基于S3C2410的桌球自动摆放控制器设计

当前的桌球摆放采用人工摆放方法,准确定位和智能控制性不好。设计一种基于S3C2410的桌球自动摆放控制器,提高桌球摆放的人工智能性。控制器的主机控制分为上位机和下位机两部分组成,上位机实现对桌球台面的视觉信息采集和桌球定位摆放方位分析,采用GT8340嵌入式控制芯片控制上位机的自动对焦高清摄像头,采用多线程串口传输方法桌球分布的标签阅读识别,实现对桌球定位摆放的多维数据的接收、存储和传输。下位机是控制器的主机控制模块,通过程序加载电路加载控制算法,进行桌球摆放信息感知和执行器控制。下位机控制器采用低功耗的S3C2410作为主控芯片,实现控制器的集成设计。测试结果表明,该控制器具有较好的控制精度,对桌球定位摆放的准确性较高,时间开销较小。     

融合自注意力和实体类型知识的实体关系联合抽取模型

从非结构化文本中抽取实体关系三元组是自然语言处理中的主要任务形式之一。目前主流的方法是采用联合式抽取,能够在训练过程中自动捕捉到实体与关系间的依赖知识,提高了实体和关系的抽取效果。但这些方法忽略了实体的类型知识,导致大量的冗余计算和错误结果的产生。鉴于此,文中提出一种融合注意力和实体类型知识的实体关系联合抽取方法。首先,采用预训练模型BERT作为编码器得到句子中各字符的向量表示,再经双向LSTM层处理得到最终的语义表示;其次,基于表示层的结果完成头、尾实体的识别;接着,通过融合不同头实体的语义信息到句子表示中,实现头实体类型约束下的潜在语义关系发现;最后,将头实体和关系分别输入自注意力模块识别出对应尾实体,得到实体关系三元组。通过在公开数据集NYT和WebNLG上的大量实验表明：文中所提模型在实体关系联合抽取任务中的F1值达到了93.2%和93.3%,与当前主流模型相比提升显著。     

运用关系型数据分析方法规划企业级IT系统模型

1前言 关系型数据分析方法是将实体以及实体间的各种联系用关系来表示的一种分析方法。它通过数据（包含实体和属性）、关系和对数据的约束三者组成的数据模型来存放和管理数据。在关系型数据分析方法中,首先要确定实体,否则无法分析关系和属性;其次是确定属性,属性是实体或关系的特征;最后确定关系,关系是实体类之间的关联。关系分为静态关系和动态关系两种。     

时空数据库中模糊实体间拓扑查询的研究

时空数据库中确定空间实体间拓扑查询的研究已日益成熟,但有关不确定实体即模糊实体间的拓扑分析和查询的研究不多。文中依据模糊学基本理论,结合复杂确定空间实体间拓扑关系推理规则,系统阐述模糊实体间的拓扑关系,并给出定量的模糊拓扑谓词。这必将有助于时空数据库拓扑查询系统的完善和发展。     

基于卷积神经网络的医疗护理实体关系抽取

针对医疗护理领域知识复杂性强、数据量大以及对准确度要求较高的问题,该研究提出一种基于卷积神经网络的医疗护理学实体关系抽取方法,实现对护理学语义关系的细粒度文本挖掘。该研究构建了医疗护理学语料标注系统,通过将医疗语料转化为向量特征矩阵,实现了对医疗语料的自动过滤和标注。通过向神经网络模型嵌入所构建的医疗关系语料库,一定程度上提高了模型疾病分类的准确度。在医疗护理学数据集上的实验表明,基于卷积神经网络的模型在指标精确度、召回率、F1值可达到89.78%、87.59%、89.77%。综上所述,该研究提出的基于卷积神经网络的医疗护理学实体关系抽取方法能够有效地抽取医疗语料数据中的实体关系,优于传统的实体关系抽取模型。     

结合Voronoi划分HMRF模型的模糊ISODATA图像分割

为了解决传统模糊聚类图像分割方法对噪声敏感及无法自动准确确定聚类数的问题,提出结合Voronoi划分HMRF模型的模糊ISODATA图像分割方法。利用Voronoi划分将图像域划分为若干子区域,以划分子区域为基本单元定义基于隐马尔科夫随机场(HMRF)模型的模糊聚类目标函数,以解决噪声敏感问题;通过迭代自组织数据分析技术算法(ISODATA)中聚类分裂、合并技术改变聚类数,以实现聚类数的自动确定。对模拟、合成图像和真实图像分割结果的定性和定量分析表明:提出算法不仅可以有效克服噪声和像素异常值对分割结果的影响,而且还能自动准确确定聚类数,实现高精度的自动变类图像分割。      

