반복패턴내의 특징점 분류 및 정합

Abstract

현재 휴대용 기기에서 사용하는 대부분의 건물 증강현실응용들은 GPS정보에 의존하여 단순히 카메라를 통해 들어온 영상의 대략적인 위치에 증강을 하기 때문에 위치 정확도와 현실감이 떨어진다. 이를 해결하기 위해 영상처리를 이용해 건물을 인식하며 기존의 건물이나 물체를 인식하기 위해서는 물체의 외관적인 특징들을 구분 짓고 구분된 특징들을 구별할 수 있어야한다. 그 중에서도 다양한 영상 변환에 강인하고 물체의 특징들을 기술하는 SIFT(Scale Invariant Feature Transform)나 SURF(Speeded Up Robust Features)등의 기술들이 널리 사용되고 있으나 특징기반 기술들이 지역적인 영역에 대해서만 특징을 기술하기 때문에 반복적인 패턴이 존재하는 물체의 경우 인식 성능이 급격하게 떨어진다. 특히 건물의 똑같은 창문이 여러 개인 경우 기술된 특징들의 유일성을 잃기 때문에 인식하기가 어렵다. 본 논문에서는 반복적인 패턴을 갖는 영상에서 나타나는 문제를 해결하기 위하여 기존의 SURF 기술자를 보완한 기술을 제안한다. 반복패턴에서 오정합되는 SURF 기술자들을 Mean Shift Clustering 방법을 이용하여 분류해낸다. 반복패턴의 경우 SURF의 64차원 기술자가 유사한 특징을 갖기 때문에, SURF 기술자들에 대하여 Mean Shift Clustering을 이용하여 유사한 특징을 갖는 기술자들을 분류할 수 있다. SURF 기술자들의 반복패턴이 분류되면 기술자들이 수렴한 모드들의 유사성을 평가하기 위해 거리측정함수를 이용한다. 그 후에 반복패턴으로 오정합되는 패턴들을 제외하고 남아있는 기술자들에 RANSAC(RANdom SAmple Consensus)을 적용하여 신뢰성 있는 기술자들을 선출한다. 선출된 기술자들을 이용하여 두 영상간의 호모그래피 변환을 계산한다. 데이터베이스(Database) 영상을 호모그래피 변환하면 질의(Query) 영상에서 검출된 기술자들과 실제좌표를 서로 비교할 수 있다. 따라서 기존에 오정합되던 문제를 수정할 수 있고, 정확한 대응점을 이용하여 증강현실에 응용할 수 있다. 또한, 다양한 건물과 반복질감 영상에 대하여 실험하고 기존 SURF와의 성능을 비교한다.|As wireless-enabled handhelds such as cellphones and cameras have been developed rapidly, various augmented reality (AR) applications are being created. But most building AR technologies so far, are less location accurate and realistic because these are augmented on the coarse location of the image from a camera depending on GPS data. To overcome this problem, I recognize buildings using the image processing. Feature-based technologies such as SIFT and SURF are wildly used, which are robust enough to diverse image transformation. It doesn't perform well in the case that there are repeated patterns because feature-based technologies describe only for a local area. Especially, if there are same windows a lot in a building, it is hard to recognize the feature of the building due to losing its uniqueness. In this thesis, I suggest the enhanced SURF descriptor to solve the problem on the images that have repeated patterns. Mismatched SURF descriptors from the repeated patterns are classified by using Mean Shift Clustering. As for the repeated patterns, 64-dimensional SURF descriptors have a similar feature. Therefore, the descriptors having a similar pattern can be classified. Once the repeated pattern of SURF descriptors are classified, distance function is used to evaluate the similarity between the converged modes. Then, RANSAC(RANdom SAmple Consensus) is applied to the remaining descriptors except for mismatched patterns and reliable ones are selected. I also calculate the homography between two images using the selected descriptors. By using homography to transform a database image, I can compare the real coordinates of the descriptors extracted from query images with those from the database image. This overcomes the problem of the mismatching problem and it can be applied to AR with the exact correspondence points. I test various buildings and repeated patterns in textiles and compare the performance of the existing SURF.; As wireless-enabled handhelds such as cellphones and cameras have been developed rapidly, various augmented reality (AR) applications are being created. But most building AR technologies so far, are less location accurate and realistic because these are augmented on the coarse location of the image from a camera depending on GPS data. To overcome this problem, I recognize buildings using the image processing. Feature-based technologies such as SIFT and SURF are wildly used, which are robust enough to diverse image transformation. It doesn't perform well in the case that there are repeated patterns because feature-based technologies describe only for a local area. Especially, if there are same windows a lot in a building, it is hard to recognize the feature of the building due to losing its uniqueness. In this thesis, I suggest the enhanced SURF descriptor to solve the problem on the images that have repeated patterns. Mismatched SURF descriptors from the repeated patterns are classified by using Mean Shift Clustering. As for the repeated patterns, 64-dimensional SURF descriptors have a similar feature. Therefore, the descriptors having a similar pattern can be classified. Once the repeated pattern of SURF descriptors are classified, distance function is used to evaluate the similarity between the converged modes. Then, RANSAC(RANdom SAmple Consensus) is applied to the remaining descriptors except for mismatched patterns and reliable ones are selected. I also calculate the homography between two images using the selected descriptors. By using homography to transform a database image, I can compare the real coordinates of the descriptors extracted from query images with those from the database image. This overcomes the problem of the mismatching problem and it can be applied to AR with the exact correspondence points. I test various buildings and repeated patterns in textiles and compare the performance of the existing SURF.