1、 1 基于图象的可配置机场跑道周边安全保障系统的识别单元 Abstract This paper proposes a novel classifier combination system that can be used by Deployable Video Based Ramp Security System for Airliner Perimeter Protection under dynamically varying environments. The proposed method adopts the concept of context-awareness and th
2、e similarity between classes, and the system working environments are learned (clustered) and identified as environmental contexts. The proposed method fitness correlation table is used to explore the most effective classifier combination for each identified context. We use t-test for classifier sel
3、ection and fusion decision and proposed context modeling. The group of selected classifiers is combined based on t-test decision model for reliable fusion. The knowledge of individual context and its associated chromosomes representing the optimal classifier combination is stored in the context know
4、ledge base. Once the context knowledge is accumulated the system can react to dynamic environment in real time. Keyword: Bayesian Classifier Context-awareness Classifier Combination 2 摘要 本论文提出了一个创造性的分类器合并系统，这个系统可以在不同的环境因素下用做可部署机场周遍安全系统的识别部分。文中所提及的方法采用了环境因素纳侦测以及类之间的相似性的概念，并且系统的运行环境会被学习和识别为环境因素。文中提及的方
5、法健壮性关系表是用来找到最有效的分类器合成方法，从而能够最好的适应已经识别的环境因素。我们利用T-TEST来做分类器的选择，结合的抉择以及环境建模。选择好的一组分类器将根据 T-TEST来做可靠的结合。单个环境因素以及与之相关的其他环境因素将被存储在环境因素库中，一旦环境因素库中的 数据达到一定量，系统将可以实时的根据环境因素的变化而做出正确的判断。 关键字： 贝叶斯分类器 环境因素侦测 分类Undergraduate Dissertation Xiamen University 1 Content Abstract.1 摘要 .2 Content.1 目录 .1 Chapter 1 Intr
6、oduction.1 Chapter 2 Model of Context-Aware Classifier Combination Scheme.3 2.1 Exhaustive Classifier Combination.3 2.2 Context-Aware Classifier Combination .5 2.3 Summary .7 Chapter 3 Bayesian Classifier and Context-Aware Classifier Combination Scheme .8 3.1 Bayesian Decision Theory.8 3.1.1. Introd
7、uction .8 3.2 Bayesian Classifier . 10 3.2.1. Multi-Category Case . 10 3.2.2. Two-Category Case . 13 3.2.3. Multivariate Density. 14 3.3 Classifier Fusion . 28 3.3.1. Nave Bayes Classifier fusion . 28 3.3.2. The training part . 28 3.3.3. Bagging and Boosting . 29 3.3.4. To combine single classifiers
8、 . 29 3.4 Summary . 30 Chapter 4 T-Test Decision Model for Classifier Combination. 32 4.1 Introduction . 32 4.2 Mathematics basic for T-Test . 33 4.3 T-Test for Classifier Combination . 34 4.4 Summary . 35 Chapter 5 Self-learning Part . 35 5.1 Introduction . 35 5.2 Scheme of Self-Learning Part . 35
9、5.3 Summary . 35 Chapter 6 Experiment on Face Recognition . 37 6.1 Experimental Introduction. 37 6.2 Experimental Results. 38 Chapter 7 Conclusion. 40 References . 41 Acknowledgement . 42 Undergraduate Dissertation Xiamen University 1 目录 Abstract.1 摘要 .2 Content.1 目录 .1 Chapter 1 简介 .1 Chapter 2 环境因
10、素侦测分类器框架模型 .3 2.1 穷举型分类器合并 .3 2.2 环境因素侦测 型分类器合并 .5 2.3 摘要 .7 Chapter 3 贝叶斯分类以及 环境因素侦测 分类器合并框架 .8 3.1 贝叶斯 决定论 .8 3.1.1. 简介 .8 3.2 贝叶斯分类器 . 10 3.2.1. 多类情况 . 10 3.2.2. 二类情况 . 13 3.2.3. 多变量密度 . 14 3.3 分类器融合 . 28 3.3.1. Nave 贝叶斯分类器 融合 . 28 3.3.2. 训练部分 . 28 3.3.3. Bagging 和 Boosting . 29 3.3.4. 将多个单个分类器合并
11、 . 29 3.4 摘要 . 30 Chapter 4 用于分类器合并的 T-Test 决定模型 . 32 4.1 简介 . 32 4.2 T-Test 的数学基础 . 33 4.3 T-Test 用于分类器合并 . 34 4.4 摘要 . 35 Chapter 5 自我学习部分 . 36 5.1 简介 . 36 5.2 自我学习部分框架 . 36 5.3 摘要 . 36 Chapter 6 用于人脸识别的测试 . 37 6.1 测试简介 . 37 6.2 测试结果 . 38 Chapter 7 结论 . 40 引用 . 41 致谢 . 42 Undergraduate Dissertatio
12、n Xiamen University 1 Chapter 1 Introduction In contrast to the huge amount of research in this active area 1, 2, 3, little work has been done on combining the specific classifier: the k nearest neighbor classifier (kNN). Classifier decision methods for identification are illustrated their better re
13、liance on recognition than single classifier and implemented in various ways. Clustering the data set into different regions is added value to recognition systems by finding specific sophisticated system for particular region in ways as selection and fusion of classifiers 4, 5, 6. This approach will
