Please use this identifier to cite or link to this item:
|Title:||Skin detection||Authors:||Thuong, Phan.||Keywords:||DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision||Issue Date:||2012||Abstract:||In recent years, skin detection is widely used in various applications of computer vision such as face detection, face recognition, image filtering or hand gesture analysis. Most of the existing skin detection methods are based on pixel-wise classification which aims to classify all image pixels as skin or non-skin individually. By evaluating many skin detection algorithms that have been proposed in previous researches and studies, it is suggested that one of the most efficient pixel-based classification methods is Bayesian classifier with histogram technique. Using this algorithm, I have developed a program to segment skin from a large image database. The program requires a large collection of diverse images to build skin and non-skin models. A training dataset has been built with more than 900 images, 10% of the images are collected from the current skin database of Asst/Prof. Kong Wai-Kin Adams; the rest are collected manually from the Web. All images are chosen such that they are in high resolution and diverse in terms of background scenes and lighting conditions. For skin dataset, different types of human skin such as yellowish, pinkish, whitish, light brown, dark brown are collected. However, more than 90% of selected skin samples are from Asian and European people; thus, the classification program will perform well if test images only contain skin types of these people. The program yields good results with high classification accuracy, up to 94.9%. Furthermore, a comparison between the classification performance in different color representations such as RGB, HSV and HS is discussed. It is shown that the selection of color space for the classifier using Bayesian classifier with histogram technique can affect segmentation results. Especially, if luminance component is removed from color space, the classification performance will degrade sufficiently.||URI:||http://hdl.handle.net/10356/49078||Rights:||Nanyang Technological University||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.