采用多层图模型推理的道路场景分割算法

发布时间：2018-11-20 21:52

【摘要】：针对传统图模型分割算法提取的物体边缘不够精细、难以适应复杂道路场景布局的问题,提出了一种基于多层图模型推理的道路场景分割(HGI)算法。该算法先将图像过分割为同质的超像素块,再采用随机森林模型训练超像素块的多类别回归器和相邻超像素的一致性回归器;然后用2种回归值计算马尔科夫随机场(MRF)模型的能量项,通过推理得到初始分割;最后为了解决超像素块包含多类别带来的分类混淆,在初始分割基础上构建像素级的全连接条件随机场模型,进行优化得到精细的分割结果。实验结果表明,采用HGI算法对人工标注数据库和真实拍摄的场景图像处理能够得到精细的分割边缘,能够解决超像素推理中的类别混淆问题,与传统的MRF图模型分割方法相比,在总体精度和平均召回率2个指标上分别提高了2%和3%。
[Abstract]:Aiming at the problem that the object edge extracted by traditional graph model segmentation algorithm is not fine enough to adapt to the complex road scene layout, a road scene segmentation (HGI) algorithm based on multi-layer graph model reasoning is proposed. The algorithm firstly divides the image into homogeneous super-pixel blocks and then uses a stochastic forest model to train multi-class regression of super-pixel blocks and consistency regression of adjacent super-pixels. Then the energy term of Markov random field (MRF) model is calculated with two regression values, and the initial segmentation is obtained by reasoning. Finally, in order to solve the classification confusion caused by the super-pixel block including multiple categories, a pixel level conditional random field model is constructed on the basis of initial segmentation, and the fine segmentation results are obtained by optimization. The experimental results show that the HGI algorithm can get fine segmentation edge of the artificial tagged database and real scene image processing, and can solve the problem of class confusion in super-pixel reasoning. Compared with the traditional MRF image model segmentation method, the proposed algorithm can solve the problem of classification confusion in super-pixel reasoning. The overall precision and average recall rate were increased by 2% and 3% respectively.
【作者单位】：西安电子科技大学综合业务网理论及关键技术国家重点实验室;
【基金】：国家自然科学基金资助项目(61502364)
【分类号】：TP391.41

【相似文献】