基于邻接表的分水岭变换快速区域合并算法

Fast region merging algorithm for watershed transform based on adjacency list

针对图像分水岭变换存在的过分割问题,提出了一种快速区域合并算法.该算法先对梯度图像进行Lee滤波以降低分水岭变换的过分割程度,在传统分水岭变换初步分割的基础上,利用区域邻接图(RAG,Region Ad jacency Graph)的邻接表数据结构描述过分割区域之间的关系,采用分级阈值合并的方式完成区域合并,以降低合并计算的复杂度.合并过程中,将基于区域相对边界强度和边界长度的相似度度量准则与传统的区域灰度均值差异度准则相结合,构成新的区域可合并度综合评价函数,以保证区域合并准确可靠.实验结果表明,与传统的分水岭变换区域合并算法相比,该算法有效地提高了区域合并的准确性,并大幅提升了运算速度.

A fast region merging algorithm was proposed to solve the oversegmentation problem produced by the watershed transform for image segmentation. Firstly, the gradient image was preprocessed with Lee filter to reduce the oversegmentation initially. Then the region adjacency graph （RAG） with its adjacency list of data structure was used to represent the image partitions after the initial partitioning with the classical watershed transform. On the basis of the adjacency lists the region merging process followed a two-step threshold merging strategy in order to reduce the computation complexity. A hybrid region dissimilarity function was presented to measure the degree of similarity between two regions for the region merging. It combined the relative boundary integrity criteria and the boundary length criteria with the traditional region homogeneity criteria to increase the contour accuracy of merged regions. Experimental results show that this algorithm improves the region merging accuracy and processing speed greatly.