基于平流矢量场和弥散流改进的几何活动轮廓模型

Improved geometric active contour model based on advection vector field and diffusion flows

几何活动轮廓通过水平集演化捕获目标边界,然而传统方法忽略了前景物体涉及的全局运动及局部形变,从本质上限制了轮廓的动态谐调范围。提出一种由隐式流驱动的几何活动轮廓方法,用于分割形变对象的动态轮廓。根据流体力学,对平流及弥散的隐式流进行建模,从而直接引导水平集的更新。根据图像序列之间有限的运动观测,采用具有平滑和稀疏项的回归算法以拟合平流矢量场。然后,引入受空间梯度约束的非均匀弥散方程,将最终叠加的隐式流用于水平集演化。提出的几何活动轮廓方法构成了一个统一的扩展框架,保留了已有模型对轮廓拓扑变化的适应性。对多个场景下的动态轮廓分割实验表明,提出方法能准确地分割具有全局运动及不规则形变的前景目标,且具有收敛速度快、稳定性强的优点。

Propose a geometric active contour approach driven by implicit flows for segmenting deformative objects. The geometric active contours capture dynamical shapes by yielding initial level-set to image features. However, the interesting objects are often associated with salient motions, which has been ignored by naive level-set methods and thus intrinsically limit the harmonizing range. According to hydrodynamics, the implicit flows involving advections and diffusions are formulated to directly guide the level-set updating. To generate the advective vector field, a regression algorithm with smoothing and sparse is embedded in terms of finite motion vectors among sequence. Also the implicit flows are simultaneously synthesized with the non-uniform diffusions restrained by spatial gradients. The proposed improved geometric active contour model based on advection vector field and diffusion flows is a unified and efficient framework, also the topological preservation is inherited from the original geometric active contours. Experimental results of applying this method to real scenes show that the method has fast convergence speed and can accurately segment deformable objects with global motion.