Color textile images usually show a few dominant colors, and the inherent fabric thread structure makes it a difficult job for automatic clustering-based techniques to extract dominant colors from textile images. Based on the two distinctive features of textile images, a probabilistic reasoning segmentation algorithm for color textile images is proposed. Due to the uniform texture of the fabric appearing in textile images, dominant colors are extracted interactively. Then a hierarchic probabilistic reasoning model is applied to capture not only the statistical dependences of color information across adjacent scales, but also those among intrascale neighbor blocks. The multiscale approach is used to avoid the conflict between boundary localization and high-resolution segmentation by deducing the maximum posterior probability for each block recursively from coarse to fine scale. No special prior distribution assumption is made about the size and shape of regions in this algorithm. That there is no need to train the multiscale contextual model prior to the segmentation is one of the big advantages of this approach. Experimental results show that the proposed algorithm can produce better segmentation results and smoother edge maps for color textile images than some state-of-the-art segmentation techniques, both supervised and nonsupervised.