基于 SIFT 的圖像匹配及其綜合應用研究

Abstract

圖像匹配是圖像處理的基礎組成部分，其作用是用 來尋找兩個或兩個以上不同條件下獲得的圖像的匹配 點。這些匹配點可以進一步用於一系列應用，例如圖像 配準，圖像檢索等。一般來說，一個匹配算法有三個組 成部分。第一是特征檢測，這個過程用於尋找一些特殊 的點，這些點也被稱為角點，興趣點或特征點。第二是 計算描述所選特征的特征描述符。第三是尋找匹配的方 法，該方法將會尋找在前面的步驟中發現的特征點之間 的匹配。本文將首先介紹一些早於 SIFT 特征檢測的算 法，然後詳細介紹了 SIFT 特征檢測以及一些使得 SIFT 匹配過程更有效的算法，例如基於 K-D 樹的搜索算法和 空間檢測算法。最後基於以上介紹的算法提出了 SIFT 的 應用，包括圖像拼接和圖像分類，在介紹 SIFT 應用的同 時，也介紹了一些其他用到的算法，例如密集 SIFT 特 征，金字塔空間匹配等。

Image matching is one of the basic tasks in image processing. It is used to find matched points in two or more images captured under different conditions. These matches can be further used to lots of applications such as image registration, image searching, image retrieving and so on. Generally, one matching algorithm has three components. One is feature detector which is used to find some points. These points are also called corner points, interesting points or feature points. The second one is feature descriptor which is used to describe the selected feature. The third part is matching metrics which shows how to use the features found in previous steps to match points. In this thesis we will first introduce some algorithms prior to SIFT feature detection algorithm. Then the SIFT feature detector algorithm used in the thesis is introduced in detail. Some matching algorithms is further introduced to make the SIFT based matching process more robust, such as K-d tree based searching algorithm and some spatial verification algorithm. Finally, some applications based on algorithms introduced above are proposed including image stitching and image classification.