Skin color Space Model

Skin color is proved to be an effective and robust face detection, positioning and tracking basis. At the same time, skin color detection can also be used in image content filtering, Content-aware video compression, Image color balance application and so on.
The feature-based face detection method is used to detect the skin color as the basis for the detection is very practical. Color processing allows for fast processing and very robust geometric transformations of face patterns. Experience has shown that facial skin has a characteristic color (easily recognizable). Using color as a feature of a face requires overcoming three major problems: 1. What color space to choose; 2. How to model the skin color distribution; 3. How to deal with the result of face color segmentation

-1. Color Space Description:

1.1.RGB Space:
RGB space is derived from the use of CRT display, with three basic colors (red, green and blue) light composition color, this space in the image storage and processing convenient application of the most extensive. However, the color-based image recognition and color analysis are very unsuitable due to the fact that the colors of different color channels have correlations, perceptual inconsistencies, and chroma and luminance blending.
1.2 Normaliz Ed RGB
By using color normalization, you can reduce storage for B-channels.

1.3 HSI, HSL, HSV (hue-saturation-brightness):
The space describes the color of the sensory values that are usually determined by the experts. Hue describes the main color, s represents saturation, and describes the richness of the color of the area proportional to the luminance. v indicates the color brightness. The apparent difference between luminance and chroma and the intuitive nature of each component make the space very popular in color segmentation. The disadvantage is the non-continuity of hue and the computational complexity of luminance.

1.4 YCRCB
The space is a nonlinear transformation of the RGB space component, and the color is subtracted by the R component and the B component of the luminance Y,rgb to obtain the other two components CR and CB.

The separation of luminance and chroma components makes the space very attractive for skin-tone models.
1.5 Perceptual Unified Color System
Skin color is not the physical property of an object, but is a human sensory characteristic. So the closer the color to the human sensory system, the better the performance. Cielab and Cieluv are two color spaces that perceive consistency. Perceptual consistency means that the distance between the two colors represented on the chromaticity graph is inconsistent with the perceived change of the color viewer, that is, the difference in numbers between colors is inconsistent with visual perception. But this brings to the time complexity of converting from RGB space to lab space.

-2.skin Modelling

The ultimate goal of skin detection is to establish a decision rule that distinguishes between skin and non-skin pixels. The distance from the color of the measured pixel to the skin tone is usually introduced as a criterion, defined by the skin tone establishment.
2.1 Explicitly defined skin areas
One way to build a skin-tone classifier is to display the skin tone boundaries defined in the color space:

When the lighting conditions are ideal:


However, if the illumination is uneven (affected by the shadow) The result of skin color segmentation may not be ideal (the paper mentions that for objects similar to skin color, such as wood, copper coloring of metal, there will be false error Positive (false positive)):


It can be seen that wood flooring, similar to skin tones, is also incorrectly detected as a skin tone area:



The obvious advantages of this method are very simple, the difficulty is to find the appropriate color space and the corresponding decision rules. [Gomez and Morales 2002] proposed a machine learning algorithm to find the appropriate color space and decision rules. First, using normalized RGB space, a constructive inference algorithm is used to set up a large number of three attributes that are superimposed by three r\g\b components and constant 1/3. This method proves that the performance exceeds the Bayes skin probability map method in the same color space.
2.2 Nonparametric Skin Distribution modelling (non-parametric skin tone distribution model)
The key idea of nonparametric method is to estimate the skin tone distribution from the training concentration.
2.2.1 Normalized lookup table (LUT)
The color space is quantified as a large number of bins, and each bins represents a specific range of color component pairs. Bins constitute a two-or three-dimensional histogram, and each bins counts the number of times the corresponding color pair appears in the training set. After the training is completed, the histogram is normalized, and the histogram value is transformed into a discrete probability distribution.

The skin [C] represents the value of the current histogram bins, which corresponds to the color vector, and norm is the normalization factor. Histogram 32 o'clock basic performance is best
2.2.2 Bayes Classifier
A more appropriate method is to calculate the conditional probability, that is, the probability of the skin in the case of a given color. The Bayesian formula is expressed as follows:

When it is only the comparison and the case, the ratio of the two gives a verdict condition:

"Statistical Color Models with application to Skin Detection" in the article Theta take is 0.4, the histogram series of 32 to achieve the best detection effect. The advantage of a lookup table is that it is fast, but consumes more memory because you want to preserve the histogram of skins and non-skin. (not finished--)

Skin color Space Model