简易解释：在曲面中插入一个平面所形成的轮廓，即是该轮廓的水平集表示，可见，该轮廓的水平集表示有多个。对于图像分割，在图像力的驱动下曲面进行更新。
轮廓的数学表达有隐式和显式两种表达。用曲面演化代替Front （C）演进。
C(t) = {(x, y)|φ(x, y, t) = 0}
∂φ/ ∂t + F|∇φ| =0 (1)
φ(x, y, 0) = φ0(x, y) 方程的本质是什么？ 几何解释是什么
edge-based level set
∂φ /∂t = g(∇I)|∇φ| (div ( ∇φ /|∇φ| _ + ν ) (2)
φ(x, y, 0) = φ0(x, y) based on mean curvature motion by Caselles et 1993
∂φ /∂t = |∇φ| (−ν + ν/ (M1 − M2) )(|∇G∗I| − M2) (3)
φ(x, y, 0) = φ0(x, y) Malladi et 1993
∂φ /∂t = |∇φ|div （g(I) ∇φ /|∇φ| _ + νg(|∇I|) = g(I)|∇φ|div ( ∇φ /|∇φ| _+∇g(I) · ∇φ + νg(|∇I|)
φ(x, y, 0) = φ0(x, y) (4) Kichenassamy et 1995 and Caselles et 1995
An abstract representation common to all edge-based partial differential equation(PDE) is as follows:
∂φ /∂t = −αA(x) · ∇φ − βP(x)|∇φ| + γZ(x)κ|∇φ| (5)
φ(x, y, 0) = φ0(x, y) A is an advection term ; P is a propagation(expansion) term ; Z is a spatial modifier term for the mean curvature k.
region-based level-sets
Region-based level-sets segment the image into objects based on region statistics (rather than just object edges) of intensity, tex- ture, or color values.
F(c1, c2, φ) =
Inside(C)
(I(x) − c1)2dx +
Outside(C)
(I(x) − c2)2dx + ν · Area(C) +μ · Length(C) (6)
by Chan and Vese 2001
3.1 DOMAIN REPRESENTATION …… 不懂
3.2 水平集函数
定义了一个抽象的水平集函数基类 itk：：LevelSetBase
所有的水平集函数类实现具体的成员函数返回the level-set value [φ(x,y)], gradient(∇φ),Hessian(∇ 2 φ),Laplacian(φ xx + φ yy ),gradient norm(|∇φ|), and mean curvature (κ = div(∇φ/|∇φ| )) given its underlying representation (continuous or discrete image or mesh). Thus,the level-set equation, term, and evolution classes are independent of the underlying domain representation which facilitates the implementation of a wide variety of level-set methods.
图像的离散化表示itk：：DiscreteLevlSetImage 被具体实现为Dense 和Sparse 情况。三种Sparse的表示：Whitaker 、Shif、Malcolm。（narrow-band）
3.2.1图像到水平集转换 BinaryImageToLevelSetAdaptorbase
3.3
RESTRICTED LEVEL-SET DOMAINS 限制水平集域
在图像子集域内进行水平集演化，划分成不同子域。 A helper base class (LevelSetDomainPartitionBase) is used to define the location and size of the level-set domains relative to Ὼ (Figure3A).
Each grid point stores a list of the active level-set function ids. For the case when there are thousands of level-sets, populating a list image by checking overlap at each pixel is time-consuming. Therefore, we further specialized into a class itk::LevelSetDomainPartitionImageWithKdTree. This class uses a Kd-tree data structure that contains the centroids of the level-set domains. The Kd- tree is used to query nearby level-set functions at each pixel and check for overlap.This enables the simultaneous evolution of thousands of level-set functions thereby expanding the applicability of level-set procedures to tracking large numbers of objects and in large images. Note that there is an initial overhead associated with building the Kd-tree that can be avoided for cases involving a small number of level-set functions.
3.4 terms
水平集方程是各项的加权和。The term base class implements functions [Evaluate(.)] for computing the contribution from a term toward the level-set update.
3.5 container-based design
Container 是什么？
Different types of terms arising from edge-based and region-based level-set methods such as the propaga-
tion, Laplacian, advection, curvature, and region-based terms described in Equations 4 and 7 derive directly from LevelSetEquationTermBase:
we used containers to store level-set function objects, equation objects, and their constitutive terms。
Leve-set container term containers level-set equation container
3.6 Level-set evolution
3.7 stopping criterion
itk::StoppingCriterionBase
3.8 user-interaction
3.9 visualization