Publications

Detailed Information

Repetitive Pattern Descriptor for Robust Image Matching and Registration : 강인한 영상 매칭과 정합을 위한 반복 패턴 묘사자

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors

하성종

Advisor
조남익
Major
공과대학 전기·컴퓨터공학부
Issue Date
2012-08
Publisher
서울대학교 대학원
Keywords
repetitive patternfeature matchingobject recognitionimage registrationimage retrieval
Description
학위논문 (박사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2012. 8. 조남익.
Abstract
After extracted features to compare two images, features with similar properties from an image are grouped into the same set. Then, features or groups of another image are examined with representatives or descriptors of the groups. To do this, the number of sets to group has to be decided first. Also, when sets grouped in reference image meet a query feature not related with the sets, its labeling problem occurs. To alleviate these problems, a new repetitive pattern descriptor is proposed in this dissertation. Also, the proposed descriptor is applied to three applications: object recognition, image registration, and image retrieval. An overview of the contributions of this dissertation is as follows.

First, features from repetitive patterns are discriminated from the ones from salient structures and a way of utilizing the descriptor for robust pattern matching is developed. Specifically, we first apply a conventional feature extraction method to a given image. Then, features are clustered based on their similarity, i.e., a classifier that tells whether a feature is from a repetitive pattern or from a salient structure is designed. For the effective use of repetitive patterns, a new descriptor based on support vector data description (SVDD) for describing clusters of similar features is defined. In other words, a set of features from a pattern is defined to be a new feature in terms of its center and radius. Then, a condition for matching of clusters is defined based on the proposed cluster description.

Second, the proposed descriptor overcomes matching failure of object with repetitive patterns. Conventional object retrieval or recognition methods based on feature matching sometimes fail when an object contains repetitive patterns, because features from repetitive patterns are too similar to each other. Specifically, when there arise many similar features in a query object due to repetitive patterns, they are usually not matched to the ones at the same positions of the reference object. In this dissertation, a new feature matching strategy to alleviate this problem is proposed. Based on the proposed description, features are categorized into ones from repetitive patterns and salient ones, and repetitive patterns are described by the proposed descriptors. For object recognition, the homography is found over the salient features by excluding repetitive features at first, which is then validated and refined by the repetitive patterns. The proposed method is tested with examples of matching buildings with repetitive patterns, and it is shown to be robuster and more reliable than the conventional methods.

Third, geometric cue of correspondences within the same pattern is added to find images that are related with a base image. The geometric cue, which is based on cross product of two vectors that consist of a point and its two nearest neighbors, has lower complexity than the conventional approaches because it does not need to calculate the inverse of a matrix. The proposed matching followed by verification by the geometric cue is applied image registration. Experiments are conducted with building and aerial images, it is demonstrated that the proposed method provides more correspondences than the conventional approaches and thus better registration results.

Finally, the proposed descriptor is applied to various features. In matching of object with repetitive patterns, the proposed approach is applied to features such as SIFT/SURF. To show its application to different features, the proposed descriptor for color features is used for image retrieval. The proposed descriptor of similar color components plays a role of pallet, and an image is consisted of some principal (grouped) colors. When the proposed matching for similar features is applied to colors of another image, there are two images with corresponding grouped ones. In order to utilize their spatial distribution together with colors itself, mutual information on the distribution is used. Maximum mutual information between two images is found, and images in database are sorted by their maximum mutual information with respect to query image. Experiments on several color image databases show that the proposed method gives better performance than the conventional methods.
Language
English
URI
https://hdl.handle.net/10371/118852
Files in This Item:
Appears in Collections:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share