【作者】 李启翮；

【导师】 萧德云；

【作者基本信息】 清华大学 ， 控制科学与工程， 2008， 博士

【摘要】 蛇模型在医学图像分割、视频分析等众多领域有着广泛的应用,尽管已有二十年的历史,在应用中仍然存在一些未解决的问题,本文针对蛇模型分割结果对初始曲线位置过于敏感、难以应用于全自动分割领域以及几何式蛇模型计算效率较低等三个问题进行了研究,取得一些具有理论意义和实用价值的成果。1.提出一种对初始化曲线位置不敏感的新型蛇模型外力场自适应压力场。首先针对蛇模型初始化问题的原因进行了深入的研究,提出了可以避免初始化问题的力场所应具备的特征:可以自适应的根据其位置的不同而对蛇模型施加不同的压力,即在目标区域内部表现为膨胀力,外部为收缩力。再根据该特征的指导,提出基于邻域和基于梯度两种构造力场的方法,前者计算复杂度较高,但通过调整所用邻域的尺寸,对分割较细的目标有良好的效果,后者计算速度快,适应范围更广。2.通过将蛇模型外力场与种子区域生长法相结合提出一种高效的全自动分割方法。基于蛇模型的图像分割难以避免人工干预,为此本文对全自动分割进行了研究。根据流场分析理论,向量流场中每个点的性质可以由该点周围向量的方向所反映,通过将这些向量的方向映射为一个标量,得到一个标量场流向标量场,然后以其为辅助图像,用一种改进的种子区域生长法进行初始分割,最后用区域合并得到最终结果。该方法计算速度快,对噪声鲁棒性好,适合快速应用。本文还将其拓展到彩色-纹理图像分割领域。3.提出一种基于双前沿蛇模型的视频目标轮廓跟踪方法。与参数式蛇模型相比,几何式蛇模型有着更好的特性,然而较大的计算负荷使其难以适应目标跟踪快速化的要求。本文对双前沿活动轮廓模型进行了研究,提出支撑区域限制和拟气球模型两项改进以提高其计算效率,得到了一种新模型Dual-frontsnake with quasi-balloon ,并在此基础上提出了一种快速、灵活的轮廓跟踪方法。该方法保持了几何式蛇模型自适应拓扑结构变化的优点,并能用于背景运动的场景,跟踪运动变化剧烈、变形较大的目标。更多还原

【Abstract】 Snake model is widely used in the domain of medical image segmentation, videoanalysis and so on. Though being studied for more than 20 years, there are still someproblems to be solved. This dissertation focused on three of the problems: the segmen-tation result of snake is heavily sensitive to the initial contours, the snake model is notappropriate for automatic segmentation, and the low computation velocity of geometricsnake. And three new methods are proposed.1. A new external force field named adaptive pressure force field which makesthe snake model not sensitive to the position of initial curve is proposed. First, theinitial problem of snake model is studied. Based on the analysis of and comparisonamong the existing external force such as balloon model, gradient vector ?ow and soon, the character of an external force which can avoid the sensitiveness to the positionof initial curve is concluded: the pressure force imposed on the snake model should beable to change adaptively according to its position, namely, it can drive the snake modelto in?ate when the evolution curve locates inside the region of the object and shrinkwhen the curve locates outside. Then, according to the character of adaptive pressureforce field, two method are proposed to construct such force field. One is based on theneighborhood, and the other is based on gradient. The time complexity of the formeris high, but it can get very good result by choosing appropriate size of neighborhood ifthe object to be segmented is thin. The latter can be calculated fast, so it can be widelyused.2. A fast automatic image segmentation method is proposed by the integrationof the external force field of snake model and seeded region growing method. Theexistence of human interaction makes snake model not appropriate for the domain ofautomatic segmentation. So the automatic segmentation method need to be studied.According to the theory of ?ow field analysis, the characteristic of each point in thefield can be re?ected by the directions of the vectors around it. Thus, an assistant image called ?ow direction scalar field is acquired by mapping the directions of thevectors of each point to be a scalar. Then the scalar field is segmented by using animproved seeded region growing method to get the initial segmentation result. And thefinal result is acquired by a region merging step. The proposed method is very e?cient,robust to noise and fit for the fast application. Besides, it is extended to the domain ofcolor-texture image segmentation.3. A video object contour tracking method based on dual-front snake model is pro-posed. Comparing with the parametric snake, geometric snake has better performance.But it is di?cult to be used in the domain of object tracking which need a high pro-cessing speed because of its heavy computation load. Two adapt the original dual-frontactive contour to object contour tracking, two improvements called support region re-striction and quasi-balloon model are made, thus a new model named dual-front snakewith quasi-balloon is acquired. Then a fast and ?exible contour tracking method isproposed based on the new model. The method is applicable for tracking fast movingobject and deformable object, and no static background is assumed. In addition, themodel can control the topology change adaptively as other geometric snakes.更多还原