基于生成对抗网络的半监督图像语义分割

提出了一种基于生成对抗网络的语义分割模型,包括一个全卷积语义分割网络以及一个判别网络,其中语义分割网络负责生成与输入图像对应的语义分割图,判别网络负责检测分割图与真实标签的区别,以促使分割网络改进分割效果。为了更好的提取全局结构信息,语义分割网络中采用了金字塔池化模块,对不同规模的空间区域进行池化操作。另外,为了应对语义分割训练数据集人工标注成本过高的问题,利用判别网络生成伪标签协助语义分割网络进行训练,从而实现了半监督训练效果。模型在PASCAL VOC2012数据集中进行了测试,结果表明该模型在全监督和半监督条件下均优于已有方法。     

模具结构对连续纤维增强热塑性复合材料性能的影响

采用自行设计的两种不同结构的熔融浸渍模具制备了连续玻璃纤维增强聚丙烯预浸带，测试了模具结构对预浸带的孔隙率、纤维断裂率、界面形貌、纤维分散均匀度和拉伸强度的影响，建立了纤维浸渍模型和纤维断裂模型，并通过理论模型对预浸带的孔隙率和断裂率进行理论预测。结果表明，本文建立的数学模型能够有效预测预浸带的浸渍程度和纤维断裂率，可用于浸渍模具结构的优化设计；在本文范围内，与波浪形模具相比，斜齿形模具的多楔形区结构可以有效地降低预浸带孔隙率和提升纤维分散程度；波浪形模具的流道圆角半径较大，楔形区个数较少，与斜齿形模具相比，可有效降低纤维断裂率并提升拉伸性能。     

基于JAVA建立树形结构的算法优化

树形结构在程序设计过程具有非常重要的作用,但树形结构的构建相对比较复杂。通过研究设计并实现一种从普通Java对象中建立树形结构的优化算法。     

基于曼哈顿距离构建非参数Q矩阵修正方法

将被试得分、理想反应距离和被试得分异常原理相结合,并加入属性计分下的被试得分特性,开发了用于多级评分情境下属性计分曼哈顿距离法(SA-MD),在不同条件下验证了SA-MD的稳定性和适宜性.通过模拟研究和实证研究表明:(i)从逻辑推导出SA-MD用于多级评分情境下Q矩阵修正更合理;(ii)在多种条件中,SA-MD 均有较优的修正效果,适用范围更广,稳定性更高;(iii)当小样本测验进行Q矩阵修正时,使用SA-MD方法可获得更优的效果.     

一种三维DNA构象的交互式参数化模拟方法

提出一种适用于移动平台的交互式参数化DNA三维动态模拟方法.依据不同的DNA模式参数,利用基于自然特征的AR识别技术和曲线,直观形象地展示DNA的多种三维构象,动态模拟碱基互补配对、碱基对边堆叠边螺旋、 DNA复制的边解旋边半保留半不连续复制等微观过程.研发DNA三维动态模拟原型系统,实验结果表明,通过视、听、触、实物等多通道交互,有利于用户随时随地自主使用,增强用户对DNA知识的理解、掌握,引导用户进行更深层次的思考,提高知识的传递效率.     

迹为整数的3×3阶正交矩阵的谱

讨论了3×3阶正交矩阵的特征值和迹的关系,证明了迹为整数的3×3阶正交矩阵的谱可由迹确定,为应用广泛的3×3阶正交矩阵的谱的计算提供了简单实用的方法.     

基于数据场联合PRI变换与聚类的雷达信号分选

基于脉冲描述字进行雷达信号分选时,传统聚类算法需要预先人工设定聚类中心和聚类数目。针对该问题,提出一种基于数据场理论联合脉冲重复间隔(pulse repetition interval,PRI)变换与聚类的雷达信号分选新方法。首先,依据数据场理论,基于势值大小实现干扰点剔除,而后利用PRI变换算法进行PRI估计,依据PRI估计值将归一化脉冲描述字数据预分类,进而以各类数据集中心间的欧氏距离小于辐射因子为准则进行类别合并,自动得到初始聚类中心和聚类数目,最后通过改进K-Means算法完成聚类分选。仿真实验表明:所提方法能够应对存在频率捷变,重频参差、抖动、参数交叠、局部脉冲丢失的复杂信号环境,分选正确率明显提升。     

A Method for Time-Series Location Data Publication Based on Differential Privacy

In the age of information sharing, logistics information sharing also faces the risk of privacy leakage. In regard to the privacy leakage of time-series location information in the field of logistics, this paper proposes a method based on differential privacy for time-series location data publication. Firstly, it constructs public region of interest(PROI) related to time by using clustering optimal algorithm. And it adopts the method of the centroid point to ensure the public interest point(PIP) representing the location of the public interest zone. Secondly, according to the PIP, we can construct location search tree(LST) that is a commonly used index structure of spatial data, in order to ensure the inherent relation among location data. Thirdly, we add Laplace noise to the node of LST, which means fewer times to add Laplace noise on the original data set and ensures the data availability. Finally, experiments show that this method not only ensures the security of sequential location data publishing, but also has better data availability than the general differential privacy method, which achieves a good balance between the security and availability of data.     

Hierarchical Community Detection Based on Partial Matrix Convergence Using Random Walks

Random walks are a standard tool for modeling the spreading process in social and biological systems.But in the face of large-scale networks,to achieve convergence,iterative calculation of the transition matrix in random walk methods consumes a lot of time.In this paper,we propose a three-stage hierarchical community detection algorithm based on Partial Matrix Approximation Convergence(PMAC) using random walks.First,this algorithm identifies the initial core nodes in a network by classical measurement and then utilizes the error function of the partial transition matrix convergence of the core nodes to determine the number of random walks steps.As such,the PMAC of the core nodes replaces the final convergence of all the nodes in the whole matrix.Finally,based on the approximation convergence transition matrix,we cluster the communities around core nodes and use a closeness index to merge two communities.By recursively repeating the process,a dendrogram of the communities is eventually constructed.We validated the performance of the PMAC by comparing its results with those of two representative methods for three real-world networks with different scales.     

基于热中子探测评价裂缝的示踪剂特性研究

针对裂缝热中子示踪探测方法,从中子双组扩散理论入手结合蒙特卡洛数值模拟方法,分析压裂前后地层热中子分布影响因素,并模拟不同类型示踪剂及含量的热中子裂缝响应规律,筛选最佳示踪剂并给出其在支撑剂中含量的上、下限。结果表明:压裂前后热中子计数变化量主要受裂缝宽度和标记支撑剂中示踪剂含量的影响;钆对裂缝宽度变化反应最灵敏,钆元素是作为标记支撑剂的最佳示踪元素;随支撑剂中Gd 2O 3含量的增加,热中子计数率降低,热中子计数变化量D增加;示踪剂中氧化钆的上、下限值随裂缝宽度呈指数降低,且当裂缝宽度为1.0 cm时,Gd 2O 3的质量分数上限为0.68%,下限为0.03%。