Graph Convolutional Networks Embedding Textual Structure Information for Relation Extraction

Deep neural network-based relational extraction research has made significant progress in recent years,andit provides data support for many natural language processing downstream tasks such as building knowledgegraph,sentiment analysis and question-answering systems.However,previous studies ignored much unusedstructural information in sentences that could enhance the performance of the relation extraction task.Moreover,most existing dependency-based models utilize self-attention to distinguish the importance of context,whichhardly deals withmultiple-structure information.To efficiently leverage multiple structure information,this paperproposes a dynamic structure attention mechanism model based on textual structure information,which deeplyintegrates word embedding,named entity recognition labels,part of speech,dependency tree and dependency typeinto a graph convolutional network.Specifically,our model extracts text features of different structures from theinput sentence.Textual Structure information Graph Convolutional Networks employs the dynamic structureattention mechanism to learn multi-structure attention,effectively distinguishing important contextual features invarious structural information.In addition,multi-structure weights are carefully designed as amergingmechanismin the different structure attention to dynamically adjust the final attention.This paper combines these featuresand trains a graph convolutional network for relation extraction.We experiment on supervised relation extractiondatasets including SemEval 2010 Task 8,TACRED,TACREV,and Re-TACED,the result significantly outperformsthe previous.

Multiphase convolutional dense network for the classification of focal liver lesions on dynamic contrast-enhanced computed tomography

BACKGROUND The accurate classification of focal liver lesions(FLLs)is essential to properly guide treatment options and predict prognosis.Dynamic contrast-enhanced computed tomography(DCE-CT)is still the cornerstone in the exact classification of FLLs due to its noninvasive nature,high scanning speed,and high-density resolution.Since their recent development,convolutional neural network-based deep learning techniques has been recognized to have high potential for image recognition tasks.AIM To develop and evaluate an automated multiphase convolutional dense network(MP-CDN)to classify FLLs on multiphase CT.METHODS A total of 517 FLLs scanned on a 320-detector CT scanner using a four-phase DCECT imaging protocol(including precontrast phase,arterial phase,portal venous phase,and delayed phase)from 2012 to 2017 were retrospectively enrolled.FLLs were classified into four categories:Category A,hepatocellular carcinoma(HCC);category B,liver metastases;category C,benign non-inflammatory FLLs including hemangiomas,focal nodular hyperplasias and adenomas;and category D,hepatic abscesses.Each category was split into a training set and test set in an approximate 8:2 ratio.An MP-CDN classifier with a sequential input of the fourphase CT images was developed to automatically classify FLLs.The classification performance of the model was evaluated on the test set;the accuracy and specificity were calculated from the confusion matrix,and the area under the receiver operating characteristic curve(AUC)was calculated from the SoftMax probability outputted from the last layer of the MP-CDN.RESULTS A total of 410 FLLs were used for training and 107 FLLs were used for testing.The mean classification accuracy of the test set was 81.3%(87/107).The accuracy/specificity of distinguishing each category from the others were 0.916/0.964,0.925/0.905,0.860/0.918,and 0.925/0.963 for HCC,metastases,benign non-inflammatory FLLs,and abscesses on the test set,respectively.The AUC(95%confidence interval)for differentiating each category from the others was 0.92(0.837-0.992),0.99(0.967-1.00),0.88(0.795-0.955)and 0.96(0.914-0.996)for HCC,metastases,benign non-inflammatory FLLs,and abscesses on the test set,respectively.CONCLUSION MP-CDN accurately classified FLLs detected on four-phase CT as HCC,metastases,benign non-inflammatory FLLs and hepatic abscesses and may assist radiologists in identifying the different types of FLLs.

A New Speech Encoder Based on Dynamic Framing Approach

Latent information is difficult to get from the text in speech synthesis.Studies show that features from speech can get more information to help text encoding.In the field of speech encoding,a lot of work has been conducted on two aspects.The first aspect is to encode speech frame by frame.The second aspect is to encode the whole speech to a vector.But the scale in these aspects is fixed.So,encoding speech with an adjustable scale for more latent information is worthy of investigation.But current alignment approaches only support frame-by-frame encoding and speech-to-vector encoding.It remains a challenge to propose a new alignment approach to support adjustable scale speech encoding.This paper presents the dynamic speech encoder with a new alignment approach in conjunction with frame-by-frame encoding and speech-to-vector encoding.The speech feature fromourmodel achieves three functions.First,the speech feature can reconstruct the origin speech while the length of the speech feature is equal to the text length.Second,our model can get text embedding fromspeech,and the encoded speech feature is similar to the text embedding result.Finally,it can transfer the style of synthesis speech and make it more similar to the given reference speech.

