A Study on the Updating Mechanism of SVM Binary Tree Architectur
- 주제(키워드) incremental clustering , SVM-BTA , updating
- 발행기관 고려대학교 대학원
- 지도교수 박대희
- 발행년도 2013
- 학위수여년월 2013. 8
- 학위구분 석사
- 학과 일반대학원 컴퓨터정보학과
- 원문페이지 42 p
- 실제URI http://www.dcollection.net/handler/korea/000000045514
- 본문언어 영어
- 제출원본 000045764311
초록/요약
In this paper, we propose a novel incremental updating mechanism of SVM-BTA (support vector machine binary tree architecture), which has been applied to face recognition based watch list identification system. It should be guaranteed that the watch list identification system can recognize faces with high accuracy and update in real time, especially when new criminal images are added to the watch list database system. For the above requirements, overall identification rate and SVM retraining time are considered as two major factors when we update the SVM-BTA architecture. Firstly, 2-means clustering algorithm is used repeatedly for SVM-BTA generation. Then, two incremental updating algorithms for SVM-BTA are proposed, which are combined with partition phase and updating phase. In the first proposed method, hierarchical 2-means incremental algorithm, vigilance parameter and separability measure are used together for the SVM-BTA incremental updating problem. In the second proposed method, fast hierarchical 2-means incremental algorithm, we define the so-called SVM-BTA important degree and use it to make an enhanced form of vigilance parameter, which can further reduce the SVM retraining time compared with the first proposed method. Two real face databases, Yale face database and KUFD (Korea University Face Database), are used in the experiment. The two proposed incremental clustering algorithms and other methods are applied to the SVM-BTA incremental updating problem for comparison. In the testing phase, we use the well-known libsvm to identify new incremental face images and the experiment results illustrate the efficiency and effectiveness of the proposed methods.
more목차
Abstract i
Contents iii
List of Figures v
1. Introduction 1
2. Related Work 3
2.1 SVM Multi-class Classifier with Binary Tree Architecture 3
2.1.1 Support Vector Machine 3
2.1.2 Support Vector Machine Binary Tree Architecture 5
2.1.3 SVM Binary Tree Architecture 7
2.2 Incremental K-Means Clustering Algorithm 9
2.2.1 The K-Means Algorithm 9
2.2.2 Incremental K-Means Clustering 11
3. Proposed Updating Mechanism 13
3.1 Hierarchical 2-Means Incremental Algorithm 13
3.1.1 Overall Updating Mechanism 13
3.1.2 Vigilance Parameter of SVM-BTA 16
3.1.3 Separabily Measure of SVM-BTA 18
3.2 Fast Hierarchical 2-Means Incremental Algorithm 21
3.2.1 Important Degree of SVM-BTA 21
3.2.2 Enhanced Form of Vigilance Parameter 23
4. Experiment 24
4.1 Data Set and Experiment Environment 24
4.2 Experiment Method 24
4.2.1 Performance Measure 24
4.2.2 Experiment Method 25
4.3 Results and Analysis 26
4.3.1 Identification Rate 26
4.3.2 Retraining Time 27
4.3.3 Overall Experiment Results 28
5. Conclusions 30
References 32
