Richards W, Dawson B, Whittington D
J Opt Soc Am A. 1986 Sep;3(9):1483-91. doi: 10.1364/josaa.3.001483.
Curvature extrema provide significant information about the shape of an image contour, such as a silhouette, and are the basis for the Hoffman-Richards codon representation for shape. This representation based on curvature easily translates into a binary string that will describe the abstract shape of any smooth image curve. The computation of the basic shape primitives requires dealing with two ever-pervasive problems: contour noise and scale. We show how contour noise can be estimated given knowledge of the shape of the filter used to compute curvature from the edge list of the contour. To handle the scale problem, we use an adaptation of Witkin's scale space. Our algorithm differs from Witkin's by using a notion of parts to set criteria for significant structures.
曲率极值提供了有关图像轮廓（如图像剪影）形状的重要信息，并且是形状的霍夫曼 - 理查兹编码表示的基础。这种基于曲率的表示很容易转化为一个二进制字符串，该字符串将描述任何平滑图像曲线的抽象形状。基本形状基元的计算需要处理两个普遍存在的问题：轮廓噪声和尺度。我们展示了如何在已知用于从轮廓的边缘列表计算曲率的滤波器形状的情况下估计轮廓噪声。为了处理尺度问题，我们采用了威特金尺度空间的一种变体。我们的算法与威特金的算法不同，我们使用部分的概念来为重要结构设定标准。
