Color Space to Detect Skin Image : The Procedure and Implication

Skin detection is one of the processes to detect the presence of pornographic elements in an image. The most suitable feature for skin detection is the color feature. To be able to represent the skin color properly, it is needed to be processed in the appropriate color space. This study examines some color spaces to determine the most appropriate color space in detecting skin color. The color spaces in this case are RGB, HSV, HSL, YIQ, YUV, YCbCr, YPbPr, YDbDr, CIE XYZ, CIE L*a*b*, CIE L*u* v*, and CIE L*ch. Based on the test results using 400 image data consisting of 200 skin images and 200 non-skin images, it is obtained that the most appropriate color space to detect the color is CIE L*u*v*.


INTRODUCTION
Skin is a part of human body.In an image, skin is able to present the existence of human.Furthermore, it can also be determined how many the skin is exposed in an image.The more skin is exposed, the more naked the human object in the image.In further process, it may indicate the presence of pornographic elements in the image.
In skin detection process, the most suitable feature is the color feature, if compared to others.Talking about color feature, it is exactly related to the color space and its descriptors.Color feature extraction in the appropriate color space will impact on the accuracy of the detection [1].Take the research done by Vasthi dan Kusumaningrum as an example,stating that the color space for fruits object is more suitable using the OHTA color space [2].For skin object, from the research by Isa dan Mariana, it is stated that RGB color space, commonly used for system input, has perceptually congruent characteristics, meaning that the RGB characteristic cannot depict the system sensitivity of human sight and thus RGB model is not quite suitable to detect skin color [3].On the other hand, in the research of Zhuo et.al., YCbCr color model was used to perform skin detection because the distribution of all skin colors and races is more concentrated in YCbCr color space, if compared to others [4].However, the two researches did not compare the used color space with others.Hence, this research will examine the most appropriate color space for skin object by comparing some color spaces available.
Descriptors for color spaces are, such as pixel, color histogram, and color moment.If the descriptor is a value of pixel, then the dimensionality will be valued high.Meanwhile for histogram and color moment, a research by Stricker dan Orengo has proven that color moment yields better result than the color histogram does [5].Hence, this research will compare some color spaces with color moment as the descriptor.

METHODS
The next stages conducted in this research can be seen in process blocks in Figure 1.To do the conversion from RGB color space to color spaces above, the following formula is used: V (Value) calculation: 2) Conversion from RGB to HSL To convert RGB to HSL (Hue, Saturation, Lightness), equation ( 1) to equation ( 6) are still relevant.Hue calculation uses equation (7), while saturation and lightness calculation are as follows: S (Saturation) calculation: L (Lightness) calculation: 3) Conversion from RGB to YIQ The conversion from RGB to YIQ uses the formula as follows

d. Color Moment calculation
Image that has been resized to 64 x 64 is then being taken the sub-image in 5 x 5.
From the sub-image, the calculation of color moment is later performed in subimage pixel i.e. mean, standard deviation andskewness (gradient).Later, this subimage moves overlapping at each line and column.
e. Classification Classification is performed by using the Support Vector Machine (SVM) method, with its Kernel is Radial Basis Function (RBF) [11].
f. Evaluation Evaluation is performed using 10-fold cross validationto determine the accuracy of each color space.

Result
The data used in this research are 400 images consisting of 200 images of human skin (100 naked man images, 100 naked woman images) dan 200 non-skin images(100 landscape images such as house, animal images, etc., and 100 images of man wearing skin color-like clothes).The skin images in this case represent all races existing on earth, i.e.Asiatic Mongoloid, Malayan Mongoloid, American Mongoloid, African Negroid, Negrito, Melanesian, Nordic, Alpine, Mediterranean,and Indic.Each race is being represented by 10 images for each sex.
The test uses K-fold cross validation with k = 10.400 data are divided into 10 fold, thus the amount of training data are 360 and testing data are 40, consisting of 10 man's skin images, 10 woman's skin images, 10 landscape images, 10 images of man wearing clothes.Table 1 shows the experiment result.Based on Table 1, it can be seen that the color space yields the highest accuracy value in skin detection is CIE L*u*v*.

Discussion
High accuracy performed by CIE L*u*v of 92,5% shows that skin color can be represented better in CIE L*u*v color space.Although some are acting as the non-skin images taken from images with akin colors of that of skins, the images between skin and non-skin show no difference in CIE L*u*v color space.
On the other hand, the highest accuracy yielded without CIE is that of YCbCr of91, 75%.It shows that the distribution of all races and skin colors is more concentrated in YCbCr space, if compared to other spaces other than those of CIE.Hence, for further development of skin detection, the suggested color space is CIE L*u*vorYCbCr.

CONCLUSION
Detecting skin colors, the color space able to produce the highest accuracy rate is CIE L*u*v of 92,5%.If it is not in CIE color space, the highest accuracy rate is 91,75% of YCbCr color space.

Figure
Figure 1.Process blocks of research plot alur penelitian