EFFICIENT NUMERICAL METHOD FOR DYNAMIC ANALYSIS OF FLEXIBLE ROD HIT BY RIGID BALL

Impact dynamics of flexible solids is important in engineering practice. Obtaining its dynamic response is a challenging task and usually achieved by numerical methods. The objectives of the study are twofold. Firstly, the discrete singular convolution （DSC） is used for the first time to analyze the impact dynamics. Secondly, the efficiency of various numerical methods for dynamic analysis is explored via an example of a flexible rod hit by a rigid ball. Three numerical methods, including the conventional finite element （FE） method, the DSC algorithm, and the spectral finite element （SFE） method, and one proposed modeling strategy, the improved spectral finite element （ISFE） method, are involved. Numerical results are compared with the known analytical solutions to show their efficiency. It is demonstrated that the proposed ISFE modeling strategy with a proper length of con- ventional FE yields the most accurate contact stress among the four investigated models. It is also found that the DSC algorithm is an alternative method for collision problems.

On the convergence and causality of a frequency domain method for dynamic structural analysis

Venanico-Filho et al. developed an elegant matrix formulation for dynamic analysis by frequency domain （FD）, but the convergence, causality and extended period need further refining. In the present paper, it was argued that： （1） under reasonable assumptions （approximating the frequency response function by the discrete Fourier transform of the discretized unitary impulse response function）, the matrix formulation by FD is equivalent to a circular convolution; （2） to avoid the wraparound interference, the excitation vector and impulse response must be padded with enough zeros; （3） provided that the zero padding requirement satisfied, the convergence and accuracy of direct time domain analysis, which is equivalent to that by FD, are guaranteed by the numerical integration scheme; （4） the imaginary part of the computational response approaching zero is due to the continuity of the impulse response functions.

Use of Local Region Maps on Convolutional LSTM for Single-Image HDR Reconstruction

Low dynamic range(LDR)images captured by consumer cameras have a limited luminance range.As the conventional method for generating high dynamic range(HDR)images involves merging multiple-exposure LDR images of the same scene(assuming a stationary scene),we introduce a learning-based model for single-image HDR reconstruction.An input LDR image is sequentially segmented into the local region maps based on the cumulative histogram of the input brightness distribution.Using the local region maps,SParam-Net estimates the parameters of an inverse tone mapping function to generate a pseudo-HDR image.We process the segmented region maps as the input sequences on long short-term memory.Finally,a fast super-resolution convolutional neural network is used for HDR image reconstruction.The proposed method was trained and tested on datasets including HDR-Real,LDR-HDR-pair,and HDR-Eye.The experimental results revealed that HDR images can be generated more reliably than using contemporary end-to-end approaches.

Deep Learning Convolutional Neural Network for ECG Signal Classification Aggregated Using IoT

Much attention has been given to the Internet of Things (IoT) by citizens, industries, governments, and universities for applications like smart buildings, environmental monitoring, health care and so on. With IoT, networkconnectivity is facilitated between smart devices from anyplace and anytime.IoT-based health monitoring systems are gaining popularity and acceptance forcontinuous monitoring and detect health abnormalities from the data collected.Electrocardiographic (ECG) signals are widely used for heart diseases detection.A novel method has been proposed in this work for ECG monitoring using IoTtechniques. In this work, a two-stage approach is employed. In the first stage, arouting protocol based on Dynamic Source Routing (DSR) and Routing byEnergy and Link quality (REL) for IoT healthcare platform is proposed for effi-cient data collection, and in the second stage, classification of ECG for Arrhythmia. Furthermore, this work has evaluated Support Vector Machine (SVM),Artificial Neural Network (ANN), and Convolution Neural Networks (CNNs)-based approach for ECG signals classification. Deep-ECG will use a deep CNNto extract critical features and then compare through evaluation of simple and fastdistance functions in order to obtain an efficient classification of heart abnormalities. For the identification of abnormal data, this work has proposed techniquesfor the classification of ECG data, which has been obtained from mobile watchusers. For experimental verification of the proposed methods, the Beth Israel Hospital (MIT/BIH) Arrhythmia and Massachusetts Institute of Technology (MIT)Database was used for evaluation. Results confirm the presented method’s superior performance with regards to the accuracy of classification. The CNN achievedan accuracy of 91.92% and has a higher accuracy of 4.98% for the SVM and2.68% for the ANN.

