基于多尺度局部区域能量最小化算法的图像分割

Algorithm of minimizing multi-scale local region energy for image segmentation

常规多尺度MRF模型中固定的四叉树结构造成图像分割结果中常产生块现象和非连续边缘.为解决这一问题,提出了一种新的多尺度MRF模型,并建立了基于区域消息传递的置信度传播(BP)算法,通过BP算法在多尺度MRF模型中对区域消息进行传递;在层间,从粗糙层向精细层进行消息传递时,利用提出的MRF模型父子区域之间的重叠,有效初始化了精细层消息的初值,避免了多尺度MRF模型层间误分类的传递;最后基于MPM准则对分割结果进行估计.实验结果表明提出的算法不仅得到了更准确的图像分割结果,而且具有较快的分割速度.

An efficient local region belief propagation （BP） algorithm based on multi-scale Markov random fields （MRF） model is proposed to solve the problem that the fixed quadtree structures of traditional hierarchical Markov random fields always results in blocky artifacts and discontinuous edges in image segmentation. The proposed algorithm builds different scales of local region messages, and the messages are propagated on the proposed MRF model through belief propagation （BP） algorithm. The proposed algorithm utilities the overlapping between parent and child regions efficiently to initialize the child region messages when passing messages from the coarser scale to the finer scale, thus avoiding the transfer of misclassification between scales in multi-scale MRF model. Finally segmentation results are estimated based on the maximized posteriori marginal （MPM） criterion. Experimental results on a wide variety of images have verified the effectiveness of the proposed algorithm.