细粒度意图之间的相似关系在高层网络上构建粗粒度意图超节点,显式建模从细粒度到粗粒度的意图层次结构;最后,将解耦得到的层次意图向量聚合为高质量的用户和项目表示,并进行内积预测和迭代优化。在Gowalla和Amazon-book两个数据集上的实验结果表明,相较于最优基线模型CLSR,HIDR的召回率(recall)分别提升了10.82%、6.63%,归一化折损累计增益(NDCG)分别提升了14.65%、9.63%,精度(precision)分别提升了10.46%和7.73%。     

融合多模态自监督图学习的视频推荐模型

现有视频推荐方法在算法框架中引入图神经网络来建模用户—视频协同关系,学习用户和视频的表示向量,但是节点中包含的冗余噪声会限制模型的建模能力。针对以上问题,提出了一种融合多模态自监督图学习的视频推荐模型(IMSGL-VRM)。首先,在图数据增强模式下构建自监督的图神经网络模型学习多模态视图下的节点特征表示,以提升节点表示的泛化能力;其次,为了得到推荐结果的多样性,设计了多兴趣提取模块从用户历史的交互视频序列中建模用户的多兴趣;最后,融合多模态的用户多兴趣表示和视频的特征表示,使用多样性可控的方式输出推荐结果,以满足视频推荐的多样性需求。在MovieLens-1M和TikTok数据集上实验,采用准确性、召回率、NDCG和多样性等指标评估模型。实验结果表明,该模型相比经典基准模型均有显著的性能提升。     

Learning concept hierarchies from textual resources for ontologies construction

Ontologies play a very important role in knowledge management and the Semantic Web, their use has been exploited in many current applications. Ontologies are especially useful because they support the exchange and sharing of information. Ontology learning from text is the process of deriving high-level concepts and their relations. An important task in ontology learning from text is to obtain a set of representative concepts to model a domain and organize them into a hierarchical structure (taxonomy) from unstructured information. In the process of building a taxonomy, the identification of hypernym/hyponym relations between terms is essential. How to automatically build the appropriate structure to represent the information contained in unstructured texts is a challenging task. This paper presents a novel method to obtain, from unstructured texts, representative concepts and their taxonomic relationships in a specific knowledge domain. This approach builds a concept hierarchy from a specific-domain corpus by using a clustering algorithm, a set of linguistic patterns, and additional contextual information extracted from the Web that improves the discovery of the most representative hypernym/hyponym relationships. A set of experiments were carried out using four different corpora. We evaluated the quality of the constructed taxonomies against gold standard ontologies, the experiments show promising results.     

Deep learning-based optimal segmentation of 3D printed product for surface quality improvement and support structure reduction

Large-sized product cannot be printed as one piece by a 3D printer because of the volume limitation of most 3D printers. Some products with the complex structure and high surface quality should also not be printed into one piece to meet requirement of the printing quality. For increasing the surface quality and reducing support structure of 3D printed models, this paper proposes a 3D model segmentation method based on deep learning. Sub-graphs are generated by pre-segmenting 3D triangular mesh models to extract printing features. A data structure is proposed to design training data sets based on the sub-graphs with printing features of the original 3D model including surface quality, support structure and normal curvature. After training a Stacked Auto-encoder using the training set, a 3D model is pre-segmented to build an application set by the sub-graph data structure. The application set is applied by the trained deep-learning system to generate hidden features. An Affinity Propagation clustering method is introduced in combining hidden features and geometric information of the application set to segment a product model into several parts. In the case study, samples of 3D models are segmented by the proposed method, and then printed using a 3D printer for validating the performance.     

3D facial model exaggeration builder for small or large sized model manufacturing

An action figure is a small human-shaped object used as a toy for children or artistic collection. In the past, the creation of action figures required intense manual effort. An initial trial to automate the process using recent scanning technology failed to yield figures with market appeal because the resulting action figures did not have sufficiently life-like shapes and expressions. The limiting factor which was not considered during this trial is the loss of individual characteristics resulting from either an increase or reduction in scale. We provide novel techniques for creating an exaggerated human face that retains all of the skin detail in the 3D scanned model, which will allow more characteristic figures to be easily created at any scale, thereby saving money and time during manufacturing. While traditional 3D printing applications utilize rigid models acquired using computer-aided design, our method demonstrates that deformable models (such as a human face) obtained from scanners are also suitable.

