一种利用对象级条件随机场的道路提取方法

An Object-based Conditional Random Fields for Road Extraction

针对传统的马尔科夫随机场不能利用观察图像中的上下文信息、条件随机场虽然能够同时利用两种上下文信息,但基于像素的条件随机场模型抗噪能力差、计算量大和效率低的问题,该文结合条件随机场和基于对象的图像分析方法,提出一种新的对象级条件随机场,并用于遥感图像道路的提取。该方法利用各个对象构建的邻接关系,建立基于对象的条件随机场模型;将道路提取问题归结为一个二类分类问题,并采用三类共29维特征进行模型训练和推断。实验结果表明:相比基于像素的条件随机场模型,本文方法提取精度更高,训练和推断的时间明显减少。

Traditional Markov Random Fields （MRF） can use contextual information in labeled image, but cannot use contextual information in observed image, though the Conditional Random Fields （CRF） can make use of these two kinds of contextual information,but the traditional pixel-based CRF is with poor noise immunity and complexity computation. This paper combines conditional random and object-based image analysis, and proposes a new object-based CRF used for extraction. This method uses the adjacency relation of each object, builds a Conditional Random Fields model based on the object, the road extraction problem is reduced to a two class classification problem,and the model training and inference are carried out by using three classes of 29 dimensional features. Experimental results show that compared to pixel-based CRF, this method is with higher accuracy and less time for training and inferring.