提升特徵點配對精準度於衛星影像套合

Improving Feature Point Matching Accuracy for Satellite Image Registration

詹宛真(Wan-Chen Chan)

指導教授 ： 張恆華

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201903946

摘要

影像套合是將兩個不同來源的相同物件，透過影像轉換的方法對齊，而衛星影像擁有不同的影像型態和複雜的影像場景，以便於各種量測觀察，故衛星影像的自動影像套合方法仍然是一個具有挑戰性的研究。本論文提出一個快速與準確的方法來實現自動衛星影像套合，提出的方法分成三大步驟：第一步驟是使用已改良的尺度不變特徵轉換方法，其使用雙線性內插法的重採樣以解決原始重採樣因旋轉而特徵點位移的問題；然後將特徵描述值正規化使兩組特徵點配對更精確與健全。第二步驟是進行特徵點配對的正確與錯誤判斷，其使用隨機抽樣組合演算法對樣本特徵點進行篩檢。一方面，在達到基本抽樣次數後，當檢覈成功次數大於當前抽樣次數的一半，則歸類為正確配對，提早抽離以加快速度；另一方面，若是挑選出的正確配對數目不足，則會以目前正確配對為標準數值，重複上述。第三步驟建立一個能以正確配對特徵點自動變化的範圍，將其均勻分類並篩選合適特徵點，最後將最合適的組合進行仿射轉換，找到最適合的轉換參數。我們將此方法與當今衛星影像套合的方法進行比較，實驗結果顯示本方法可以得到精確且節省時間的效果。該策略可應用在多種衛星影像，例如：城市、河流、山谷、海岸，觀察城市發展、河流侵蝕變化和海平面上升。

關鍵字

遙測影像 ； 影像套合 ； 尺度不變特徵轉換 ； 仿射轉換

並列摘要

Image registration is defined as the same object of two different sources alignment by some kind of image transformation. Since satellite imagery has different image types and complex image scenes for a variety of measurement observation, developing automatic image registration methods for satellite imagery is still a big challenge. This thesis proposes a fast and accurate method to achieve automatic satellite image registration, which is composed of three major steps：the first step is to use the improved scale-invariant feature transform(SIFT)method, in which bilinear interpolation resampling is adopted to solve the original resampling problem due to rotation and displacement. Then, we normalize the feature description values to make two feature point sets more precise and robust. The second step is to remove the misjudgment of the feature point pairs. We use a random sampling combination algorithm to sample feature points. On the one hand, when the number of successful inspection is greater than half of the current sampling number after reaching the basic sampling number. It is classified as correct pairing. Doing this stop the iteration earlier and speed up the whole process. On the other hand, if the number of correct pairs is not enough, we repeat the above step with the current correct pairing as the standard value. The third step is to build a measurement domain that automatically changes with respect to the correct pairing feature points. We make them evenly and filter the appropriate feature points. Finally, we extract the most suitable combination to find the most suitable conversion parameters. By comparing this new approach with other satellite image registration methods, it is suggested that our method is precise and time-saving that can be applied to a variety of satellite imagery, such as cities, rivers, valleys, and coasts. It can be used to observe urban development, changes in river erosion, and rise of the sea level.

並列關鍵字

remote sensing image ； image registration ； SIFT ； affine transformation

參考文獻

[1] H. E. M. Meier, A. Hoglund, K. Eilola, and E. Almroth-Rosell, "Impact of accelerated future global mean sea level rise on hypoxia in the Baltic Sea," (in English), Climate Dynamics, vol. 49, no. 1-2, pp. 163-172, Jul 2017.

[2] J. L. Moigne, N. S. Netanyahu, and R. D. Eastman, Introduction (Image Registration for Remote Sensing). UK:Cambridge University Press, 2011.

[3] D. G. Lowe, "Distinctive image features from scale-invariant keypoints," International journal of computer vision, vol. 60, no. 2, pp. 91-110, 2004.

[4] Q. Li, G. Wang, J. Liu, and S. Chen, "Robust scale-invariant feature matching for remote sensing image registration," IEEE Geoscience and Remote Sensing Letters, vol. 6, no. 2, pp. 287-291, 2009.

[5] B. Li and H. Ye, "RSCJ: Robust sample consensus judging algorithm for remote sensing image registration," IEEE Geoscience and Remote Sensing Letters, vol. 9, no. 4, pp. 574-578, 2012.