基于HOURGLASS网络语义关键点提取的光学图像空间目标姿态估计方法

空间目标姿态估计是有效实现各类航天任务的重要前提,基于空间光学观测图像的目标姿态估计关键一环在于快速准确地建立起观测图像与空间目标之间的“二维特征点-三维实体结构”映射关系。传统的方法往往将这一任务分解为特征提取和特征关联两个步骤序贯进行,然而在空间目标光学观测场景中,高动态的光照变化和目标的相对高速运动特点会显著降低图像特征提取的可靠性,影响后续特征关联匹配的正确率并最终降低对空间目标的姿态估计精度。针对这一问题,本文提出了一种基于语义关键点提取的光学图像空间目标姿态估计方法,利用Hourglass网络端到端地提取包含语义信息的关键点,直接实现了光学图像中二维特征点与目标三维实体结构的关联映射,并在此基础上利用EPnP算法求解待估计的目标姿态值。实验结果表明,本文所提的方法能较好地兼顾算法精度与效率,其在仿真数据集上的姿态估计最小误差为0.83°,且在数据降质的情况下平均误差依然优于传统方法。      

潜在语义分析中主题数的确定方法

潜在语义分析的主要思想是通过奇异矩阵分解的方法将高维向量空间模型表示的文档映射到低维的潜在语义空间。在采用潜在语义分析对自然语言文档进行分类的过程中,一个关键的问题是如何确定主题数。通常的做法是在降维过程中缩减保留奇异值数目,使得保留的奇异值的平方和达到所有奇异值平方和的90%。此保留奇异值的数目即主题数,但这种方式并不够准确有效。为能够更加准确地确定主题数,文中提出了另一种自动确定主题数的有效方法。测试结果表明,该方法能够自动有效确定主题数。     

基于深度学习的远程监督关系抽取方法研究

为更好判定远程监督语句中的实体语义关系,实现对语言处理信息的准确提取,提出基于深度学习的远程监督关系抽取方法.借助远程监督方法,获取关系三元组中已存储的信息参量,再通过待学习数据标注的方式,完成远程监督关系的抽取数据集构建.在此基础上,设计监督执行框架,利用已定义的句子级别特征条件,实现对待抽取标签的学习处理,完成基于...     

不确定区域间方向关系的相似性度量方法

空间关系相似性在众领域中具有广泛应用,确定区域间方向关系相似性的研究已较完善,然而这些方法不能直接用于不确定区域间方向关系的相似性度量.本文以扩展的方向关系矩阵EDRM(Extended Direction Relation Matrix)为模型,提出了一种不确定区域间方向关系的相似性度量方法,据此实现了算法SA-EDRM;随后基于目标对象的位置和比例尺这两种基本的空间变化生成不同场景,对算法SA-EDRM进行测试,由实验结果验证了算法的有效性;最后以农业GIS中的具体应用实例说明本文工作的实用价值.     

结合场景描述的文本生成图像方法

研究基于场景描述文本生成对应图像的方法,针对生成图像常常出现的对象重叠和缺失问题,提出了一种结合场景描述的生成对抗网络模型。首先,利用掩模生成网络对数据集进行预处理,为数据集中的对象提供分割掩模向量。然后,将生成的对象分割掩模向量作为约束,通过描述文本训练布局预测网络,得到各个对象在场景布局中的具体位置和大小,并将结果送入到级联细化网络模型,完成图像的生成。最后,将场景布局与图像共同引入到布局鉴别器中,弥合场景布局与图像之间的差距,得到更加真实的场景布局。实验结果表明,所提模型能够生成与文本描述更匹配的图像,图像更加自然,同时有效地提高了生成图像的真实性和多样性。     

An MPEG-4 Renderer for High Quality Video Composition and Texture Mapping

This paper presents the hardware architecture of a co-processor supporting the real time rendering of all 2D natural or synthetic visual objects proposed by the MPEG-4 standard as well as sprite decoding. It enables the composition and the transformation of natural video objects and the texture mapping on triangles, allowing the 2D-mesh decoding. This architecture is able to render scenes that are compliant with MPEG-4 Main Profile, Level3 and Hybrid Visual Profile. The co-processor is designed to be used in a shared memory system architecture. It consists in a dedicated implementation that seeks the best compromise between cost and performances. In a first step, a software model is used to guarantee the visual quality of the rendered scene and to validate the algorithmic choices. Then, the complexity and performances of this novel architecture are evaluated. Finally, a behavioral model validates the architectural choices.     

Image-based rendering and 3D modeling: A complete framework

Multi-viewpoint synthesis of video data is a key technology for the integration of video and 3D graphics, as necessary for telepresence and augmented-reality applications. This paper describes a number of important techniques which can be employed to accomplish that goal. The techniques presented are based on the analysis of 2D images acquired by two or more cameras. To determine depth information of single objects present in the scene, it is necessary to perform segmentation and disparity estimation. It is shown, how these analysis tools can benefit from each other. For viewpoint synthesis, techniques with different levels of tradeoff between complexity and degrees of freedom are presented. The first approach is disparity-controlled view interpolation, which is capable of generating intermediate views along the interocular axis between two adjacent cameras. The second is the recently introduced incomplete 3D technique, which in a first step extracts the texture of the visible surface of a video object acquired with multiple cameras, and then performs disparity-compensated projection from the surface onto a view plane. In the third and most complex approach, a 3D model of the object is generated, which can be represented by a 3D wire grid. For synthesis, this model can be rotated to arbitrary orientations, and original texture is mapped onto the surface to obtain an arbitrary view of the processed object. The result of this rendering procedure is a virtual image with very natural appearance.     

