基于图割优化的能量最小化模型图像分割方法

Energy minimization model for image segmentation via graph cut optimization

针对主动轮廓模型中利用梯度下降法求解能量函数容易陷入局部极小的不足,设计了一个离散化最小能量函数模型。该模型以Chan-Vese模型为基础,利用图割方法优化能量泛函,实现能量的全局最优解。新模型首先将图像映射为图,将基于像素的能量泛函转换为可用图表示的离散化能量函数,通过计算节点及其邻域关系权值,迭代求解最小化能量并将其作用于形变轮廓曲线,直至达到稳定状态。新模型改进了主动轮廓模型对弱边界图像初始轮廓敏感的问题,提高了分割精度和运行速度。

Aiming at the drawbacks of active contour models which used gradient descent and result in local minimum easily,this paper proposed a discrete energy minimization model for image segmentation.It designed the new model based on Chan-Vese model and optimized the energy function via graph cut method.It could find a global minimum rather than a local one.To construct the new model,first step was to map the image for a graph,and then changed the level set energy function into a discrete form which should be proved graph-representable.Using the model traversed each node and its neighborhood,it computed the weights of the edges with these nodes,and got the new labels of pixels and updated the initial contour,until the energy remained constant and the contour reached the boundary of object.The major advantages of this model included the existence of global minimum and its insensitivity to initialization.Numerical implementations show that the model improves the accuracy and speeds for image segmentation.