图像分割的定义-湖南大学.ppt

应用 * 算法的不足和改进 因为水平集方法计算量太大，为了提高收敛速度，提出了很多种算法；比如之前提出的窄带模型；我们可以先用某些算法对图像进行粗分割，得到初始轮廓，这样可以大大降低水平集的迭代次数。 针对不好处理的尖角问题，可以对边缘检测函数在尖角顶点附近进行修正，例如用截断函数来选定尖角顶点附近的位置。 把基于边缘的水平集模型和基于区域的水平集模型相结合，利用二者的优点，找出更好的算法。 * * Contour Propagation in Level Set Taking the derivative of the level set function with respect to time (using the chain rule), we will get the evolving expression of the contour as Through numerical methods we can get the solution of the geometric PDE, so we should first build the PDE function that including the level set function, and that is a kind of function that we call the energy function and then minimize it. Energy function in level set Parametric active contour model (Kass 1988): Find C(p) which minimizes: Energies are curve integrals Internal energy External energy Image gradient is max. on C Energy Components Internal energy: constraints the smoothness, elasticity and rigidity of the curve, such as the first order derivative, second order derivative, length and area components. They are the common parts. External energy: attracts the contour towards the object in the image, like the boundary-based energy, including the local information (gradient, i.e. edge) and the region-based energy, including the both the local and global information. Different models employ different external energy part—designable. Region based energy Image divided in 2 regions: I Ωin Ωin Ωout Region: Background C Contour: interface between 2 regions Region: objects to be segmented Region based energy Find the regions Ωin and Ωout that minimize the criterion: Region-based terms I Ωin Ωin Ωout C Region based energy, fitting term, minimal variance Region based energy (Chan-Vese Model) * 一些实验结果 简单例子 /watch?v=hs4VoYE9ZkY /watch?v=2k-h9jntFLw /watch?v=UQZAcijZTq8 * * 算法的不足和改进 水平集算法的核心是解决水平集进化的问题。通过分析比较目前已有的水平集模型，发现主要存在有两大问题： 绝大部分水平集模型均涉及到需要在进化过程中对水平集函数重新初始化的问题； 许多模型都需要借助原始图像的梯度信息来保证水平集进化如期收敛。 文献[4]解决了第一个问题，但没有解决第二个问题。而第二个问题会直接导致算法在处理由梯度定义的弱边缘图像特征时显得无能