Retrieving 3D CAD model by freehand sketches for design reuse

Reuse of existing designs can help companies save up a significant amount of time and money in large-scale product development. Searching out similar CAD models of products is a key to facilitating design reuse. This paper proposes a method for retrieving CAD models of mechanical parts by freehand sketches. Users can draw three2D outline sketches and a skeleton sketch to express their query intent. The 2D outline sketches describe the geometrical information of models and the skeleton sketch contains the topological information. A relevance feedback mechanism is introduced to combine the two similarity degrees measured by 2D outlines or skeleton into a unified one. Experiments are conducted to evaluate the performance of the proposed method.     

非刚性三维模型检索特征提取技术研究

三维模型特征描述符是一种简洁且信息量丰富的表示方式.特征提取是许多三维模型分析处理任务的关键步骤.近年来,针对非刚性三维模型特征提取技术的研究引起了人们的广泛关注.本文首先汇总了常用的非刚性三维模型基准数据集和算法评价标准.然后在广泛调研大量文献和最新成果的基础上,将非刚性三维模型特征分为人工设计的特征描述符和基于学习的特征描述符两大类并分别加以介绍.对每类方法所包含的典型算法,尤其是最近几年基于深度学习的特征提取算法的基本思想、优缺点进行了分析、对比和总结.最后进行总结,并对未来可能的发展趋势加以展望.     

Deep learning-empowered digital twin for visualized weld joint growth monitoring and penetration control

This paper presents an innovative digital twin to monitor and control complex manufacturing processes by integrating deep learning which offers strong feature extraction and analysis abilities. Taking welding manufacturing as a case study, a deep learning-empowered digital twin is developed as the visualized digital replica of the physical welding for joint growth monitoring and penetration control. In such a system, the information available directly from sensors including weld pool images, arc images, welding current and arc voltage is collected in pulsed gas tungsten arc welding (GTAW-P). Then, the undirect information charactering the weld joint geometry and determining the welding quality, including the weld joint top-side bead width (TSBW) and back-side bead width (BSBW), is computed/estimated by traditional image processing methods and deep convolutional neural networks (CNNs) respectively. Compared with single image source, weld pool image or arc image, the CNN model performs better when taking the 2-channel composite image combined by both as the input and the state-of-the-art accuracy in BSBW prediction with mean square error (MSE) as 0.047 mm² is obtained. Then, a decision-making strategy is developed to control the welding penetration to meet the quality requirement and applied successfully in various welding conditions. By modeling the weld joint cross section as an ellipse, the developed digital twin is visualized to offer a graphical user interface (GUI) for users perceiving the weld joint growth intuitively and effectively.     

激光直接制造和再制造中的三维CAD模型直接分层技术

在分析SolidWorks软件平台下CAD模型数据的内部表达方法以及拓扑信息和几何信息提取方法的基础上,研究三维CAD模型直接分层技术.对SolidWorks进行二次开发,调用SolidWorks应用程序接口函数中的曲面一曲面求交函数对CAD模型曲面与分层平面求交,得到的交线首尾相连形成轮廓轨迹;同时研究了光栅填充扫描算法及程序实现.为实现切片数据的通用化,设计了记录切片数据的文件格式,用直线、圆弧或圆描述分层轮廓.对上述的直接分层不仅进行了软件模拟,还用于直接制造.制作的试件与STL间接分层试件比较结果表明,采用直接分层的试件的精度和表面质量优于STL间接分层.     

System for reconstruction of three-dimensional micro objects from multiple photographic images

We introduce a system to reconstruct a three-dimensiojnal (3D) polygonal model of 3D micro objects with outer dimensions ranging from several hundred microns to several millimeters from multiple two-dimensional (2D) images of an object taken from different views. The data acquisition system consists of a digital microscope that captures still images at a resolution of 1600 × 1200 pixels and a computer-controlled turntable. We employ the shape-from-silhouette (SFS) method to construct a voxel-based 3D model from silhouette images. The concave shapes are further carved by using the space carving technique. In order to make the resulting model compatible with a commercial CAD/CAM system, the voxel model is converted into a triangular mesh using the marching cubes algorithm. Because the mesh generated from the voxel model by using the marching cubes algorithm inherits the staircase effect, the mesh is adjusted to recover the object precisely by using silhouette images. Finally, we evaluate the accuracy of the proposed method. The reconstructed models of complex micro objects indicate the effectiveness of the 3D shape reconstruction system for micro objects.     

