Angular Radial Transform을 이용한 새로운 모양 서술 방법

Abstract

In this dissertation, a new shape descriptor that utilizes both phase and magnitude information of angular radial transform (ART) coefficients is proposed. It was proven that ART has the more desirable property than Zernike moments (ZMs) for shape description in MPEG-7 core experiments. The proposed method called ART Descriptor with Phase and Magnitude (ARTPM) employs both phase and magnitude information of ART coefficients to calculate the similarity distance between two shapes. In addition, it provides the rotation angle between two shapes under comparison. The experiment of this dissertation consists of two parts: the performance evaluation the proposed descriptor in terms of rotation angle estimation and image retrieval. In the first part, comparison results of rotation angle estimation of the proposed method with three ZMs based methods on an experimental dataset in terms of root mean square error are presented. In the second part, experiments are carried out to compare image retrieval performance of the proposed ARTPM to invariant Zernike descriptor, traditional ART descriptor and invariant ART descriptor under various image transformations on different datasets derived primarily from the standard MPEG7 CE-1 shape dataset. . Through the experiment, the proposed method shows the best performance with the smallest average error (ε=1.232°) but the computational time is close to the state of the art for rotation angle estimation. The proposed method also outperforms existing methods on image retrieval comparison: the average BEP of the proposed method is 66.01%, while the average BEPs of the IZMD, invariant IARTD descriptor and the traditional ART descriptor are 41.65%, 57.69% and 36.51%,, respectively. The major contribution of this dissertation is that a new shape description method which utilizes both phase and magnitude information of ART coefficients was proposed. Efficient shape description methods are critical for effective image retrieval. The proposed shape description method not only provides better image retrieval performance than the previous methods but also invariant to rotation due to its rotation angle estimation method.|영상에 포함된 객체의 형태는 내용 기반 영상검색에 중요한 정보를 가지고 있기 때문에, 이를 이용하여 영상을 검색하는 방법들이 활발히 연구되어 왔다. 그중에서도 최근에 제안된 저니키 모멘트의 위상과 크기를 이용하는 회전불변 서술자(IZMD: Invariant Zernike moment descriptor)을 이용한 영상 검색 방법과 ART 계수의 위상과 크기를 이용한 회전불변 서술자 (IARTD: Invariant ART descriptor)은 기존의 영역 기반 형태 기술 방법들 중에서 우수한 성능을 보이고 있다. 본 논문에서는 기본의 방법들 보다 향상된 영상 검색 성능을 얻기 위해서 ART(Angular Radial Transform)의 크기와 위상을 이용해서 영상을 정합하는 방법을 제안한다. ART는 MPEG-7 표준화 과정에서 ZM 보다 객체의 형태를 서술하는데 우수한 성능을 보이는 것이 증명 되었으며, 신호처리의 관점에서 위상 정보가 신호의 특성을 표현하는데 크기 정보보다 효과적이다. 따라서, 제안하는 방법은 ART의 위상과 크기 정보를 이용하여 두 객체의 형태의 유사도를 계산하며, 이 과정에서 두 객체 사이의 회전각도를 추정 할 수 있다. 이 논문의 실험은 크게 두 가지로 나누어서 이루어졌다. 하나는 제안하는 방법을 이용하여 두 객체 사이의 회전각도를 추정하는 성능을 기존의 방법들과 비교한 것이고, 다른 하나는 제안하는 방법의 영상 검색 성능을 기존의 영역 기반 형태 기술 방법들의 성능과 비교하는 것이다. 성능 비교를 위한 실험은 MPEG-7 CE-1의 형태 데이터 셋과 이에 다양한 잡음을 이용하여 생성한 데이터셋을 이용하여 정량적인 비교분석을 수행하였다. 실험결과, 제안하는 방법을 이용한 회전각도 추정한 결과는 기존의 회전각도 추정방법들에 의해서 얻어진 결과보다 높은 정확도를 제공하였다. 또한, 제안한 방법을 이용한 영상 검색결과 또한 기존의 저니키 모멘트나 ART를 이용한 방법들보다 유사한 이미지의 검색 성능이 증가되는 것을 알 수 있었다.; In this dissertation, a new shape descriptor that utilizes both phase and magnitude information of angular radial transform (ART) coefficients is proposed. It was proven that ART has the more desirable property than Zernike moments (ZMs) for shape description in MPEG-7 core experiments. The proposed method called ART Descriptor with Phase and Magnitude (ARTPM) employs both phase and magnitude information of ART coefficients to calculate the similarity distance between two shapes. In addition, it provides the rotation angle between two shapes under comparison. The experiment of this dissertation consists of two parts: the performance evaluation the proposed descriptor in terms of rotation angle estimation and image retrieval. In the first part, comparison results of rotation angle estimation of the proposed method with three ZMs based methods on an experimental dataset in terms of root mean square error are presented. In the second part, experiments are carried out to compare image retrieval performance of the proposed ARTPM to invariant Zernike descriptor, traditional ART descriptor and invariant ART descriptor under various image transformations on different datasets derived primarily from the standard MPEG7 CE-1 shape dataset. . Through the experiment, the proposed method shows the best performance with the smallest average error (ε=1.232°) but the computational time is close to the state of the art for rotation angle estimation. The proposed method also outperforms existing methods on image retrieval comparison: the average BEP of the proposed method is 66.01%, while the average BEPs of the IZMD, invariant IARTD descriptor and the traditional ART descriptor are 41.65%, 57.69% and 36.51%,, respectively. The major contribution of this dissertation is that a new shape description method which utilizes both phase and magnitude information of ART coefficients was proposed. Efficient shape description methods are critical for effective image retrieval. The proposed shape description method not only provides better image retrieval performance than the previous methods but also invariant to rotation due to its rotation angle estimation method.