A Hybrid Color Space for Skin Recognition for Real-Time Applications

Abstract

Color space conversion utilized by many researchers in order to enhance skin recognition performance by projecting the skin color cluster to a more distinctive distribution. In spite of the substantial research effort in this area, finding a suitable color space for face and skin recognition is still an unsolved issue. Deviation of skin tone under different lighting condition, dissimilarity of skin color among different ethnics and races, various camera sensors characteristics, presence of skin-like color objects in image background and variation of skin color tone among different body limbs are among the major challenges in skin recognition. Majority of these challenges are expected to be mitigated through color space conversion. This paper proposes a new hybrid color space by applying Principal Component Analysis technique to skin color cluster in ten existing conventional color spaces including RGB, YCbCr , YUV, nRGB, i1i2i3, YIQ, XYZ, YPbPr , YES, YCgCr . The proposed hybrid color space which termed P1P2P3 consist of the three major Principal Components of these conventional color spaces components. Using Algebraic simplification these principal component has been reformulated in terms of RGB color space. Parametric pixel wised skin detection techniques have been employed in order to evaluate the proposed color space effect on skin detection performance. Three popular supervised classifiers including Multilayer Perceptron, Support Vector Machine and Random Forest has been employed to generate a parametric model of skin color cluster using the proposed color space. Experiment results shows the proposed hybrid color space P1P2P3 with F -score and False Positive Rate 0.960 and 0.041 respectively performed better than the existing conventional color spaces in term of pixel wised skin recognition.