基于图卷积神经网络的三维点云分割算法Graph⁃PointNet

三维点云无序不规则的特性使得传统的卷积神经网络无法直接应用,且大多数点云深度学习模型往往忽略大量的空间信息。为便于捕获空间点邻域信息,获得更好的点云分析性能以用于点云语义分割,文中提出Graph⁃PointNet点云深度学习模型。Graph⁃PointNet在经典点云模型PointNet的基础上,结合二维图像中聚类思想,设计了图卷积特征提取模块取代多层感知器嵌入PointNet中。图卷积特征提取模块首先通过K近邻算法搜寻相邻特征点组成图结构,接着将多组图结构送入图卷积神经网络提取局部特征用于分割。同时文中设计一种新型点云采样方法多邻域采样,多邻域采样通过设置点云间夹角阈值,将点云区分为特征区域和非特征区域,特征区域用于提取特征,非特征区域用于消除噪声。对室内场景S3DIS、室外场景Semantic3D数据集进行实验,得到二者整体精度分别达到89.33%和89.78%,平均交并比达到64.62%,61.47%,均达到最佳效果。最后,进行消融实验,进一步证明了文中所提出的多邻域采样和图卷积特征提取模块对提高点云语义分割的有效性。

基于轻量化PointNet网络的林果园喷雾作业靶标实时识别方法

为了进一步提高喷雾机器人靶标检测的精准性、实时性和应用部署的实用性,该研究提出一种基于轻量化PointNet网络的林果园喷雾作业靶标实时识别方法。首先通过区域提取降采样、地面分割和改进DBSCAN聚类等点云预处理方法提取原始点云中的靶标;然后通过移动最小二乘上采样将靶标点云转化为满足点云识别网络输入要求的点云数据;最终通过在PointNet网络中引入残差模块和改进循环剪枝算法轻量化PointNet网络,完成林果树靶标的实时识别。试验结果表明,在ModelNet40数据集上,轻量化PointNet网络可达89.7%的准确率;在实际苗圃环境的试验中,该研究方法对靶标的识别准确率可达92.49%,同时误识率与拒识率分别为13.4%和6.47%,相较PointNet网络识别准确率提升了4.38个百分点,误识率和拒识率分别降低了7.2和4.07个百分点;轻量化PointNet网络识别准确率仅比PointNet++网络低1.14个百分点,误识率和拒识率分别高了0.9和1.12个百分点。但是轻量化PointNet网络的模型参数量较PointNet网络和PointNet++网络的模型参数量显著减少,仅为PointNet网络的11.5%,PointNet++网络的27.02%;运算量相较PointNet网络、PointNet++网络分别减少13.3和76.79个百分点。该研究提出的轻量化PointNet网络具有较高的实时性、精确性和鲁棒性,能够满足林果园喷雾作业的靶标识别需求,可为林果园喷雾作业靶标实时识别提供参考。

基于PointNet优化网络的铁路站台语义分割

铁路站台点云语义分割是对铁路侵界现象进行检测的关键环节。文中以新型激光扫描测量系统采集的具有三维空间信息的点云数据为基础,在获取初步分割结果的基础上,设计PointNet网络整体结构提取点云数据全局特征,采用多层次金字塔结构对网络进行局部特征提取优化,实现铁路站台点云数据语义分割。研究表明,所提方法对实验点云数据的分割准确率达到84.5%,在铁路工程应用中的点云总体分割精度达到75.34%,在铁路检测中实现了大范围多尺度点云数据的可靠语义分割,满足铁路侵界现象检测分析需求。

Graph Transformers研究进展综述

随着图结构数据在各种实际场景中的广泛应用,对其进行有效建模和处理的需求日益增加。Graph Transformers(GTs)作为一类使用Transformers处理图数据的模型,能够有效缓解传统图神经网络(GNN)中存在的过平滑和过挤压等问题,因此可以学习到更好的特征表示。根据对近年来GTs相关文献的研究,将现有的模型架构分为两类:第一类通过绝对编码和相对编码向Transformers中加入图的位置和结构信息,以增强Transformers对图结构数据的理解和处理能力;第二类根据不同的方式(串行、交替、并行)将GNN与Transformers进行结合,以充分利用两者的优势。介绍了GTs在信息安全、药物发现和知识图谱等领域的应用,对比总结了不同用途的模型及其优缺点。最后,从可扩展性、复杂图、更好的结合方式等方面分析了GTs未来研究面临的挑战。

