使用多重描述符之圖像塊匹配於場景對位

Abstract

在這篇論文中，我們提出一個方法來解決基於影像的場景對位。場景對位的目的在於建立一對影像之間的密集對應，過去已經有許多研究專注於影像對位這個問題，在兩張包含相同場景但是從鄰近的視角或在連續的時間點上所拍攝的影像之間，進行密集的立體對應和光流場的估計。然而，要對兩張包含不同場景或物體的影像做場景對位，仍然是個具有挑戰性的問題，因為在一般影像中存在著多種不同的影像變異，我們並沒有辦法事先知道有哪些變異發生在它們之間，所以在這一個研究中，我們提出利用多重描述符來克服不同影像變異的問題，並且介紹一個準則，可以在 SIFT、geometric blur、DAISY 和 LIOP 之間選出適合的描述符。另外，不同場景間嚴重的影像變異以及影像的高解析度，會造成計算成本變得更高，為了改善在這些情況下的效率，我們採用一個階層式的結構，以由粗糙到精細的方法，先快速地對整張影像估計出大概的對應。我們的方法主要是基於一個現有的技術，即 PatchMatch filter，它是一個通用且快速的架構，目的在於處理一般多重標號的問題，所以我們將自己提出比較不同描述符的準則結合到這個架構中。在最後的實驗中，我們將提出的方法測試在不同且具有挑戰性的資料集合上，結果顯示我們的方法藉由利用不同的描述符所提供的互補資訊，可以適合用在一些普通的影像對上。

In this thesis, we introduce a general method for image-based scene alignment. Scene alignment aims to establish dense correspondence for a pair of images. Much research effort has been made to estimate a dense stereo and optical flow field for two given images that have the same scene but were captured from distinct viewpoints or at different time. However, to align images with different scenes or objects, it is still challenging. There are diverse image variations between general images. It is difficult to know what kind of image variations occurs across the images in advance. We hence propose to utilize multiple descriptors to deal with the problem of different image variations. A criterion is presented to select proper descriptors among SIFT, geometric blur, DAISY, and LIOP. Moreover, serious image variations as well as high image resolutions make the computational cost becomes much higher. To improve efficiency in this circumstances, we adopt a hierarchical structure to estimate the approximate correspondences in coarse-to-fine manner. This work is based on an exiting technique, ie PatchMatch filter, which is a generic and fast computational framework for general multi-labeling problems. We integrate the aforementioned criterion into the framework. Experiments on different challenging datasets show that our approach is suitable for general images by leveraging the complementary information from the different descriptors.