DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Kim, Jin-Hyung | - |
dc.contributor.advisor | 김진형 | - |
dc.contributor.author | Lee, Seong-Hun | - |
dc.contributor.author | 이성훈 | - |
dc.date.accessioned | 2013-09-12T01:46:17Z | - |
dc.date.available | 2013-09-12T01:46:17Z | - |
dc.date.issued | 2013 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=515420&flag=dissertation | - |
dc.identifier.uri | http://hdl.handle.net/10203/180354 | - |
dc.description | 학위논문(박사) - 한국과학기술원 : 전산학과, 2013.2, [ vi, 78 p. ] | - |
dc.description.abstract | Text contained in scene images provides the semantic context of the images. For that reason, robust extraction of text regions is essential for successful scene text understanding. However, separating text pixels from images still remains a challenging issue because of uncontrolled lighting conditions and complex backgrounds. In addition, any prior knowledge about text regions is usually unavailable in the scene image. To robustly extract text regions in the scene image, we propose a two-stage probabilistic framework that combines top-down knowledge of the text and bottom-up image processing. To deal with the various conditions of scene images, bottom-up image processing produces multiple image segmentations which represent different types of interpretations of the scene images. Our image segmentation algorithm seamlessly combines color, texture, and edge to isolate text regions from backgrounds without the loss of small details of text regions. Even though single segmentation cannot find all text regions, the set of all segmented regions obtained by multiple segmentations could contain all text regions. The proposed two-stage conditional random field approach generates multiple proposals of text regions and integrates them into textlines by utilizing the properties and hierarchical structures of the scene text. A region-oriented representation of the image is used to build a random field in each stage of the CRF model for identifying the possibilities of the text regions at local and global levels. In the first stage, proposals of text regions are generated by removing apparent non-text regions in each segmentation by using a local CRF model. The local CRF model couples to local image features such as color, edge, and textures as well as global character contexts such as compactness, aspect ratio, and compatibility between characters. In the second stage, the proposed system selectively integrates the multiple proposals to find plausible combinations of text ... | eng |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | Scene Text Extraction | - |
dc.subject | Two-Stage CRF Models | - |
dc.subject | Multiple Image Segmentations | - |
dc.subject | Component | - |
dc.subject | 자연 영상 내 글자 추출 | - |
dc.subject | 2단계 CRF 모델 | - |
dc.subject | 다중 영상 분할 | - |
dc.subject | 컴포넌트 | - |
dc.subject | 글자 후보 통합 | - |
dc.subject | Character Proposal | - |
dc.title | A probabilistic framework integrating multiple proposals of text regions for scene text extraction | - |
dc.title.alternative | 자연 영상 내 글자 추출을 위한 확률 모델 기반의 글자 후보 통합 시스템에 관한 연구 | - |
dc.type | Thesis(Ph.D) | - |
dc.identifier.CNRN | 515420/325007 | - |
dc.description.department | 한국과학기술원 : 전산학과, | - |
dc.identifier.uid | 020065120 | - |
dc.contributor.localauthor | Kim, Jin-Hyung | - |
dc.contributor.localauthor | 김진형 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.