DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Yoo, Chang-Dong | - |
dc.contributor.advisor | 유창동 | - |
dc.contributor.author | Baek, Seung-Ryul | - |
dc.contributor.author | 백승렬 | - |
dc.date.accessioned | 2013-09-12T01:56:17Z | - |
dc.date.available | 2013-09-12T01:56:17Z | - |
dc.date.issued | 2011 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=467869&flag=dissertation | - |
dc.identifier.uri | http://hdl.handle.net/10203/180751 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 전기 및 전자공학과, 2011.2, [ vi, 42p ] | - |
dc.description.abstract | This thesis studies a method to increase the performance of the image categorization based on the bag-of-visual words (BOV) model. The BOV model represents images as histograms of visual-words describing only their appearance while ignoring their spatial structure, and these histograms are classified based on the classification algorithm such as latent dirichlet allocation (LDA), probablistic latent semantic allocation (pLSA), or support vector machines (SVM). SVM using kernel functions are known to be the state-of-the-art classifier for the BOV model. Recently, the histogram intersection kernel (HIK) is introduced and known to be relatively faster and performs better for histogram features than other popular kernels. Currently, the HIK-based classifier such as intersection kernel SVM (IKSVM) is widely used for image recognition and classification tasks. The performance of the image categorization based on the BOV model mainly depends on the histogram representation and the classification algorithm. The focus of this paper is on the former. The objective of this thesis is learning a discriminative histogram representation for widely used histogram intersection kernel (HIK). The histogram feature space is modeled as the multivariate Gaussian distribution and we propose a learning criterion which can obtain a discriminative histogram representation by increasing inter-class distances of histogram features while decreasing intra-class distances of histogram features. The learning criterion is formulated as the linear programming (LP) problem, which can be optimized by the conventional linear programming solver. The original formulation are successful for binary or ternary classification problem, however, often fails to find the feasible solution and cannot complete the classification process for large database. So, we map histogram feature space into higher dimensional space by incorporating generalized HIK and exponential formulation is obtained. We show that... | eng |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | object recognition | - |
dc.subject | bag of words model | - |
dc.subject | computer vision | - |
dc.subject | machine learning | - |
dc.subject | 물체인식 | - |
dc.subject | 단어군집모델 | - |
dc.subject | 컴퓨터 시각 | - |
dc.subject | 기계학습 | - |
dc.subject | 이미지분류 | - |
dc.subject | image categorization | - |
dc.title | Learning a discriminative histogram representation for histogram intersection kernel (HIK) | - |
dc.title.alternative | 히스토그램 교차 커널을 위한 분별력 있는 히스토그램 표현 학습 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 467869/325007 | - |
dc.description.department | 한국과학기술원 : 전기 및 전자공학과, | - |
dc.identifier.uid | 020093235 | - |
dc.contributor.localauthor | Yoo, Chang-Dong | - |
dc.contributor.localauthor | 유창동 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.