信息熵约束下的视频目标分割

Video object segmentation via information entropy constraint

大部分基于图论的视频分割方法往往先通过分析运动和外观信息获得先验显著性区域,然后用最小化能量模型来进一步分割,这些方法常常忽略对外观信息精细化分析,建立的目标模型对复杂场景的鲁棒性不佳。根据信息熵能够度量样本纯度,信息熵最小化和能量模型最小化具有一致的目标,提出一种信息熵约束下的视频目标分割方法。首先在经典光流法基础上结合点在多边形内部原理获得第一阶段的分割结果;然后以超像素为基本分割单元,获得均匀的运动和表现;最后在能量函数中引入信息熵约束项,构建前景背景像素标记的优化问题,通过最小化能量函数得到更精确的分割结果。在公开数据集上的实验结果表明目标模型中引入信息熵约束项能够有效提高视频目标分割的鲁棒性。

In most of the graph-based segmentation methods,prior saliency regions are often obtained by analyzing motion and appearance information and then the energy model was minimized for further segmentation.These methods often ignore refined analysis of appearance information,and are not robust to complex scenarios.Since information entropy can measure sample purity and information entropy minimization has a consistent goal with energy model minimization,a video object segmentation via information entropy constraint was proposed.Firstly,the segmentation results of the first stage were obtained by combining with optical flow vector and the point-in-polygon principle from the computational geometry.Secondly,the uniform movement and performance were gained through presenting superpixel as the basic division unit.Finally,video segmentation was formulated as a pixel labeling optimization problem with two labels by introducing information entropy constraint into energy function,and more accurate segmentation results were obtained by minimizing the energy function.The experimental results on public datasets show that the proposed method can effectively improve the robustness of video object segmentation.