抽取

In the first approach the metadata extraction is viewed as a classification problem. Three statistic classification models, including Support Vector Machines, Perceptron with Uneven Margins, and Maximum Entropy are studied. The approach suffers from the problems of ignoring the variance in lengths of documents and ignoring the dependency of text units within documents. To solve the problems, our second approach takes the metadata extraction as a sequence data labeling problem and makes full use of the sequential information.

论文首先研究了基于分类的元数据抽取方法，考察了支持向量机、非对称间隔感知机和最大熵模型在ADME中的应用；然后针对分类方法中存在的忽略文档之间的差异、不考虑文档上下文等问题，提出了基于序列数据标注模型的抽取方法，并结合最大熵马尔可夫模型、投票感知机模型和条件随机场模型进行了研究；最后本文根据前两种方法的设计思想和优缺点，研究提出一种新的结构预测模型—— CDASP（Coordinate Descent Algorithm for Structure Prediction），并将其应用于元数据的自动抽取中。

The key issue of decimation is how to realize the digital filter before decimation, especially for design and implementation of multistage decimation filter.

实现抽取的关键问题是如何实现抽取前的数字滤波，特别是多级抽取时滤波器的设计与实现。

Four half band filter decimation modules are designed which can cascade connected to implement 2K decimator.

设计和实现多个半带滤波器抽取模块，组合成多级抽取器实现2~K倍抽取。

This method basic thought is: First, extracts the time and the place phrase which in the news essential factor the event occurs; Then, through extraction time and place phrase extraction news content; Finally, through calculates their similarity according to the study news content to judge the news homepage the heavy multiplicity.

该方法的基本思想是：首先，抽取新闻要素中关于事件发生的时间和地点短语；然后，通过抽取的时间和地点短语抽取新闻的内容；最终，根据学习的新闻内容通过计算它们的相似度来判断新闻网页的重复度。

In order to further extract open contour segment, such as circular arc, from a closed contour, a technique that combines parameters identification and contour vertices extraction was put forward.

为了抽取开放的规则轮廓段，提出了特徵参数识别和特徵轮廓抽取交互进行、有机结合的技术。为了抽取开放的自由曲线段，提出了基于转角映射的特徵连接点辨识技术。

There are several different types of sample surveys: Simple random sampling: every member of the target population is equally likely to be selected, and the selection of a particular member of the target popula- tion has no effect on the other selections.

有以下几种抽样统计：简单随机抽样：不同目标人群中的每位成员具有相同的被抽取几率，抽取目标人群中的任意成员对其他成员的抽取没有影响。

Methods: During the April 2003 to December 2004, from the patients with different position colorectal cancer beyond Dukes C stage who were performed laparoscopic radical operation random draw 20 patients as experiment group, and from patients with different position colorectal cancer beyond Dukes C patients who were performed open radical operation 20 patients took randomly as contrast group. We draw thepatients blood about 5 ml from antecubital vein at the time before and after the operation about 30 minutes respectively, preserved the blood in deep hypothermia circumstance,cytrokeratin-20(CK-20) as the micrometastasis maker, using the RT-PCR technique,G3PDH as the intra-contrast gene, quantitatively detect the amounts of CK-20 in the blood with the gelatum imagery system and the computer analyzing software on the gelatum imagery system ,in order to compare the influence of the laparoscopic and the open operation on the micrometastasis.

在我院2003年4月—2004年12月间不同部位的临床分期在Dukes C期之内的接受了腹腔镜结直肠癌根治术结直肠癌患者中随机抽取20例作为试验组，另外在同期的接受开腹手术患者中随机抽取20例作为对照组，分别抽取术前30min和术后30min的肘前静脉血5ml，集中低温保存，以细胞角蛋白(cytekeratin-20)CK-20为微转移的标志物，应用RT-PCR技术，以G3PDH为内对照基因，应用凝胶成像系统及其计算机凝胶成像系统分析软件来定量检测手术前、后患者外周静脉血中CK-20含量，比较腹腔镜和开腹的手术方式对肿瘤微转移的影响。

The dissertation focuses on the topics for automatic, accurate, and efficient metadata extraction from general and unstructured documents, on the basis of the state of the work on metadata extraction. After analyzing the limitations of current work, we propose a model for Automatic Document Metadata Extraction, and formulize the problem under the framework of statistical learning.

首先，论文针对目前元数据抽取在处理的文档对象和采用的抽取方法方面所存在的问题和不足，以统计学习理论和信息抽取领域相关研究为指导，结合非结构化文档的特点，提出了文档元数据自动抽取（Automatic Document Metadata Extraction，ADME）模型，给出其形式化描述，进而介绍模型工作机理和处理流程。

Compared with methods appeared in literatures, our algorithm, with high accuracy, sensitivity and correctness, complete adaptively the correct distinction between interferential fringes and backgrounds. In chapter four, methods for the computer operatable description of line and region features are discussed, which contains: Computation curve's curvature in quantity, Fringes' separation distance , Contour's smoothness of region, Similarity of varied curves.

在第三章中，本文研究的是以线条为直接目标的图象分割方法——条纹抽取，在研究分析了许多条纹抽取方法以后，针对激光全息焊点缺陷检测这一课题中的复杂数字光学图象LHII条纹抽取进行了多方面的研究，主要有：分别从点对应、邻域特性对应推导并提出了几个信息分离算法；结合光学图象形成方面的知识，提出了变周期余弦分割二值化方法；结合当前条纹走向、毗邻条纹走向及可靠条纹点约束条纹跟踪，在许多方面，改进并发展了常规的条纹跟踪算法，实现了无人-机交互的全自动条纹跟踪抽取。

In degraded PSM matching classification, errors modification method of degraded PSM is presented and the matching tensor of every canonical scattering centers is given. In Chapter 4, the theory of wideband millimeter-wave radar target identification is introduced. It is indicated that, for HR radar target identification, description of temoral relations among features and pattern recogntion adaptive to variation of target angles are of key importance. A rule-based pattern recogniton method of sequential reasoning is proposed, which uses a series of rules to describe relations of features variation caused by target angles variation and has the advantages of adaption to unrandom variation and false probability control in classification over traditional statistical pattern recognitiop method. Feature extraction is crucial step in target identification. In range profiles identification application, features are extracted by means of range domain pre-processing algorithm, spacial and amplitude visual computation directly from range profiles and transform algorithm based on range profiles. Visual or transformed features are either sufficiently convinced or necessarily convinced and both of them are effective and robust to range profiles identification.

在第四章，首先对宽带毫米波体制背景下目标识别方法的一些特点进行了阐述并指出，在毫米波雷达目标识别中，特征之间动态关系的描述以及能适应目标姿态角变化的模式识别方法的研究乃是要解决的关键问题；进而提出了一种基于规则库的序贯推理模式识别新方法，在这种方法的研究中，主要包括序贯推理规则库的规则排列与特征选用顺序以及规则库的收敛等问题；这种方法克服了传统的统计模式识别方法特征利用效率不高、难以适应特征值的非随机性变化的缺点，把姿态角变化所引起的特征的变化用一系列规则加以表示，其优点是能适应特征值的非随机性动态变化，并能控制分类过程中的差错概率α；特征抽取是目标识别中的关键步骤，在基于目标距离像的特征抽取方法研究中，提出了距离空间域的预处理算法、距离空间域与幅度域的直观特征抽取方法以及基于目标距离像变换分析的特征抽取方法。

In the United States, chronic alcoholism and hepatitis C are the most common ones.

在美国，慢性酒精中毒，肝炎是最常见的。

If you have any questions, you can contact me anytime.

如果有任何问题，你可以随时联系我。