基于Graph Transformer的半监督异配图表示学习模型

现有的图卷积网络(GCN)模型基于同配性假设,无法直接应用于异配图的表示学习,且许多异配图表示学习的研究工作受消息传递机制的限制,导致节点特征混淆和特征过度挤压而出现过平滑问题。针对这些问题,提出一种基于Graph Transformer的半监督异配图表示学习模型HPGT(HeteroPhilic Graph Transformer)。首先,使用度连接概率矩阵采样节点的路径邻域,再通过自注意力机制自适应地聚合路径上的节点异配连接模式,编码得到节点的结构信息,用节点的原始属性信息和结构信息构建Transformer层的自注意力模块;其次,将每个节点自身的隐层表示与它的邻域节点的隐层表示分离更新以避免节点通过自注意力模块聚合过量的自身信息,再把每个节点表示与它的邻域表示连接,得到单个Transformer层的输出,另外,将所有的Transformer层的输出跳连到最终的节点隐层表示以防止中间层信息丢失;最后,使用线性层和Softmax层将节点的隐层表示映射到节点的预测标签。实验结果表明,与无结构编码(SE)的模型相比,基于度连接概率的SE能为Transformer层的自注意力模块提供有效的偏差信息,HPGT平均准确率提升0.99%~11.98%;与对比模型相比,在异配数据集(Texas、Cornell、Wisconsin和Actor)上,模型节点分类准确率提升0.21%~1.69%,在同配数据集(Cora、CiteSeer和PubMed)上,节点分类准确率分别达到了0.8379、0.7467和0.8862。以上结果验证了HPGT具有较强的异配图表示学习能力,尤其适用于强异配图节点分类任务。

基于TF-PointNet++的林木点云数据语义分割算法研究

【目的】林木点云数据具有边界模糊、空间分布不均的特点,经典PointNet++网络对林木点云邻域特征的关注有局限性,存在原点云的邻接关系未能捕捉全面、不能很好地进行分割的问题,因此改善对林木的分割效果十分重要。【方法】提出了一种基于Transformer-PointNet++的点云语义分割网络,在经典PointNet++网络的中间层引入了Transformer结构,利用自注意力机制优化特征提取,核心思想是将点云中的每个点作为输入,在下采样层以及输出层利用自注意力机制对点云进行全局的特征交互和信息聚合,改善了经典PointNet++对于邻域特征的信息关注。在此基础上,采用了开源ModelNet40点云集合和自建林木点云数据集,分别投入到经典PointNet++网络、DST-PointNet++、B-PointNet++和TF-PointNet++网络进行实验,用于测试改进后网络模型的效果。【结果】TF-PointNet++在开源ModelNet40数据集和自建林木点云数据集上皆有更好的表现。在ModelNet40上,OA达到了94.08%,较PointNet++提升了4.04个百分点;在自建数据集上,OA达到了82.52%,较PointNet++提升了26.24个百分点,mIoU达到了85.67%,提升了11.21个百分点。【结论】TF-PointNet++网络可以使模型更好地理解和利用整个点云的上下文信息,从而提升语义分割的准确性和全局一致性。本研究证明了TF-PointNet++网络在处理语义分割任务中的有效性和先进性,对于提升林木点云数据的语义分割效果具有重要意义。

基于DenseNet与PointNet融合算法的三维点云分割

点云分割对于智能驾驶、物体检测和识别、逆向工程等任务非常重要。PointNet是一种能够直接处理点云数据的方法,近年来在点云分割任务中得到广泛应用,但其分割精度较低,而PointNet++的计算成本又较高。针对以上问题,提出一种融合DenseNet和PointNet的算法,用于点云分割,并引入三分支混合注意力机制,以提高PointNet在提取局部特征方面的能力。基于密集连接卷积网络(DenseNet)思想,提出用DenseNet-STN和DenseNet-MLP结构来替代PointNet中的空间变换网络(STN)和多层感知机(MLP);同时,使用Add连接代替密集块(DenseBlock)中的Concat连接,以提高对点特征间相关性的准确性,同时不显著增加模型复杂度。DenseNet-PointNet能够提高复杂分类问题的泛化能力,实现对复杂函数更好的逼近,从而提高点云分割的准确率。有效性和消融实验结果表明,本文算法具有良好的性能。点云分割实验结果表明,DenseNet-PointNet在大多数类别中的交并比(IoU)都高于PointNet的IoU,并在部分类别中也高于PointNet++,参数量是PointNet++的47.6%,浮点运算量(FLOPs)是PointNet++的49.1%。实验结果验证了DenseNet-PointNet的可行性和有效性。

