Image segmentation is a first step to vision system and used as a pre-processing for many applications such as pattern recognition, image classification, picture coding or target tracking. In the previous papers, we reported an unsupervised image segmentation method based on Bayesian classifier and applied it to object-to-object color transformation. Although Bayesian decision rule is a robust tool to classify the objects statistically with the minimum error in average, it needs to preset some appropriate class centers before starting the classifier. The location of initial seed points much influences the segmentation accuracy. This paper discusses a better way to set the initial seeds and reports the Bayesian discriminator works better when coupled with k-means classifier for correcting the location of seed points. In addition, the paper introduces a new application of proposed model into scene color interchanges between segmented objects.