Image segmentation is a first step to vision system and used for many applications such as pattern recognition, image classification, understanding, or picture coding. In the previous work, we reported unsupervised image segmentation by our k-means Bayesian classifier and applied it to automatic scene color interchange. Although Bayesian decision rule is a robust tool based on the minimum error criterion, it needs to preset any appropriate class centers before starting the classifier. Since the location of initial seed points much influences segmentation accuracy, the better color clustering is a key to success for image segmentation.This paper presents a novel approach to a non-parametric color clustering by introducing Parzen window and discusses how to estimate the probability density function and how to preset the reliable seed points. The paper proposes a Particle model as an alternative and fast algorithm for Parzen window model. Experimental results applied to unsupervised image segmentation are demonstrated.