GraphMLP-Mixer:基于图-多层感知机架构的高效多行为序列推荐方法

在多行为序列推荐领域,图神经网络(GNNs)虽被广泛应用,但存在局限性,如对序列间协同信号建模不足和处理长距离依赖性等问题.针对这些问题,提出了一种新的解决框架GraphMLP-Mixer.该框架首先构造全局物品图来增强模型对序列间协同信号的建模,然后将感知机-混合器架构与图神经网络结合,得到图-感知机混合器模型对用户兴趣进行充分挖掘.GraphMLP-Mixer具有2个显著优势:一是能够有效捕捉用户行为的全局依赖性,同时减轻信息过压缩问题;二是其时间与空间效率显著提高,其复杂度与用户交互行为的数量成线性关系,优于现有基于GNN多行为序列推荐模型.在3个真实的公开数据集上进行实验,大量的实验结果验证了GraphMLP-Mixer在处理多行为序列推荐问题时的有效性和高效性.

基于PointNet++的焊装夹具零件识别

焊装夹具是汽车白车身焊接生产线中重要的组成部分,有效的管理和归纳焊装夹具零件设计数模能够显著提高设计效率。将原始设计数模离散为点云,利用点云数据和PointNet++深度学习网络探讨了一种焊装夹具零件智能分类方法,并对比各模型的分类精度,选取运行效率和精度最高的单尺度分组(SSG)模型完成焊装夹具零件的分类。训练结果表明,该方法在验证集上的准确率为97.5%,型块、连接块、定位销、销座、支座的验证集类内准确率分别为92.5%、97.5%、100%、97.5%和100%。这些结果表明该方法具有较高的识别精度,能够满足焊装夹具零件分类的精度要求。

基于改进PointNet的空调散热器V形槽3D点云分割算法

针对给空调散热器自动化点胶时无法准确识别散热器V形槽位置的问题,基于PointNet网络的散热器V形槽语义分割方法,首先针对散热器点云V形槽区域与内部区域特征相似的问题,设计一种通过提取点云边缘将点云边缘区域的点云与内部区域的点云分别进行预处理的方法,实现突出点云边缘区域特征的目的。其次,在PointNet网络最大池化函数的基础上,引入平均池化函数,增加网络所提取的全局特征的特征信息,减少因最大池化引起的信息丢失,并去除T-Net变换网络,减少网络的复杂度。从实验室平台采集空调散热器样本进行实验,结果表明,改进算法的平均交并比(mean intersection over union, mIoU)达到78.17%,总体精度(overall accuracy, OA)达到了92.01%,相较于PointNet提高了9.73%和6.37%,验证了算法的有效性。

基于GraphSAGE网络的藏文短文本分类研究

文本分类是自然语言处理领域的重要研究方向,由于藏文数据的稀缺性、语言学特征抽取的复杂性、篇章结构的多样性等因素导致藏文文本分类任务进展缓慢。因此,该文以图神经作为基础模型进行改进。首先,在“音节-音节”“音节-文档”建模的基础上,融合文档特征,采用二元分类模型动态网络构建“文档-文档”边,以充分挖掘短文本的全局特征,增加滑动窗口,减少模型的计算复杂度并寻找最优窗口取值。其次,针对藏文短文本的音节稀疏性,首次引入GraphSAGE作为基础模型,并探究不同聚合方式在藏文短文本分类上的性能差异。最后,为捕获节点间关系的异质性,对邻居节点进行特征加权再平均池化以增强模型的特征提取能力。在TNCC标题文本数据集上,该文模型的分类准确率达到了62.50%,与传统GCN、原始GraphSAGE和预训练语言模型CINO相比,该方法在分类准确率上分别提高了2.56%、1%和2.4%。

A STABILITY RESULT FOR TRANSLATINGSPACELIKE GRAPHS IN LORENTZ MANIFOLDS

In this paper,we investigate spacelike graphs defined over a domain Ω⊂M^(n) in the Lorentz manifold M^(n)×ℝ with the metric−ds^(2)+σ,where M^(n) is a complete Riemannian n-manifold with the metricσ,Ωhas piecewise smooth boundary,and ℝ denotes the Euclidean 1-space.We prove an interesting stability result for translating spacelike graphs in M^(n)×ℝ under a conformal transformation.

