Xiong et al. developed an improved color constancy solution, GSI, by identifying all of those potential gray surfaces and average them in the RGB space.[1] This strategy assumes that the gray surface, no matter what the lighting color is, have an S axis of nearly zero in the LIS coordinate system that roughly correspond to the variation of intensity, illumination and reflectance.[1,2] However, this assumption is often violated, where some non-gray surfaces under specific illumination would be mistaken as gray ones. Simply averaging all detected pixels would bias the illumination estimation. To overcome this problem, the GSI is extended by analyzing color distribution of all identified gray surfaces in perspective chromaticity space. We employ an unsupervised cluster technique to extract the group with maximal data distribution density among them. The center of the selected cluster is then used to approximate the illumination colors. The advantage of the cluster technique is that we don't need a large training data set to establish the relationship between the statistical properties of image data and the lighting color incident on it. The experiments based on two real image data set show that this method is comparative to other elaborate existing color constancy methods and has lower costs than most existing color constancy methods.