Image segmentation by pixel-to-pixel connection using mean-shift vectors
저자
발행사항
포항 : 포항공과대학교 대학원, 2017
학위논문사항
학위논문(박사)-- 포항공과대학교 대학원 : 전자전기공학과 2017. 8
발행연도
2017
작성언어
영어
주제어
발행국(도시)
경상북도
형태사항
; 26 cm
소장기관
Image segmentation is a technique to decompose an image into contextually consistent regions by grouping the pixels having similar characteristics. As a result of the decomposition, the number of data is dramatically reduced, and the effect of noise is alleviated. Furthermore, high level information is able to be extracted from the segmented regions. Image segmentation is therefore widely used as pre-processing of various image processing and computer vision algorithms, such as image in-painting, image enhancement, object tracking, 2D-to-3D conversion, motion estimation, and stereo matching.
Mean-shift-based image segmentation, one of the representative image segmentation methods, provides an efficient way to locate the modes without estimating the underlying density where the modes denote the local maxima of underlying density of pixels. For each pixel, it computes the weighted mean of the pixels within bandwidths. In mean-shift-based segmentation, the weights are selected so that the weighted mean is always located in the way of maximal increase of density. After the weighted mean is computed, each pixel is shifted to its weighted mean. This process is called the mean-shift procedure. The mean-shift procedure is repeated until the weighted mean is converged. Because the weighted mean is located in the way of maximum increase of density, the converged point will be the mode of the pixel.
Mean-shift-based segmentation improves the homogeneity of intra-segment pixels. The mode is the local maximum of the underlying density, and thus the modes of intra-segments pixels are located nearby. Hence, by replacing the range components of each pixel with those of its modes, the homogeneity of intra-segment pixels is greatly improve. As a result of improving the homogeneity, the inhomogeneity of inter-segment pixels is also improved. Consequently, the boundary of objects becomes more distinguishable, which helps mean-shift-based segmentation produce accurate segmentation results. Furthermore, when repeating the mean-shift procedure, mean-shift-based segmentation continuously considers the range spectrum of surrounding area. Consequently, it provides consistent results for natural images that have various and arbitrary range spectra. However, its computational complexity is extremely high, and thus it may impede the processing of the entire system that uses mean-shift-based segmentation as pre-processing. Mean-shift-based image segmentation therefore needs to be accelerated by exploiting parallelism. Unfortunately, as the mean-shift procedure is iterated, data locality is degraded. As a result, memory bandwidth is seriously decreased. The reduced memory bandwidth significantly degrades the effectiveness of parallel processing.
To reduce the computational complexity and to improve the effectiveness for parallel processing, it is required to eliminate the iterative process. There are several approaches for non-iterative mode-seeking. However, the existing non-iterative mode-seeking methods do not improve the homogeneity of intra-segment pixels, which degrades segmentation accuracy. Furthermore, the existing methods considers the range spectrum of much confined area compared to the conventional mean-shift-based segmentation. As a result, the existing non-iterative mode-seeking methods have difficulty handling the arbitrary and various range spectra of natural images. This dissertation thereby presents a novel non-iterative mode-seeking approach for image segmentation, which improves homogeneity of intra-segment pixels and effectively considers the range spectra.
As the conventional mean-shift-based segmentation does, the proposed method computes the weighted means. However, unlike the conventional mean-shift-based segmentation, the proposed method computes the weighted mean only once for each pixel. There are four pixels which encloses each weighted mean in spatial domain. Then, each pixel is connected to the pixel having maximal density estimate among the four pixels. As a result, pixels are connected to each other by their mean-shift vectors. Finally, regions are formed by grouping the connected pixels. The proposed method not only generates segmented regions through pixel-to-pixel connection, but also can improve the homogeneity of intra-region pixels by replacing the range components of each pixel to those of its weighted mean.
The proposed method however does not consider the range spectra of images, and thus is hard to handle the various and arbitrary range spectra of natural images like the existing non-iterative mode-seeking methods. In order to handle the range spectra of images, the proposed method analyzes the range spectra of images, and then modifies the range domain to be anisotropic; the range domain is expanded in the direction of narrow range spectra, while is compressed in the direction of wide range spectra.
Whereas the mean-shift vector has sub-pixel accuracy, the proposed method connects pixel-to-pixel. Therefore, the pixel-to-pixel connection is considered as the approximation of the mean-shift vector. Because of the loss of accuracy, the proposed method cannot sufficiently aggregate pixels. In order words, the proposed method may not aggregate the pixels which are aggregated by the conventional mean-shift-based method. That causes the over-segmentation problem. In order to alleviate the over-segmentation problem, the greedy merging process is introduced in this dissertation. The greedy merging process analyzes the attribution of each region, and then the regions with similar attributes are merged with the greedy threshold compared to that of the regions with dissimilar attributes.
