Hierarchical multiscale image segmentation

Abstract

We present a new image segmentation algorithm by developing a hierarchical tree through a non-linear scale-space approach. Aiming at preserving edges better than the Gaussian blurring, this segmentation process is accomplished in three stages. Our method starts creating a scale-space from the original image and a whole family of blurred images is generated using a modified version of the wavelet a trous algorithm. In the following steps, the hierarchical tree of blurred images is built and the linking procedure is performed so that the segmentation process is concluded. Experiments show that the proposed method works quite satisfactory and produces good results.