面向影像分割的多尺度快速区域合并方法

摘要

影像分割的区域合并技术中,传统的采用区域邻接图的方法存在着数据结构和算法复杂、难以扩充和维护、可考虑的特征因子有限以及空间浪费严重等问题.对此,提出了一种新的区域合并方法,提供了一套采用面向对象技术解决区域合并问题的新框架.在该框架下,区域的相异度指标、属性、邻接关系和行为可根据需要自由定义、扩充和修改,算法的稳定性和可维护性得到提升,合并过程被充分简化.在此基础上提出了多尺度合并区域的改进方法,并对等级队列的构建机制进行了优化.最后通过多尺度的对比实验,证明该方法不但能保证区域合并的精度,而且可以显著提高执行效率.%As a region merge technology for image segmentation, traditional method of Region Adjacency Graph (RAG) has lots of shortness such as complicated data structure and algorithm, difficulty of expansibility and maintenance, limits on available feature factors and great amount of memory waste. Therefore, a new approach is proposed which provides a new framework for region merge using object-oriented technique. In this frame work, the dissimilarity index, properties, adjacency and actions of regions can be defined, extended and merge process is fully simplified. Additionally, an improved method of obtaining multi-scale merge results is proposed and the regulation of building hierarchical queue is also optimized. Experiments have shown that on the precondition of guaranteed merge (segmentation) effect, operational speed of the approach is much faster than RAG and Nearest neighbor graph(NNG).