CAJ | 학술논문

提出了一种从遥感影像中分割出道路的新方法.通过比较多个方向上纹理和灰度的一致性,可以在每一点处筛选出一致性最优方向,此方向上的Gabor滤波响应以及作为衡量一致性指标的方差值共同构成该方法的特征矢量.鉴于MRF的独特优点,选择其作为分割模型,并结合最大期望(EM)原理实现非监督分割,构建MRF的相应能量项时加入了方向信息以提高分割质量,同时用置信传递法(BP)代替传统的模拟退火法以提高能量最小化过程的速度.实验表明新方法可以更为有效地从高分辨率遥感影像中分割出道路目标.
제출료일충종요감영상중분할출도로적신방법.통과비교다개방향상문리화회도적일치성,가이재매일점처사선출일치성최우방향,차방향상적Gabor려파향응이급작위형량일치성지표적방차치공동구성해방법적특정시량.감우MRF적독특우점,선택기작위분할모형,병결합최대기망(EM)원리실현비감독분할,구건MRF적상응능량항시가입료방향신식이제고분할질량,동시용치신전체법(BP)대체전통적모의퇴화법이제고능량최소화과정적속도.실험표명신방법가이경위유효지종고분변솔요감영상중분할출도로목표.
A new method is given for differentiating roads from other objects in high resolution remotely sensed images. By comparing the consistency of gray value and one dimensional texture in different directions, the most consistent direction at every pixel is found. The responses of a group of Gabor filters as well as the variance indicating the consistency in this direction construct the feature vector. The Markov random field（MRF）principle is utilized to model the road segmentation issue as it has a number of advantages, meanwhile the expectation-maximization algorithm is introduced to estimate the model parameters automatically and produce unsupervised segmentation results. In order to reduce the computing cost, traditional simulated annealing scheme is replaced by the belief propagation (BP) approach. Experiments demonstrate that the proposed method can partition roads from other targets in high resolution RS images more efficiently than some commonly used algorithms.