A Value for Games Defined on Graphs

Given a graph g=( V,A ) , we define a space of subgraphs M with the binary operation of union and the unique decomposition property into blocks. This space allows us to discuss a notion of minimal subgraphs (minimal coalitions) that are of interest for the game. Additionally, a partition of the game is defined in terms of the gain of each block, and subsequently, a solution to the game is defined based on distributing to each player (node and edge) present in each block a payment proportional to their contribution to the coalition.

Heterophilic Graph Neural Network Based on Spatial and Frequency Domain Adaptive Embedding Mechanism

Graph Neural Networks(GNNs)play a significant role in tasks related to homophilic graphs.Traditional GNNs,based on the assumption of homophily,employ low-pass filters for neighboring nodes to achieve information aggregation and embedding.However,in heterophilic graphs,nodes from different categories often establish connections,while nodes of the same category are located further apart in the graph topology.This characteristic poses challenges to traditional GNNs,leading to issues of“distant node modeling deficiency”and“failure of the homophily assumption”.In response,this paper introduces the Spatial-Frequency domain Adaptive Heterophilic Graph Neural Networks(SFA-HGNN),which integrates adaptive embedding mechanisms for both spatial and frequency domains to address the aforementioned issues.Specifically,for the first problem,we propose the“Distant Spatial Embedding Module”,aiming to select and aggregate distant nodes through high-order randomwalk transition probabilities to enhance modeling capabilities.For the second issue,we design the“Proximal Frequency Domain Embedding Module”,constructing adaptive filters to separate high and low-frequency signals of nodes,and introduce frequency-domain guided attention mechanisms to fuse the relevant information,thereby reducing the noise introduced by the failure of the homophily assumption.We deploy the SFA-HGNN on six publicly available heterophilic networks,achieving state-of-the-art results in four of them.Furthermore,we elaborate on the hyperparameter selection mechanism and validate the performance of each module through experimentation,demonstrating a positive correlation between“node structural similarity”,“node attribute vector similarity”,and“node homophily”in heterophilic networks.

基于PointNet++的工井点云语义分割模型研究

受地下工井空间狭窄、环境复杂的影响,采集的工井点云数据存在空间分布不规则且不均匀、数据量大、难以实现多目标高效的语义分割等问题。有鉴于此,本文提出了一种基于PointNet++的工井点云语义分割方法。首先,采集地下工井点云数据并加入语义标签,制作模型训练需要的数据集;其次,为提高点云分割任务的性能,引入一种基于深度学习的PointNet++网络模型,并利用多分辨率分组(multi-scalegrouping,MSG)和随机输入(random inputdropout,DP)策略,实现了地下工井地面、顶、爬梯、墙、电缆线和支架的语义分割;最后,采用精确度、召回率、交并比和F1分数作为评价指标对分割效果进行评价。结果表明,与PointNet网络模型相比,本文方法各类别的评价指标均得到了显著的提升,并且地下工井地面、井顶、井墙和电缆线等类别的评价指标均超过80%,显示分割性能良好,有利于地下电缆工井场景的多目标快速精准分割,为地下工程精细化管理奠定了基础。

Weighted Forwarding in Graph Convolution Networks for Recommendation Information Systems

Recommendation Information Systems(RIS)are pivotal in helping users in swiftly locating desired content from the vast amount of information available on the Internet.Graph Convolution Network(GCN)algorithms have been employed to implement the RIS efficiently.However,the GCN algorithm faces limitations in terms of performance enhancement owing to the due to the embedding value-vanishing problem that occurs during the learning process.To address this issue,we propose a Weighted Forwarding method using the GCN(WF-GCN)algorithm.The proposed method involves multiplying the embedding results with different weights for each hop layer during graph learning.By applying the WF-GCN algorithm,which adjusts weights for each hop layer before forwarding to the next,nodes with many neighbors achieve higher embedding values.This approach facilitates the learning of more hop layers within the GCN framework.The efficacy of the WF-GCN was demonstrated through its application to various datasets.In the MovieLens dataset,the implementation of WF-GCN in LightGCN resulted in significant performance improvements,with recall and NDCG increasing by up to+163.64%and+132.04%,respectively.Similarly,in the Last.FM dataset,LightGCN using WF-GCN enhanced with WF-GCN showed substantial improvements,with the recall and NDCG metrics rising by up to+174.40%and+169.95%,respectively.Furthermore,the application of WF-GCN to Self-supervised Graph Learning(SGL)and Simple Graph Contrastive Learning(SimGCL)also demonstrated notable enhancements in both recall and NDCG across these datasets.

