自适应特征加权的Gibbs随机场影像分割方法

Image Segmentation by Combining Adaptively Weighted Features with Gibbs Random Field

针对现有分割算法很少同时兼顾不同特征分量区分能力的差异和相邻像素间的相关性的问题,提出了一种结合Gibbs随机场的特征加权遥感影像分割方法.该方法首先依据训练样本计算各特征分量的区分能力,确定不同地物类别相应的特征分量的权重;然后利用加权最小距离分类法对影像进行初始分割,并利用Gibbs随机场来描述像素的空间相关性;最后综合Gibbs随机场描述的标记场和加权最小距离分类法描述的特征场来获取影像的最终分割结果.实验结果表明,Gibbs随机场能够有效地描述空间相关性,根据区分能力确定的权重强化了区分能力强的特征分量.

Fewexisting image segmentation methods simultaneously take into account both the distinguishability of different features and the relationship between neighboring pixels. In this paper,a novel image segmentation algorithm is proposed by combining the adaptively weighted features with the Gibbs random field. First,the distinguishability of each component of image features for each land-cover type is defined as a weight parameter,which is determined by the corresponding component of the training samples belonging to the same land-cover type. Second,the initial segmentation is obtained by using the minimum distance classifier,and the spatial correlations of neighboring pixels are modeled by the Gibbs random field. Finally,the label field,which is modeled as the label prior of Gibbs random field,and feature field,which is represented as the normalized weighted distance of weighted features,are combined together to generate the segmentation result. Experimental results demonstrate that the Gibbs random field can effectively describe the spatial relationship,and the adjusted weight can strengthen the distinguishability of the feature component,which can distinguish different land-cover objects accurately.