학술논문
유역변환기반의 마커 클러스터링을 이용한 질감 영상 분할 / Textured Image Segmentation Using Marker Clustering in Watershed Transform
Document Type
Dissertation/ Thesis
Author
Source
Subject
Language
Korean
Abstract
영상 분할(image segmentation)은 영상을 명확한 속성을 가지는 다른 영역들로 분리하는 작업으로 여러 응용분야에서 영상 분석의 기반이 되는 주요 기술이다. 응용의 수준이 높아질수록 복잡한 질감 영역을 분할할 수 있는 기법이 요구되며, ‘질감 측정-분할기법’ 연계의 형태로 구성된다. 대부분의 질감 측정은 다양한 특성 영상(feature images)을 가지기 때문에 영상 분할을 완성하기 위해 이 특성들을 어떻게 통합하느냐가 주요 이슈가 된다. 그러한 점에서 클러스터링 기법은 그 특성들에 대한 복잡한 처리 없이 우수한 분할 성능을 제공한다. 하지만 클러스터링 기법은 높은 계산 복잡도와 특유의 분할 오류를 가지므로 응용이 제한된다. 본 연구에서는 클러스터링의 계산 복잡도를 낮추기 위해 마커 클러스터링을 고려한다. 마커-제어 유역변환(marker-controlled watershed transform)에서 마커는 관심 객체에 대한 마킹이자 영역이 확장되는 시작영역으로, 마커에 해당하는 화소만을 클러스터링에 적용하여 복잡도를 줄일 수 있다. 또 마커 클러스터링은 유역변환에 의해 확장되어 완성되므로 클러스터링 특유의 분할 오류도 감소시킬 수 있다. 제안한 기법에서는 가보 필터(Gabor filter)의 질감 에너지에서 마커를 얻고, 마커 화소만을 FCM(fuzzy C-means) 클러스터링에 적용하며, 유역변환 과정에서 마커들이 생성한 영역들을 마커 간 클러스터 소속정도에 따라 병합하여 질감을 분할한다. 제안한 기법은 Brodatz의 영상 조합에 대한 실험에서 클러스터링 특유의 분할 오류를 개선하였으며, 영상 분할 시간 비교에서 기존의 FCM 클러스터링 연계보다 영상의 복잡도가 높을수록, 빠르고 변화폭이 작은 분할 시간을 보여주었다.
Image segmentation which partitions an image into different regions each having certain properties is a principal technology for image analysis in various application fields. The higher level application demands that it partitions more complex textured regions. Generally, the solutions consist of ‘Texture measurement-Segmentation method’ and the main issue is how to integrate texture having multi-feature. In the point of view, clustering supplies superior segmentation performance without complex processing. But the application is restricted because of its higher computational complexity and typical errors from partition. In this thesis, marker clustering for minimizing them is considered. In marker-controlled watershed transform, it regards a marker as a marking about interest object and a start region for extending single region. Hence it can minimize the complexity with the clustering to only restricted pixels called marker. And it can reduce typical errors from partition because the marker clustering is completed by watershed transform. In proposed method, the markers are taken from Gabor texture energies and are applied to FCM(fuzzy C-means) clustering. The segmentation is completed by merging partitioned regions which were generated from all markers in watershed transform according to cluster membership between markers. In the experimental result with mosaic images mixture from Brodatz' album, it improved typical partition errors and shows faster speed than previous FCM clustering chains in the comparison of performance time.
Image segmentation which partitions an image into different regions each having certain properties is a principal technology for image analysis in various application fields. The higher level application demands that it partitions more complex textured regions. Generally, the solutions consist of ‘Texture measurement-Segmentation method’ and the main issue is how to integrate texture having multi-feature. In the point of view, clustering supplies superior segmentation performance without complex processing. But the application is restricted because of its higher computational complexity and typical errors from partition. In this thesis, marker clustering for minimizing them is considered. In marker-controlled watershed transform, it regards a marker as a marking about interest object and a start region for extending single region. Hence it can minimize the complexity with the clustering to only restricted pixels called marker. And it can reduce typical errors from partition because the marker clustering is completed by watershed transform. In proposed method, the markers are taken from Gabor texture energies and are applied to FCM(fuzzy C-means) clustering. The segmentation is completed by merging partitioned regions which were generated from all markers in watershed transform according to cluster membership between markers. In the experimental result with mosaic images mixture from Brodatz' album, it improved typical partition errors and shows faster speed than previous FCM clustering chains in the comparison of performance time.