基于TCN和迁移学习的混凝土坝变形预测方法

混凝土坝变形测点数据丢失或者新增测点测量时间太短都会导致这部分测点的数据量不足,使得变形预测精度受到影响。为了提高这些小数据量测点的变形预测精度,提出了将时域卷积网络(TCN)与迁移学习相结合的变形预测方法。以数据量充足的测点为源域,以缺少数据的测点为目标域,将在源域上训练好的TCN模型的结构和参数迁移到目标域模型中,固定其中的冻结层参数,利用目标域中的数据对目标域模型可调层参数进行调整。同时,采用动态时间规整选择与目标域数据序列相似度最高的监测数据作为最佳源域数据,提升迁移学习效果。工程实例分析表明:迁移学习后的目标域模型的均方根误差和平均绝对误差与利用足量数据训练的TCN模型的预测误差相比,差异仅分别为1.73%和8.09%,小数据量情况下TCN预测模型的精度得到了提高。

基于Transformer和动态3D卷积的多源遥感图像分类

多源遥感数据具有互补性和协同性,近年来,基于深度学习的方法已经在多源遥感图像分类中取得了一定进展,但当前方法仍面临关键难题,如多源遥感图像特征表达不一致,融合困难,基于静态推理范式的神经网络缺乏对不同类别地物的适应性。为解决上述问题,提出了基于跨模态Transformer和多尺度动态3D卷积的多源遥感图像分类模型。为提高多源特征表达的一致性,设计了基于Transformer的融合模块,借助其强大的注意力建模能力挖掘高光谱和LiDAR数据特征之间的相互作用;为提高特征提取方法对不同地物类别的适应性,设计了多尺度动态3D卷积模块,将输入特征的多尺度信息融入卷积核的调制,提高卷积操作对不同地物的适应性。采用多源遥感数据集Houston和Trento对所提方法进行验证,实验结果表明:所提方法在Houston和Trento数据集上总体准确率分别达到94.60%和98.21%,相比MGA-MFN等主流方法,总体准确率分别至少提升0.97%和0.25%,验证了所提方法可有效提升多源遥感图像分类的准确率。

基于时序图卷积的动态网络链路预测

动态网络链路预测广泛的应用前景,使得其逐渐成为网络科学研究的热点.动态网络链路演化过程中具有复杂的空间相关性和时间依赖性,导致其链路预测任务极具挑战.提出一个基于时序图卷积的动态网络链路预测模型(dynamic network link prediction based on sequential graph convolution, DNLP-SGC).针对网络快照序列不能有效反映动态网络连续性的问题,采用边缘触发机制对原始网络权重矩阵进行修正,弥补了离散快照表示动态网络存在时序信息丢失的不足.从网络演化过程出发,综合考虑节点间的特征相似性以及历史交互信息,采用时序图卷积提取动态网络中节点的特征,该方法融合了节点时空依赖关系.进一步,采用因果卷积网络捕获网络演化过程中潜在的全局时序特征,实现动态网络链路预测.在2个真实的网络数据集上的实验结果表明,DNLP-SGC在precision, recall, AUC指标上均优于对比的基线模型.

自适应动态滤波网络地震随机噪声压制方法

由于地质及环境条件复杂,导致地震勘探采集资料信噪比相对较低,对后续的研究带来不利影响,因此地震勘探数据处理中对随机噪声的压制一直备受关注。现有算法无法较好实现对噪声的有效压制和对有效信号的极大保留,为此,将传统方法和深度学习相结合,提出了基于自适应动态滤波网络(Adaptive Dynamic Filtering Net,ADFNet)的方法压制地震资料中的随机噪声。该网络以编码器—解码器为架构,首先引入通道注意力机制(Attention Mechanism,AM)的思想,通过通道AM对空洞卷积多尺度数据特征集成,为网络提供了精准且丰富的特征表示;然后,引入动态卷积,以较低的计算复杂度实现对地震资料高频特征的学习,从而保留更丰富的细节信息。合成数据和实际数据的实验结果均表明,ADFNet可有效压制地震资料中的随机噪声,同时保留更丰富的地震数据细节,处理后的地震数据信噪比得到显著提升。

