Histogram Thersholding for Automatic Color Segmentation Based on K-means Clustering

Abstract. Color segmentation method has been proposed and developed by many researchers, however it still become a challenging topic on how to automatically segment color image based on color information. This research proposes a method to estimate number of color and performs color segmentation. The method initiates cluster centers using histogram thresholding and peak selection on CIE L*a*b* chromatic channels. k-means is performed to find optimal cluster centers and to assign each color data into color labels using previously estimated clusters centers. Finally, initial color labels can be split or merge in order to segment black, dark, bright, or white color using luminosity histogram. The final cluster is evaluated using silhouette to measure the cluster quality and calculate the accuracy of color label prediction. The result shows that the proposed method achieves up to 85% accuracy on 20 test images and average silhouette value is 0.694 on 25 test images. Keywords: Automatic color segmentation; Histogram thresholding; Cluster centers initialization; k-means clustering