Probabilistic Formulation for Skin Detection Sanun Srisuk 42973003 Seminar I 1 Outline Problem Statement Literature Review Proposed Skin Detection Experimental Results Conclusions 2 Problem Statement Face Detection Methods •Shape Analysis •Fuzzy Pettern Matching •Neural Networks •SVM •Hausdorff Distance 3 Literature Review Skin Detection using Fuzzy Theory Skin Detection using Color Statistics 4 Skin Detection using Fuzzy Theory • • • Wu et al. [12] propose a method for skin detection using fuzzy theory. SCDM and HCDM are skin and hair color models. The perceptually uniform color system (UCS) is used for color representation. 5 Skin Detection using Fuzzy Theory (cont.) SCDM 1. Manually select skin regions in each image. 2. Prepare a table of 92x140 entries to record the two dimensional chromatic histogram of skin regions, and initialize all the entires with zero. 3. Convert the chromaticity value of each pixel in the skin regions to UCS, and then increase the entry of the chromatic histogram corresponding to it by one. 4. Normalize the table by dividing all entries with the greatest entry in the table. 6 Skin Detection using Fuzzy Theory (cont.) HCDM Skin and Hair Color Detectors 7 Skin Detection using Fuzzy Theory (cont.) The method proposed in the Wu et al. scheme sometimes fails to detect the real face. Reasons under concern include the followings. •Illumination: This is because, the luminance information is used to detect the hair part of faces, the variance of the illumination color will affect the detection result. •Hairstyle: Faces with special hairstyles, such as skinhead, or wearing a hat, may fail to be detected. This is because the shape of the skin-hair pattern of such a face in the image may become quite different from the head-shape model. 8 Skin Detection using Color Statistics Wang et al. [11] present a fast algorithm that automatically detects face regions in MPEG compressed video. •Bayesian minimum rule is used to classify skin or nonskin class. •Classification is performed in YCbCr color model. 9 Skin Detection using Color Statistics (cont.) Bayesian decision rule Minimum cost decision rule 10 Skin Detection using Color Statistics (cont.) where 11 Skin Detection using Color Statistics (cont.) This algorithm can detects 84 of 91 faces (92%), including faces of different sizes, frontal and side-view faces. Detected face regions are marked by white rectangular frames overlaid on the original video frames. There are eight false alarms in this experiment. The algorithm is restricted in several aspects. •It can only be applied to color images and videos, because of the use of chrominance information. •The smallest faces that are detectable by this algorithm are about 48x48 pixels (3x3 macroblocks) 12 Proposed Skin Detection In this paper, we •propose a method for color model selection using bayesian estimation. •present an algorithm for color model combination using fuzzy concept. •create 1-D and 2-D histograms for skin pixel classification. 13 Proposed Skin Detection probability of skin or nonskin given C1 and C2 where are chrominance components. denotes the skin and nonskin classes. be the 2-D histogram of skin and nonskin areas. 14 Proposed Skin Detection represents the probability of given class is the a priori probability of class 15 Proposed Skin Detection (cont.) Maximum a posteriori (MAP) A decision function for selecting the chrominance be the selected chrominance components. 16 Proposed Skin Detection (cont.) 17 Proposed Skin Detection (cont.) membership function 18 Proposed Skin Detection (cont.) Skin detection function where is the skin color likeness function. are the minimum and maximum values of the range of human skin in selected chrominance components. is the range from a to b and from c to d. 19 Proposed Skin Detection (cont.) is the weighting coefficient associated with chrominance component is the probability generated by 1-D histogram. 20 Results Skin detection under varying illuminations 21 Results Skin detection under different races 22 Results Original Image Our proposed method 23 Results HSV [6] YCbCr [3] 24 Results Original Image Our proposed method 25 Results HSV [6] YCbCr [3] 26 Results Original Image Our proposed method 27 Results HSV [6] YCbCr [3] 28 Results Original Image Our proposed method 29 Results HSV [6] YCbCr [3] 30 Conclusions The skin and nonskin probabilities are created from 1-D and 2-D histograms. Bayesian estimation is used to select appropriate well-known color models. Skin detection is performed by fuzzy membership function and normalized by 1-D histogram. The method is proposed for robust skin detection under varying illuminations and different races. 31 32