尺度空間之螺旋特徵與平面影像對位

Abstract

在電腦視覺的領域中，影像對位 (image registration) 是一個基礎且重要的問題。其衍生出的議題及應用相當廣泛，包括：立體匹配 (stereo matching)、三維結構重建 (3D structure reconstruction)、物體識別 (object recognition)、移動追蹤 (motion tracking) 等。在本論文中，我們著重於平面影像對位 (planar image registration)。利用創新的影像特徵描述 (image descriptor)：螺旋型特徵 (spiral descriptor)，進行影像匹配 (image matching) 以及影像對應 (image correspondences) 的修正。平面影像的校正以及最佳化也在本論文中一並提出。

此影像對位系統主要包含兩項技術，影像對應的偵測或選取以及單應性矩陣 (homography matrix) 的最佳化。對於影像之對應，我們利用可靠的影像對應法，例如：SIFT，自動取得，但其無法適用於所有情況。所以對於較困難的情況，我們使用人工點選對應點的方式，之後再利用所提出的螺旋型特徵進行修正。此螺旋型的特徵點位置位於尺度空間中 (scale space)，此特徵描述由螺旋型之外觀建立，以達到縮放以及位移不變性 (invariant)。使用動態規劃 (dynamic programming) 進行特徵匹配 (descriptor matching) 適合於旋轉不變性。對於平面影像對位，我們提出一個創新的方法以提升精確度。首先，我們利用自動偵測影像對應或人工點選對應點並且使用螺旋型特徵進行自動修正的方法來求得單應性矩陣。此矩陣可分解成其參數，並且利用非線性的最佳化方法進行調整。最後，最佳之單應性矩陣可以提供高精確度的平面影像對位。

我們所提出的螺旋型特徵可以自動且穩定的進行影像匹配。對於較困難的情況，我們使用人工點選對應點並且自動修正的方式以獲得影像對應。此平面影像對位系統不僅提升了對位精確度，並且提供方便使用者進行平面影像對位的方法。

Image registration is a fundamental problem in computer vision, and it also has been used to many research issues including stereo matching, 3D structure reconstruction, object recognition, and motion tracking. In this thesis, we focus on planar image registration. A novel image feature descriptor: spiral descriptor is proposed for image matching and correspondences refinement. The planar image registration and its optimization method is also proposed in this thesis.

The image registration system involves two major techniques, image correspondence detection/selection and homography matrix optimization. For image correspondences, we obtain them automatically using reliable image matching methods like SIFT, but it is impossible for all cases. Therefore, we manually select pairs of corresponding points and refine using proposed spiral descriptor for the hard cases of image matching. The spiral feature points are localized in scale space and the descriptors are built along spiral-shape profile, which can achieve scaling and translation. The dynamic programming technique is used to match spiral descriptors and it is suitable for rotation invariant. For planar image registration, we propose a novel method to promote the registration accuracy. First, we estimate the homography matrix by either detecting the image correspondences automatically or selecting image corresponding points manually and refining using proposed spiral descriptor. The initial homography matrix is decomposed into its parameters and the non-linear optimization process adjusts these parameters using iterative process. Finally, the optimal homography can produce high registration accuracy for planar images.

The proposed spiral descriptor can match images automatically and robustly. For the hard cases of image matching, we select the correspondences manually and refine the positions automatically. The proposed planar image registration system not only promote the registration accuracy but also provides convenience process for user doing image registration.