A Temporal Knowledge Graph Embedding Model Based on Variable Translation

Knowledge representation learning(KRL)aims to encode entities and relationships in various knowledge graphs into low-dimensional continuous vectors.It is popularly used in knowledge graph completion(or link prediction)tasks.Translation-based knowledge representation learning methods perform well in knowledge graph completion(KGC).However,the translation principles adopted by these methods are too strict and cannot model complex entities and relationships(i.e.,N-1,1-N,and N-N)well.Besides,these traditional translation principles are primarily used in static knowledge graphs and overlook the temporal properties of triplet facts.Therefore,we propose a temporal knowledge graph embedding model based on variable translation(TKGE-VT).The model proposes a new variable translation principle,which enables flexible transformation between entities and relationship embedding.Meanwhile,this paper considers the temporal properties of both entities and relationships and applies the proposed principle of variable translation to temporal knowledge graphs.We conduct link prediction and triplet classification experiments on four benchmark datasets:WN11,WN18,FB13,and FB15K.Our model outperforms baseline models on multiple evaluation metrics according to the experimental results.

An End-To-End Hyperbolic Deep Graph Convolutional Neural Network Framework

Graph Convolutional Neural Networks(GCNs)have been widely used in various fields due to their powerful capabilities in processing graph-structured data.However,GCNs encounter significant challenges when applied to scale-free graphs with power-law distributions,resulting in substantial distortions.Moreover,most of the existing GCN models are shallow structures,which restricts their ability to capture dependencies among distant nodes and more refined high-order node features in scale-free graphs with hierarchical structures.To more broadly and precisely apply GCNs to real-world graphs exhibiting scale-free or hierarchical structures and utilize multi-level aggregation of GCNs for capturing high-level information in local representations,we propose the Hyperbolic Deep Graph Convolutional Neural Network(HDGCNN),an end-to-end deep graph representation learning framework that can map scale-free graphs from Euclidean space to hyperbolic space.In HDGCNN,we define the fundamental operations of deep graph convolutional neural networks in hyperbolic space.Additionally,we introduce a hyperbolic feature transformation method based on identity mapping and a dense connection scheme based on a novel non-local message passing framework.In addition,we present a neighborhood aggregation method that combines initial structural featureswith hyperbolic attention coefficients.Through the above methods,HDGCNN effectively leverages both the structural features and node features of graph data,enabling enhanced exploration of non-local structural features and more refined node features in scale-free or hierarchical graphs.Experimental results demonstrate that HDGCNN achieves remarkable performance improvements over state-ofthe-art GCNs in node classification and link prediction tasks,even when utilizing low-dimensional embedding representations.Furthermore,when compared to shallow hyperbolic graph convolutional neural network models,HDGCNN exhibits notable advantages and performance enhancements.

GATiT:An Intelligent Diagnosis Model Based on Graph Attention Network Incorporating Text Representation in Knowledge Reasoning

The growing prevalence of knowledge reasoning using knowledge graphs(KGs)has substantially improved the accuracy and efficiency of intelligent medical diagnosis.However,current models primarily integrate electronic medical records(EMRs)and KGs into the knowledge reasoning process,ignoring the differing significance of various types of knowledge in EMRs and the diverse data types present in the text.To better integrate EMR text information,we propose a novel intelligent diagnostic model named the Graph ATtention network incorporating Text representation in knowledge reasoning(GATiT),which comprises text representation,subgraph construction,knowledge reasoning,and diagnostic classification.In the text representation process,GATiT uses a pre-trained model to obtain text representations of the EMRs and additionally enhances embeddings by including chief complaint information and numerical information in the input.In the subgraph construction process,GATiT constructs text subgraphs and disease subgraphs from the KG,utilizing EMR text and the disease to be diagnosed.To differentiate the varying importance of nodes within the subgraphs features such as node categories,relevance scores,and other relevant factors are introduced into the text subgraph.Themessage-passing strategy and attention weight calculation of the graph attention network are adjusted to learn these features in the knowledge reasoning process.Finally,in the diagnostic classification process,the interactive attention-based fusion method integrates the results of knowledge reasoning with text representations to produce the final diagnosis results.Experimental results on multi-label and single-label EMR datasets demonstrate the model’s superiority over several state-of-theart methods.

A Generalization of Torsion Graph for Modules

Let R be a commutative ring with identity and M an R-module. In this paper, we relate a graph to M, say Γ(M), provided tsshat when M=R, Γ(M)is exactly the classic zero-divisor graph.