基于鲁棒高阶条件随机场的视频自动分割

Automatic Video Segmentation Based on Robust Higher Order Conditional Random Field

针对交互式分割方法存在用户标注繁琐和过分割现象,以及仅考虑二元项不能获得图像中准确的物体边界等问题,结合鲁棒高阶条件随机场,提出一种视频自动分割方法。采用基于超像素显著性特征的分割方法对视频初始帧进行自动分割,其结果作为初始化种子建立模型。根据颜色信息设计高斯混合模型,基于纹理、形状等特征,利用联合Boosting算法训练Jointboost强分类器模型,通过条件随机场提高分割准确度。引入基于超立体像素的高阶项,增加像素与区域的关联,提高分割边界的平滑度。实验结果表明,该方法明显地提高了分割效果。

This paper presents an automatic video segmentation method based on robust higher order Conditional Random Field（CRF） ,which alleviates the problem that interactive segmentation is time-consuming and labor-intensive, and oversegmentation is generated in unsupervised segmentation, and simple pairwise-pixel segmentation cannot get accurate boundary. It utilizes the saliency based segmentation of the first frame of video as initial seeds instead of user labeling. The Gaussian mixture model and a strong jointboost classifier model are respectively learned on the features of color,texture and shape, the combination of both in CRF improves the accuracy of segmentation. It adds higher order potential based on supervoxel to solve the shortcoming of oversmoothing of pairwise-pixel segmentation. Experimental results demonstrate that the method is more effective and efficient than the state-of-art methods.