图像分割和目标提取是图像处理领域的基本问题，分割结果对图像配准、修补、特征提取等后续处理产生重要的影响。本文从水平集方法的基本思想出发，综述了几种典型的水平集动态轮廓模型，它们涉及图像的边缘、区域、纹理和先验形状。然后提出了含边缘信息的纹理分割和先验形状分割模型，并应用于病原菌的提取。与原有模型相比，该方法的能量泛函改造和增加了关于图像梯度的正则约束，加强了目标边缘对分割轮廓的吸引，同时消除了一些由噪声、阴影和杂质造成的影响。贝叶斯框架将水平集动态轮廓模型的能量泛函统一起来，为图像各种信息的融合提供了可能。本文关注贝叶斯框架下划分约束的选择，考虑图像梯度的影响，而已有的方法大多只关心划分轮廓长度最短。还探讨了纹理图像分割对象的选取及其各通道关系的假设。这些能量泛函运用变分方法得到一类偏微分方程组，可用AOS求解，文中给出详细的求解过程及其误差精度。实验表明改进后的纹理分割模型能直接处理大量单个病原菌的提取，得到比原有模型更好的分割效果。先验形状分割模型通过相同的改变，可以解决一些病原菌的粘连问题。

Segmentation and selection of regions of interest inside an image is one of the most important steps for image processing, the results of which exert a significant influence on such subsequent tasks as image registration, surface inpainting and feature extraction. In this thesis, based on level set methods, some typical active contour models for level sets are reviewed, involving the edge, region, texture and shape prior of an image. Mathematical models on texture and shape prior with the edge information are proposed and applied in the segmentation of bacteria images. Compared with existing models, the energy functionals of these methods add a modified regularized term with respect to image gradient, which increases the attraction of object edges to the active contour and eliminates the effect led by noises, shadow, and impurities of images. These active contour models for level sets are generalized by Bayesian formulation, which make it possible to integrate diverse features of images. We focus on the choice of the partition constraint in Bayesian framework, assuming it is related with image gradient, while most of other methods favor image partition with the shortest interface. The input for the segmentation of images with texture and the relation of its channels are considered in this thesis. The solution and its error accuracy of a class of Partial Differential Equations resulting from variational techniques, which are solved by AOS, are illustrated in details. Experimental results show that the improved segmentation model based on texture features directly deals with numbers of images containing a single bacterium, obtaining better detection results. With the same modification, the model based on shape prior are able to tackle the adhesion between bacteria.