In this thesis, we propose a fast unsupervised multiresolution color image segmentation algorithm which takes advantage of gradient information in an adaptive and progressive framework. This gradient-based segmentation method is initialized by a vector gradient calculation on the full resolution input image in the CIE L*a*b* color space. The resultant edge map is used to adaptively generate thresholds for classifying regions of varying gradient densities at different levels of the input image pyramid, obtained through a dyadic wavelet decomposition scheme. At each level, the classification obtained by a progressively thresholded growth procedure is combined with an entropy-based texture model in a statistical merging procedure to obtain an interim segmentation. Utilizing an association of a gradient quantized confidence map and non-linear spatial filtering techniques, regions of high confidence are passed from one level to another until the full resolution segmentation is achieved. Evaluation of our results on several hundred images using the Normalized Probabilistic Rand (NPR) Index shows that our algorithm outperforms state-of the art segmentation techniques and is much more computationally efficient than its single scale counterpart, with comparable segmentation quality.

Library of Congress Subject Headings

Image processing--Digital techniques; Computer vision

Publication Date


Document Type


Department, Program, or Center

Electrical Engineering (KGCOE)


Dianat, Sohail

Advisor/Committee Member

Amuso, Vincent

Advisor/Committee Member

Saber, Eli


Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TA1632 .V36 2009


RIT – Main Campus