基于自编码器的动态协作中继系统

鉴于大多数现有端到端自编码器(AE)仅适用于点对点的通信场景,提出一种基于AE的动态协作通信系统,将基于深度学习的AE扩展到多点通信系统。构建了3个神经网络子系统,分别用于学习发送端、中继节点和接收端的最佳编码、传输和解码,通过三者的联合训练达到多点通信系统的最佳传输性能。其中,发送端和接收端使用一维卷积层进行信号特征的提取及学习,中继节点通过引入密集层和一维卷积层,支持放大转发(AF)和解码转发(DF)两种经典的中继协作方式。仿真实验表明,在加性高斯白噪声以及瑞利衰落信道条件下,提出的模型采用两种不同的协作方式,其误码性能均优于单一点到点通信系统,验证了系统方案的可行性和有效性。此外,该系统支持动态的节点拓扑结构,在无需额外训练的条件下,本系统支持中继节点数量实时变化。

密度导向的点云动态图卷积网络

针对现有主流网络对于点云局部特征提取的能力不足,以及在特征提取过程中未考虑点云密度的问题,提出一种密度导向的点云动态图卷积网络.首先提出点云局部密度指数的概念,衡量点及其邻域点在相应的空间位置中的密集程度;然后利用局部密度指数动态赋予每个点一个膨胀因子,提出密度导向的动态点分组方法对点云构建局部图结构,对每个局部图结构构造动态边缘卷积模块进行特征的提取与聚合,既提取了点云的几何特征,又实现了置换不变性;最后采用残差网络的思想优化图神经网络的过平滑问题.实验结果表明,在分类数据集ModelNet40与ScanObjectNN上,所提网络的分类准确率分别为93.5%和82.2%;在分割数据集ShapeNet与S3DIS上,该网络的平均交并比分别为85.6%和60.4%,均高于DGCNN等主流网络;所提网络在多项任务中的精度都得到显著提升,且在处理密度不均的点云时有较好的鲁棒性,验证了所提算法的可行性与有效性.

基于深度与实例分割融合的单目3D目标检测方法

针对单目3D目标检测在视角变化引起的物体大小变化以及物体遮挡等情况下效果不佳的问题,提出一种融合深度信息和实例分割掩码的新型单目3D目标检测方法。首先,通过深度-掩码注意力融合(DMAF)模块,将深度信息与实例分割掩码结合,以提供更准确的物体边界;其次,引入动态卷积,并利用DMAF模块得到的融合特征引导动态卷积核的生成,以处理不同尺度的物体;再次,在损失函数中引入2D-3D边界框一致性损失函数,调整预测的3D边界框与对应的2D检测框高度一致,以提高实例分割和3D目标检测任务的效果;最后,通过消融实验验证该方法的有效性,并在KITTI测试集上对该方法进行验证。实验结果表明,与仅使用深度估计图和实例分割掩码的方法相比,在中等难度下对车辆类别检测的平均精度提高了6.36个百分点,且3D目标检测和鸟瞰图目标检测任务的效果均优于D4LCN(Depth-guided Dynamic-Depthwise-Dilated Local Convolutional Network)、M3D-RPN(Monocular 3D Region Proposal Network)等对比方法。

基于动态深度可分离卷积神经网络的管道泄漏孔径识别

针对传统模型为提高管道泄漏检测的精度而导致的模型结构复杂度、参数量和计算量大的问题,提出一种基于动态深度可分离卷积神经网络的管道泄漏孔径识别方法;动态卷积层将提取到的泄漏信号特征经过通道注意力权值计算和动态权值融合,通过动态深度可分离卷积层获得更强的特征表达能力,利用全局平均池化层降低网络模型参数,通过全连接层识别管道泄漏孔径。结果表明:新方法具有较高的识别精度,克服了传统模型资源开销大、功耗高的问题,降低了模型的训练时间,提升了管道泄漏孔径的识别速度,可用于工业中的管道泄漏程度监测。

基于动态网络的文本敏感信息感知脑响应检测模型

针对文本敏感信息感知过程复杂和个体差异大造成敏感信息感知脑响应潜伏期不确定性的问题,提出了一种基于动态卷积神经网络的脑响应检测模型——DyCNN_CBAM。该模型通过增加的动态卷积模块,让每层的卷积参数在训练的时候随着输入可变,可提升模型的尺寸与容量。然后在模型第一、二层后增加的注意力机制模块,自动计算贡献度较高的时空信息。实验结果表明:该模型比现有的单尺度模型平均分类准确率提高了4%,F1分数提高6.7%,同时比现有多尺度网络平均分类准确率提高了2%,F1分数提高1.2%。此外,在公开数据集上取得最好的F1分数。由此说明,该网络更够适应文本敏感信息感知脑信号潜伏期抖动性,有效地提升了文本敏感信息检测模型的稳定性。

