摘要： 在图像分割中,将条件随机场(CRF)模型及其高阶模型广泛用作能量函数,后者以二阶CRF模型为基础,通过引入高阶势函数反映各分割块内像素标记的一致性,使分割的目标边缘更加精确,但能量最小化的计算效率不理想。针对该问题,提出一种基于鲁棒PnPotts高阶CRF模型的改进型图像分割算法。根据给定的标记集合运行最大流/最小割算法得到局部最优解,再用局部最优解修改节点的标记,对未确定标记的节点运行α扩展算法,并在每次迭代过程中动态更新图的流和边的剩余容量,使得每次迭代的时间快速减少。实验结果表明,与α扩展算法相比,改进算法在保持原有分割效果的基础上,相同图像的能量最小化收敛速度比原算法快2倍~3倍。

关键词: 高阶条件随机场模型, 图像分割, 能量最小化, 最大流/最小割, 局部最优解, α扩展算法

Abstract: In image segmentation,Conditional Random Field(CRF) model and its higher-order model are widely used as energy function.The latter is based on second-order model by introducing higher-order potential function to reflect consistency of pixel labels of each segment,so that the segmented object boundary is more accurate.But the computing efficiency of energy minimization is not ideal.Aiming at this problem,this paper presents an improved image segmentation algorithm based on robust PnPotts high-order CRF model,which computes local optimal solution by running max-flow/min-cut algorithm according to the set of labels.The local optimal solution is used to modify the label of nodes and the α expansion run for the remaining nodes.The flows and residual capacities of the side are dynamically updated in each iteration process,so the running time of each iteration rapidly decreases.Experimental results show that compared with the α expansion algorithm is run,the improved algorithm can not only maintain the original segmentation effect,but also achieve 2~3 times of speed-up in the time of energy minimization on the same image.

Key words: high-order Conditional Random Field(CRF) model, image segmentation, energy minimization, max-flow/min-cut, local optimal solution, α expansion algorithm

中图分类号: 

• TP301.6