一种基于全连接CRF的前景-背景分割方法

Foreground-Background Segmentation Based on Full Connected Conditional Random Fields

像素级的前景-背景分割通常被当做一个基于条件随机场的能量最小化问题,但是基于的只是局部连接的条件随机场,全连接的条件随机场因为复杂度高而不被采用。通过使用均值场近似技术将邻居节点间的约束转换成低通滤波操作,虽然简化了全连接条件随机场的计算,但也丢失了大量的相关性信息。为了克服这种信息丢失,对临近的像素间的二元约束进行保持,只将空间距离较远的像素间的二元约束转换成低通滤波,并添加了局部的光滑项进行分割边缘约束,然后使用图割算法对最后的能量函数进行优化。实验结果显示,算法由于充分利用了全局约束信息,对具有复杂边缘、细小枝状边缘、凹陷边缘的物体具有较好的分割效果。

Pixel-level fore/back-ground segmentation is often treated as the minimization of energy function based on conditional random fields（CRF） , but it is just based on the locally connected conditional random fields, not the fully con nected because of huge complexity of fully connected CRF. The binary constraints between the pixels are converted into lowpass filtering operations by mean-field approximation, while simplifying the computation, but losing a large amount of correla- tion constraint. In this paper, to avoid the loss of correlation constraint, only binary constraints between the pixels at the far distance are considered to transform into the low-pass filtering operations, while the binary constraints between the adjacent pixels are kept. In addition, the local smooth term is added to constrain the edges. Then, the graph cut algorithm is applied to optimize the final Energy function. Experimental results show that the algorithm has good segmentations benefiting from more fully utilized of global binary constraints, especially on the objects with complex edges, dendritic components or depressed edges.