基于动态蛇形卷积和非跨步卷积的绝缘子缺陷检测

针对复杂场景下绝缘子缺陷检测存在小目标识别困难的问题,提出基于动态蛇形卷积和非跨步卷积的绝缘子缺陷检测方法。首先,算法引入动态蛇形卷积,构造出符合绝缘子特点的特征提取模块,提高对绝缘子及其缺陷的特征提取能力。然后,采用“空间-深度”的非跨步卷积,减少融合过程中的特征丢失。最后,为进一步降低模型复杂度,对模型进行通道剪枝,减少冗余部分。在绝缘子缺陷数据集上进行实验对比,与基准算法相比,绝缘子的破损、污闪以及自爆缺陷的识别率分别提升了5.7%、2.4%和0.8%,改进算法在绝缘子的检测率上提升了0.5%。同时平均精度均值较改进前提升了2.3%,模型大小降低了50.07%。实验结果表明,改进算法在提高绝缘子缺陷小目标检测精度的同时,有效降低了模型大小,对绝缘子缺陷检测的研究具有一定的参考和应用价值。

基于Sobel算子桥接的双编码器路面裂缝检测网络

为提高道路路面裂缝的检测精度,针对路面裂缝的多态性和噪声复杂等问题,提出了一种基于Sobel算子桥接的双编码器路面裂缝检测网络,双编码器由原图编码和梯度编码两部分组成,以解决单编码器容易丢失梯度信息的问题。首先,原图编码结果通过桥接Sobel算子计算8个方向产生梯度编码的编码信息;然后,将原图编码结果与梯度编码结果通过一个多尺度的边缘信息弥补模块,以增强裂缝的边缘信息;最后,引入动态通道图卷积获得通道之间存在的拓扑关系,以突出重要通道的语义特征。研究结果表明:所提出的方法在DeepCrack、CamCrack789和CFD这3个基准数据集上取得较好的结果;综合指标ODS在DeepCrack、CamCrack789和CFD数据集分别为87.75%、85.05%、78.83%。

基于YOLOv8的疫苗胚蛋活性视觉检测

胚蛋活性检测对疫苗生产的质量与安全至关重要。传统机器视觉检测方法严重依赖人为设计的特征提取算法,对图像质量和环境条件要求高,检测结果稳定性和容错性差,导致实际检测过程中的通用性受到限制,为克服这种缺陷,该研究提出一种基于改进YOLOv8的疫苗胚蛋活性检测模型。采用自主设计图像采集装置,采集孵化10~11 d的胚蛋图像,通过几何变换、颜色调整、图像增强等方式构建并扩充数据集;采用Shuffle Net V2替换YOLOv8模型的骨干网络,在保持准确率的同时显著减少了计算复杂度,能更好地部署到嵌入式设备中;在YOLOv8颈部网络的卷积层后添加动态蛇形卷积层,通过其自适应地聚焦于细长和迂回的局部结构,准确地捕捉管状结构的性质特征,从而提高胚蛋检测的准确率;使用EIOU(embedding intersection over union)损失函数,用于适应研究中边界框对齐和形状相似的场景,构建了符合试验中胚蛋图像的网络模型,以实现疫苗胚蛋活性快速、无损、批量检测。试验结果表明,改进YOLOv8模型精确率、召回率、平均精度均值(m AP_(50-95))分别达99.2%、98.2%、96.9%,对比原始YOLOv8模型分别提高了2.0、0.3、1.5个百分点,模型计算复杂度与推理时间相较与原模型分别降低60.9%、60.5%。说明此模型可以更好地实现疫苗胚蛋活性无损检测,为自动化批量检测提供理论依据。

改进YOLOv7的航拍图像目标检测

针对无人机航拍图像中尺度变化、小目标且密集遮挡情况导致检测精度降低的问题,提出一种改进YOLOv7的航拍图像目标检测算法。设计一种联合动态卷积的加权采样模块,从多个维度捕捉特征,提高模型特征提取能力;增加浅层特征检测头,保留更多的细节信息,增强对小目标特征的利用能力;在特征融合部分构建一种具有残差结构的多尺度特征聚合模块(C2-Res2Block),使模型融合丰富的多尺度信息。使用MPDIoU度量替换传统IOU计算边界回归损失,提高模型对密集遮挡目标的定位能力。通过在无人机航拍数据集VisDrone2019上进行实验表明,改进后的算法较原模型mAP@0.5提高了4.3个百分点,mAP@0.5:0.95提高了2.4个百分点,参数量减少了6.81×106,与目前主流的目标检测算法相比也取得更高的检测精度,有效提高了对无人机航拍图像的检测精度,并明显改善对航拍目标的误检和漏检情况。