14、 use the ensemble as an individual means and covariance in a class. In this paper, we propose the method Bayesian based similarity between classes that is clustered using unsupervised method for varying illuminant. In additional method, we discuss about an adaptive run-time framework of classifier c
15、ombination by employing the concept of context-awareness and the genetic algorithm. The proposed method can construct most effective classifier combination by selecting preprocessing, feature representation. System contexts associated with the system working environments are learned by clustering th
16、em. Context learning can be performed by an unsupervised learning method such as SOM, FuzzyArt, etc. We assume that an environmental context changes continuously. That is, we exclude an impulse style of context changes in this paper. The structures of classifier combination are encoded as artificial
17、 chromosomes, and genetic algorithm is used to explore a most effective classifier combination for each identified context. The knowledge of an individual context and its associated chromosomes of effective classifier combination are stored in the context knowledge base. There is no general approach
18、 or theory for efficient classifier combination, yet 6, 7. We assume that classifier systems include preprocessing, feature representation, and class decision components. The design of classifier combination can be divided into two stages: 1) the generation of possible classifiers using the classifi
19、er components and 2) the decision of classifier combination scheme using an aggregation method. The classifier components are a set of (combined) preprocessing, feature representation, and class Undergraduate Dissertation Xiamen University 2 decision methods. We assume that s (combined) preprocessin
20、g, t different feature representation, u class decision, methods are provided, and we focus on combining them to produce a output of high performance in terms of accuracy. The proposed method primarily aims at robust object recognition under uneven environments by selection and fusion of different c
21、lassifiers. We method searches the group of most effective classifier system for each environmental context by proposed t-test method. The chosen best classifiers are then combined using t-test based decision model derived from the reliability condition of combination. The main difference of the pro
22、posed classifier combination method from other methods is that it can combine classifiers in accordance with the identified context in reliable way. The proposed method adopts the strategy of context knowledge accumulation so that classifier combination scheme is able to adapt itself under changing
23、environments based on experience. Once the context knowledge base is constructed, the system can react to changing environments at run-time. The context retrieve part is omitted in this paper because we can easily use some OpenCV functions like cvHaar and cvContour and so on to get the background in
24、formation of images, because GE will provide us all the crucial information about the image, like human, human segmentation, important things. Undergraduate Dissertation Xiamen University 3 Chapter 2 Model of Context-Aware Classifier Combination Scheme 2.1 Exhaustive Classifier Combination There is
25、no general approach or theory for efficient classifier combination, yet 8. We discuss about the framework of context-aware classifier combination with the capability of adaptation using accumulated context knowledge. The implementation issue of the proposed adaptive framework employing the context-awareness and the genetic algorithm will be discussed in the next section. A
