K-Means, Mean Shift, and SLIC Clustering Algorithms: A Comparison of Performance in Color-based Skin Segmentation

摘要

Commonly used in computer vision, segmentation is grouping pixels into meaningful or perceptually similar regions. In this work, we are going to evaluate the performance of three popular data-clustering algorithms, the K-means, mean shift and SLIC algorithms, in the segmentation of human skin based on color.ududThe K-means algorithm Iteratively aims to group data samples into K clusters, whereudeach sample belongs to the cluster with the nearest mean. The mean shift algorithm is a non-udparametric algorithm that clusters data iteratively by finding the densest regions (clusters) in a feature space. An enhanced version of the classic K-means algorithm, the SLIC limits the search region to audsmall area around the cluster reducing the algorithm complexity to be only dependent onudthe number of pixels in the image. It also provides control over the compactness of theudclusters.ududColor-based skin segmentation algorithms depend on both a color space at which segmentation is performed and a classification method used to determine whether a pixel is skin or non-skin. We have implemented the K-means, mean shift andudSLIC algorithms in the RGB color space to detect human skin. Our method beginsudby clustering images using these algorithms and then segmenting the clustered regionsudoccupied by skin. Pixels in the clusters are classified as skin or non-skin using the Kovacudmodel.ududWe have evaluated the algorithms' performance on the SFA database (controlled environ-udment) and on another database created for testing on an uncontrolled environment. The performance has been evaluated using time complexity, F1 score, recall, and precision. We have found that on average the mean shiftudalgorithm triumphs over the three algorithms in terms of performance while the SLIC algorithms holds an advantage being the fastest.The K-means algorithm has a good performance when the number of clusters K is between 10 and 15, whereas the mean shift algorithm has good performance when the bandwidth h is between 0.03 and 0.06. The SLIC algorithm maxes out its performance at around k = 100 and the number of clusters can be increased to K = 300 without remarkably increasing the complexity.