基于高阶CRF模型的图像语义分割

图像语义分割方法大多基于点对条件随机场模型, 不能定位到单个目标, 并且难以利用全局形状特征, 造成误识。针对这些问题, 提出一种新的高阶条件随机场模型, 将基于全局形状特征的目标检测结果和点对条件随机场模型统一在一个概率模型框架中, 同时完成图像分割、目标检测与识别的任务。利用目标检测器和前背景分割算法获取图像中目标区域, 在目标区域上定义新的高阶能量项。新的高阶条件随机场模型就是高阶能量项和点对条件随机场模型的加权混合模型, 其最优解即为图像语义分割结果。在MSRC-21类数据库上进行的实验验证了该模型能够显著提升图像语义分割性能, 并定位到单个目标。

Image semantic segmentation based on higher-order CRF model

Current image semantic segmentation methods mostly use pairwise conditional random field (CRF) models, which can not distinguish instances of objects and tend to recognize wrongly for lack of global shape features. To solve of these problems, this paper proposed a new higher-order CRF model, which incorporated the pairwise CRF model and object detection based on global shape features into a unified probabilistic framework, and completed image segmentation, object detection and recognition tasks all at the once. It defined new higher-order energy terms on the object regions which were segmented out by the object detector and foreground-background segmentation algorithm. The proposed higher-order CRF model was the weighted combination of the higher-order energy terms and pairwise CRF model. The experiments conducted on the MSRC 21-class database show that the new higher-order CRF model can improve image semantic segmentation and locate instances of objects.