在系统回顾和比较了基于颜色的皮肤检测的方法和技术(其中包括:颜色空间选择、肤色建模方法、动态跟踪模型以及光照不变性与自适应模型)的基础上;基于一个包含1 894张图片的大样本库,着重比较了肤色在14个3维颜色空间和14个2维色度平面中的分布紧致性、肤色与非肤色类之间的可分辨性,以及肤色概率图(SPM)、高斯混合模型(GMM)、自组织映射图(SOM)和支持向量机(SVM)在这些颜色空间中的皮肤分类性能。比较结果表明:(1)颜色空间的变换并不能改善肤色紧致性、肤色-非肤色可分辨性以及分类等性能,但RGB及线性变换空间却具有较好的类可分辨性和分类性能;(2)去除亮度信息将明显降低肤色和非肤色之间的可分辨性和分类性能;(3)Bayes决策下的3维SPM的分类性能是最优和空间无关的,而其余分类器则普遍存在类似的“空间偏好性”;(4)同时采用肤色和非肤色模型的分类器的分类性能优于仅使用肤色模型的分类性能。

This paper presents a comprehensive review and evaluation for color-based skin detection.The literature is reviewed in four categories: colorspace selection,skin color modeling,dynamic modeling and illumination invariance and adaptation.Using a large data set of 1 894 images,we examine whether the colorspace transformation can increase the compactness of skin distribution and the discriminability between skin and nonskin distributions in fourteen 3D colorspaces and fourteen 2D chrominance planes.We also evaluate the classification performance of skin probability map(SPM),Gaussian mixture model(GMM),self organizing map(SOM) and support vector machine(SVM) in said colorspaces.The results reveal that 1) the colorspace transformation cannot improve performance in general,the discriminability and testing performance in RGB and linear colorspaces are better than in other colorspaces,2) the Absence of the luminance component decreases discriminability and performance significantly,3) the performance of Bayes SPM in 3D colorspaces is superior to that of others,4) except 3D Bayes SPM,the 'color preference' of other detector is intrinsic and quite similar,5) the detector using skin and nonskin model simultaneously is better than the detector using skin model alone.