Manifold Constrained Transfer of Facial Geometric Knowledge for 3D Caricature Reconstruction

3D caricatures are important attractive elements of the interface in virtual environment such as online game. However, very limited 3D caricatures exist in the real world. Meanwhile, creating 3D caricatures manually is rather costly, and even professional skills are needed. This paper proposes a novel and effective manifold transfer algorithm to reconstruct 3D caricatures according to their original 2D caricatures. We first manually create a small dataset with only 100 3D caricature models and use them to initialize the whole 3D dataset. After that, manifold transfer algorithm is carried out to refine the dataset. The algorithm comprises of two steps. The first is to perform manifold alignment between 2D and 3D caricatures to get a "standard" manifold map; the second is to reconstruct all the 3D caricatures based on the manifold map. The proposed approach utilizes and transfers knowledge of 2D caricatures to the target 3D caricatures well. Comparative experiments show that the approach reconstructs 3D caricatures more effectively and the results conform more to the styles of the original 2D caricatures than the Principal Components Analysis (PCA) based method.     

A public fragile watermarking scheme for 3D model authentication

A public fragile watermarking scheme based on the sensitivity of vertex geometry is proposed for 3D model authentication. In the 3D fragile watermarking embedding, slightly perturbing the positions of a subset of vertices is usually needed to keep them in some predefined relationship with their neighboring vertices. Two problems frequently arise in the embedding stage: the causality problem and the convergence problem. The causality problem arises while the neighboring relationship of a former processed vertex is influenced by the perturbing of its latter processed neighboring vertices. The convergence problem means that the original model has been heavily distorted before some vertices reach the predefined relationship. In this paper, we propose a multi-function vertex embedding method and an adjusting-vertex method to overcome these two problems. The proposed method does not need the original model and watermarks for authentication; moreover, the key for extracting watermarks is relatively smaller than that of previous works. For some artistic or technical models, sometimes it is very important to control the distortion ratio caused by watermark embedding. Our method can control the average distortion by the keys used in watermark embedding. Experimental and analytic results on various kinds of 3D models show the effectiveness of the scheme.     

Analytical model for three-dimensional light-field displays based on voxel construction

Three-dimensional light-field displays (3D-LFDs) can provide viewers with glass-free 3D images. The display effect of reconstructed 3D images is closely related to the number of 3D object points that the 3D display system can construct, which can be regarded as the information quantity of the 3D-LFDs. However, the information quantity presented by a 3D-LFD is limited by fixed hardware parameters such as display panels and optical components. It is difficult to present an ideal 3D visual experience with the increase in display depth. In this paper, an analytical model based on voxel construction for 3D-LFDs is proposed, which can estimate the information quantity that a 3D-LFD can reproduce. The variation regularity of information quantity with display depth is analyzed and the influence of hardware parameters on display effect is discussed. With the proposed strategies, the display effect can be improved in a targeted manner without blurring and aliasing by adjusting the hardware parameters in display systems and matching the information quantity at different display depths.     

网络环境下道路三维整体建模与简化方法

为实现网络环境下道路工程的三维可视化,对其中的关键技术：顾及约束的整体模型构建及模型简化方法进行了研究。基于约束Delaunay三角网构建理论,建立了外形与内部拓扑关系均为整体的道路三维模型。提出了顾及道路约束边界的半边折叠误差度量方法,采用半边折叠操作,在服务器端对道路模型进行整体简化,并建立操作层次树存储操作记录;提出了约束边优先细化,延迟简化的视相关策略,结合视相关重构准则,减少网络可视化所需传输的数据量,在客户端实现了道路三维模型的快速重构。基于上述原理方法开发了相关系统,已在高速公路的网络建设管理中成功应用。      

An architecture for universal construction via modular robotic components

A set of modular components is presented for use in reconfigurable robotic construction systems. The set includes passive and active components. The passive components can be formed into static structures and adaptable grids carrying electrical power and signals. Passive and active components can be combined into general purpose mobile manipulators which are able to augment and reconfigure the grid, construct new manipulators, and potentially perform general purpose fabrication tasks such as additive manufacturing. The components themselves are designed for low-cost, simple fabrication methods and could potentially be fabricated by constructors made of the same components. This work represents a step toward a Cyclic Fabrication System, a network of materials, tools, and manufacturing processes that can produce all of its constituent components. These and similar systems have been proposed for a wide range of far-term applications, including space-based manufacturing, construction of large-scale industrial facilities, and also for driving development of low-cost 3D printing machines.