Image segmentation through graph-based clustering from surface normals estimated by photometric stereo

A method for segmenting 2D images based on 3D shape information is proposed. First, a robust photometric stereo technique estimates the 3D normals of the objects present in the scene for every image pixel. Then, the image is segmented by grouping its pixels according to their estimated normals through graph-based clustering. Differently from other image segmentation algorithms based on intensity, colour or texture, the regions of which are determined by the visual appearance of the depicted objects, the regions obtained with the proposed technique only depend on the 3D shapes of those objects. This can be advantageous for higher level scene understanding algorithms. This technique is especially suited to poorly illuminated scenarios and utilises a conventional camera and six inexpensive strobe lights.     

Real-Time Generation of Dynamic Infrared Scene

Although many models have been put forward to realize static infrared scene, they could not generate dynamic infrared scene real time in interactive way. In this paper a new method is proposed to solve the problem. We first model the targets and background of infrared scene based on the hybrid way of geometry and multi-spectral texture images. Then considering the attenuation effect of atmosphere and the noise mechanic of infrared image sensor, we present an infrared depth image model to generate dynamic images of the objects in the scene from different viewpoint. The complexity of infrared dynamic scene is thus reduced greatly and the reality of infrared scene is improved. Finally, real-time walkthrough for infrared scene is successfully realized and the average walkthrough speed is larger than 25 frames per second.     

A Method for 3D Scene Description and Segmentation in an Object Record

ＡＭｅｔｈｏｄｆｏｒ３ＤＳｃｅｎｅＤｅｓｃｒｉｐｔｉｏｎａｎｄＳｅｇｍｅｎｔａｔｉｏｎｉｎａｎＯｂｊｅｃｔＲｅｃｏｒｄ￥ＣｈｅｎＴｉｎｇｂｉａｏ（ＤｅｐａｒｔｍｅｎｔｏｆＲａｄｉｏＥｎｇｉｎｅｅｒｉｎｇ，ＮａｍｉｎｇＵｎｉｖｅｒｓｉｔｙｏｆＰｏｓｔｓ...     

3D information extraction using Region-based Deformable Net for monocular robot navigation

This paper extends the Region-based Deformable Net (RbDN) technique described in [1] to extract the 3D information of all the objects in the scene from a single moving camera. The technique is used for segmenting real-time video sequences captured from a single moving camera. The deformation process tracks the changes in the location and the shape of the segments across the frames. These changes along with the camera displacement are used to estimate the 3D information. The algorithm is completely autonomous and does not require pre-knowledge, training, or assumption about the contents of the sequence. It can handle the difficult case where the motion of the camera is parallel to its optical axis. It can also estimate the distances to objects that are more than 100 m away as long as the camera displacement is over 10% of the expected distance to the objects.     

Object Oriented Motion-Segmentation for Video-Compression in the CNN-UM

Object-oriented motion segmentation is a basic step of the effective coding of image-series. Following the MPEG-4 standard we should define such objects. In this paper, a fully parallel and locally connected computation model is described for segmenting frames of image sequences based on spatial and motion information. The first type of the algorithm is called early segmentation. It is based on spatial information only and aims at providing an over-segmentation of the frame in real-time. Even if the obtained results do not minimize the number of regions, it is a good starting point for higher level post processing, when the decision on how to regroup regions in object can rely on both spatial and temporal information. In the second type of the algorithm stochastic optimization methods are used to form homogenous dense optical vector fields which act directly on motion vectors instead of 2D or 3D motion parameters. This makes the algorithm simple and less time consuming than many other relaxation methods. Then we apply morphological operators to handle disocclusion effects and to map the motion field to the spatial content. Computer simulations of the CNN architecture demonstrate the usefulness of our methods. All solutions in our approach suggest a fully parallel implementation in a newly developed CNN-UM VLSI chip architecture.     

基于单线激光雷达的三维形貌重建方法研究

现有三维扫描激光雷达存在成本高、扫描范围有限等缺点,为解决这些问题,本文提出一种基于单线激光雷达的三维形貌重建系统,并对系统进行误差补偿以提高系统的准确度.该系统采用可编程逻辑控制器控制的步进电机旋转单线激光雷达的扫描平面,实现对实际空间场景的扫描.通过上位机控制扫描装置的工作过程同时获取原始点云数据,进而针对原始点云...