An adaptive Bayesian segmentation algorithm for color images is presented, which extends the adaptive clustering approach of Pappas to multichannel images. A scalar segmentation label field is generated for the multichannel data, which is modeled as a vector field, where the components of the vector field (each individual channel) are assumed to be conditionally independent given the segmentation labels. The class conditional probability model for the vector image field is taken as a multivariate Gaussian with a space-varying mean function. A Gibbs random field is employed as the a priori probability model for the segmentation label field that imposes a spatial connectivity constraint on the labels. The space-varying class means associated with the image segments can be used to form an estimate of the actual image from noisy observations. Experimental results are provided to demonstrate the benefits of using adaptivity via the space-varying means and the spatial connectivity constraint. We also discuss the effects of the color space within which the clustering is performed on resulting segmentations.