In experiments, the performance of the proposed method was evaluated in terms of segmentation accuracy and processing speed on Berkeley segmentation dataset. The segmentation accuracy of the proposed method outperformed those of the existing non-iterative mode-seeking-based segmentation methods, Medoid-shift and Quick-shift. Furthermore, by removing the iterative process, the processing speed of the proposed method was accelerated 14.103 times on CPU and 18.040 times on GPU in average when compared to the conventional mean-shift-based segmentation. Moreover, the proposed method demonstrated better segmentation accuracy though modifying the range domain and applying the greedy merging process using region attributes. In comparison to the representative and state-of-the-art benchmark methods, the proposed method demonstrated superior segmentation accuracy. Only the hierarchical segmentation methods using spectral clustering showed better accuracy than the proposed method, but the proposed method provided 52.301 times faster processing speed.
분석정보
서지정보 내보내기(Export)
닫기소장기관 정보
닫기권호소장정보
닫기오류접수
닫기오류 접수 확인
닫기음성서비스 신청
닫기음성서비스 신청 확인
닫기이용약관
닫기학술연구정보서비스 이용약관 (2017년 1월 1일 ~ 현재 적용)
학술연구정보서비스(이하 RISS)는 정보주체의 자유와 권리 보호를 위해 「개인정보 보호법」 및 관계 법령이 정한 바를 준수하여, 적법하게 개인정보를 처리하고 안전하게 관리하고 있습니다. 이에 「개인정보 보호법」 제30조에 따라 정보주체에게 개인정보 처리에 관한 절차 및 기준을 안내하고, 이와 관련한 고충을 신속하고 원활하게 처리할 수 있도록 하기 위하여 다음과 같이 개인정보 처리방침을 수립·공개합니다.
주요 개인정보 처리 표시(라벨링)
목 차
3년
또는 회원탈퇴시까지5년
(「전자상거래 등에서의 소비자보호에 관한3년
(「전자상거래 등에서의 소비자보호에 관한2년
이상(개인정보보호위원회 : 개인정보의 안전성 확보조치 기준)개인정보파일의 명칭 | 운영근거 / 처리목적 | 개인정보파일에 기록되는 개인정보의 항목 | 보유기간 | |
---|---|---|---|---|
학술연구정보서비스 이용자 가입정보 파일 | 한국교육학술정보원법 | 필수 | ID, 비밀번호, 성명, 생년월일, 신분(직업구분), 이메일, 소속분야, 웹진메일 수신동의 여부 | 3년 또는 탈퇴시 |
선택 | 소속기관명, 소속도서관명, 학과/부서명, 학번/직원번호, 휴대전화, 주소 |
구분 | 담당자 | 연락처 |
---|---|---|
KERIS 개인정보 보호책임자 | 정보보호본부 김태우 | - 이메일 : lsy@keris.or.kr - 전화번호 : 053-714-0439 - 팩스번호 : 053-714-0195 |
KERIS 개인정보 보호담당자 | 개인정보보호부 이상엽 | |
RISS 개인정보 보호책임자 | 대학학술본부 장금연 | - 이메일 : giltizen@keris.or.kr - 전화번호 : 053-714-0149 - 팩스번호 : 053-714-0194 |
RISS 개인정보 보호담당자 | 학술진흥부 길원진 |
자동로그아웃 안내
닫기인증오류 안내
닫기귀하께서는 휴면계정 전환 후 1년동안 회원정보 수집 및 이용에 대한
재동의를 하지 않으신 관계로 개인정보가 삭제되었습니다.
(참조 : RISS 이용약관 및 개인정보처리방침)
신규회원으로 가입하여 이용 부탁 드리며, 추가 문의는 고객센터로 연락 바랍니다.
- 기존 아이디 재사용 불가
휴면계정 안내
RISS는 [표준개인정보 보호지침]에 따라 2년을 주기로 개인정보 수집·이용에 관하여 (재)동의를 받고 있으며, (재)동의를 하지 않을 경우, 휴면계정으로 전환됩니다.
(※ 휴면계정은 원문이용 및 복사/대출 서비스를 이용할 수 없습니다.)
휴면계정으로 전환된 후 1년간 회원정보 수집·이용에 대한 재동의를 하지 않을 경우, RISS에서 자동탈퇴 및 개인정보가 삭제처리 됩니다.
고객센터 1599-3122
ARS번호+1번(회원가입 및 정보수정)