Multiresolution adaptive and progressive gradient-based color-image segmentation

摘要

We propose a novel unsupervised multiresolution adap-ntive and progressive gradient-based color-image segmentation algo-nrithm (MAPGSEG) that takes advantage of gradient information innan adaptive and progressive framework. The proposed methodol-nogy is initiated with a dyadic wavelet decomposition scheme of annarbitrary input image accompanied by a vector gradient calculationnof its color-converted counterpart in the 1976 Commission Interna-ntionale de l’Eclairage (CIE) L*a*b* color space. The resultant gradi-nent map is used to automatically and adaptively generate thresholdsnto segregate regions of varying gradient densities at different reso-nlution levels of the input image pyramid. At each level, the classifi-ncation obtained by a progressively thresholded growth procedure isnintegrated with an entropy-based texture model by using a uniquenregion-merging procedure to obtain an interim segmentation. A con-nfidence map and nonlinear spatial filtering techniques are combined,nand regions of high confidence are passed from one resolution levelnto another until the final segmentation at the highest (original) reso-nlution is achieved. A performance evaluation of our results on sev-neral hundred images with a recently proposed metric called the nor-nmalized probabilistic Rand index demonstrates that the proposednwork computationally outperforms published segmentation tech-nniques with superior quality.