Abstract
摘要 影像套合通常可分成四個步驟：選取特徵點、特徵點配對、估 測轉移函數參數、套合影像。而其中最困難的部份就是如何有效的選取特 徵點並將特徵點正確地配對。 我們選取在影像上灰階值較高的區塊之 中心點作為特徵點，利用主值分析的方法(PCA)，去分析這些特徵點的分 佈，求出一組新的基底，將兩張影像分別"轉正"後，再進行配對。兩張影 像上的比例大小、旋轉、平移這些變異都可經由轉正的過程移除，然而兩 張影像間如果有多餘或是消失的特徵點，我們轉正的結果就會有一些誤差 ，所以我們提出了一些方法將這些多餘的特徵點儘可能的消除，最後我們 也提出了一個角特徵的分析法來對配對點作進一步的確認；確認所有的配 對關係之後，使用最小平方差估計法(least squares estimation)來作轉 換函數的估測，並利用所得的轉換函數將影像作套合的運算。在論文中我 們也對模擬資料和實際衛星影像作一些實驗，配對的結果相當的良好。 Abstract Image registration plays an important role in the fields such as change detection, establishing the geographical information system, medical image and military detection. In general, the approach to image registration consists of four steps: feature-points selection, feature-points pairing, mapping function estimation, and image registration. The most difficult parts are to select and pair feature-points in different images. We extract the centroids of the most conspicuous parts in each image as feature-points. The method we adopt is to analyze the distribution of all feature-points in each image to get the principle axes by Principle Components Analysis. Then we normalize the distribution of all the points and obtain their normalized coordinates according to the new axes in each image. Scale, rotation and translation variations between two images can be removed by the normalization process. The influence of spurious and missing points can be alleviated by the outliers discard strategy. In this thesis, experimental results of simulation data and satellite images are presented. It is shown that the satellite images are overlaid well by the method we adopt even though the overlapping regions are not very large.