自動化影像套合之研究

Abstract

本研究針對影像套合流程中的各個步驟如特徵萃取、影像匹配、整體一致性檢查與系統誤差測試等，分別加以討論。然後結合各個步驟，發展出一套可適用於各式的影像，且以最少人工介入為標的的自動化影像套合方法。 經研究之後，該自動化流程如下：1.以TDGO興趣運算元找出明顯的特徵點位。2.以鬆弛法特徵匹配建立兩張影像的起始對應關係。如果鬆弛法匹配失敗，則改由人工選取兩對共軛點以建立起始對應關係。3.以標準化互相關匹配與最小二乘法匹配，匹配共軛點至次像元精度。4.利用中心透視投影之八參數轉換做為兩張影像之對應模式，並以Tau測試剔除具有較大誤差的共軛點位。5.以Ajne、Moore與Spearman三種系統誤差測試，檢驗數學模式是否適當。控制點之選取，採疊代方式逐漸加密，以找出適當的轉換參數。 實驗部分，分別利用航空立體像對、135相機拍攝掃瞄之近景影像與數位相機拍攝之影像作為實驗數據，進行測試。

The objective of this research is to investigate used procedures for image registration, such as feature extraction, image matching, consistency check and system error test. An automated image registration scheme is then developed. The developed procedures include：1.Use the TDGO interest operator to find the feature point; 2.Utilize the relaxation method to establish the initial correspondence between the two registering images. If the relaxation method fails, manually digitize 2 pair of control points to set the initial correspondence between the two images for registration; 3.Use Normalized Cross-Correlation and Least-Squares Matching to match the image to the subpixel; 4.Utilize projective transformation as the mapping function and use Tau test for gross error detection; 5.Utilize the Ajne, Moore and Spearman statistical tests to examine the suitability of the mapping function. In order to obtain a better mapping function, the selection of matching points can be implemented iteratively. Several image pairs from different sources, including scanned aerial photographs, images taken with a digital camera, are used in the experiments.