Meaningful regions segmentation in CBIR
Shikui Wei, Yao Zhao, Zhenfeng Zhu · 2005
In this paper, a new approach to fully automatic image segmentation is proposed to get the meaningful regions of general-purpose image. In order to avoid image over segmenting, the original input image is first smoothed by Gaussian filters with different scales. Then an improved ISODATA clustering algorithm with parameters selecting dynamically is proposed to cluster the image pixels into different regions. To eliminate those fragmentary regions, a region merging strategy is also presented. The final experimental results show that the proposed approach can effectively separate the objects from background of general-purpose image.