• 作者： 王獻章; 傅楸善
• 作者服務機構： 國立臺灣大學資訊工程研究所
• 中文摘要： 在電腦影像處理中，如何辨識影像中的物體及如何描述物體，是一項非常重要的課題。近二十年來，“偵測支配點”在處理這個問題上已成為主要的步驟之一。如何精確地找到我們需要的支配點，正是本論文研究的主題。 在本論文中，我們提出一個新的方法偵測支配點，稱之為：RCV-Cosine方法，其中RCV為Rated Composite Vector（比例合成向量）的縮寫，意即利用一個演算法，針對在2D影像中物體邊界上的每一點，求出左右比例合成向量，藉以餘弦夾角公式，求出近似的曲率值。準確地估計曲率值是找出支配點重要的第一步，也是我們提出RCV-Cosine方法所追求的目標。 支配點的形成，即由那些具區域性最大曲率值的點獲得。我們將討論由我們提出的演算法求出的支配點將適合做圖形辨識，並且它適合在微量的放大縮小及旋轉下辨識物體。
• 英文摘要： Dominant point detection is an effectual step in pattern recognition, feature description, and shapereconstruction. In this paper, we propose a novel measurement of curvature to detect the dominant pointsalong the digital planar closed curve. We call it the rated composite vectors for cosine (RCV-cosine)measurement because of an analogy to the k-cosine measurement advocated by Rosenfeld and Johnstonfor describing angle detection. We define the RCV as a composite vector obtained by iteratively andadaptively rating the vectors described in our algorithm. We estimate an“angle”with RCV replacinga fixed window size k for cosine measurement. For each point in the curve, we find both the left andthe right RCVs that are similar to the left and the right tangents, and thus calculate the angle, called theRCV-cosine value, between two RCVs through cosine measurement. The dominant points are obtainedby the local maximum of the RCV-cosine values. The proposed algorithm is robust on raw synthetic imagesand extracts the dominant points effectively and correctly. We present several experimental results andshow that our approach is invariant to translation, rotation and slight scaling, and that it is closer to humanperception. Furthermore, the approximation errors of shape reconstruction are discussed.
• 中文關鍵字： dominant point detection; rated composite vectors; k-cosine; kl-cosine; curvature; corner detection; pattern recognition; shape reconstruction
• 英文關鍵字： --