On selecting colour components for skin detection Giovani Gomez Dept. of Computing, ITESM-Morelos AP 99-C, 62050, Morelos. M´ exico gegomez@campus.mor.itesm.mx Abstract We used a data analysis approach for selecting colour components for skin detection. The criterion for this selection was to achieve a reasonable degree of generalisation and recognition, where skin points exhibit a well defined cluster. After evaluating each component of several colour models, we found that a mixure of components can cope well with such require- ments. We list the top components, and from these we select one colour space: H-GY-Wr (Wr [15]). A nearly convex area of this space contains 97% of all skin points, whilst it encompass 5.16% of false posi- tives. Even simple rules over this well-shaped space can achieve a high recognition rate and low overlap to non-skin points. This is a data analysis approach that will help to many skin detection systems. 1 Introduction Skin colour detection is a very important step in many vision systems, like gesture recognition, hand tracking, video indexing, region of interest, face de- tection, etc. (see e.g. [2, 4, 6, 9, 11, 12, 13, 14, 15] to name just a few). Pixel based skin detection can nar- row the search space prior to high-level layers, how- ever this is not an easy task. Skin pixels can vary with ambient light, such as colour lamps acting as fil- ters, brightness and specularities, shadows, daylight, etc. Since different cameras return different values for the same scene, pixel-based skin detection becomes a cumbersome task. There has been a growing interest in using proba- bilistic methods for skin detection. One widely-used choice is the Skin Probability Map, or SPM for short [1, 2, 8, 17], which has been assesed [1] as the best one in terms of accuracy and efficiency. Despite of the long history of skin detection, there are no works addressing which colour components have more discriminatory power. Another few works [1, 17, 16] have surveyed single and thus limited colour models for skin detection. Due to this lack of infor- mation, most vision systems just work on HSV, raw RGB, or normalised RGB. This paper explores individual contributions of colour components, and it shows that a hybrid space (i.e. a mixure of components) can achieve a better overall recognition rate compared to any single model. By comparing this hybrid model to a standard SPM in RGB, we shall show a substantial difference between both overall recognition rates. 2 Skin Probability Maps A popular version of a SPM [1, 2, 8, 17] is a lookup table where RGB values directly address a voting slot. Two 3D histograms are computed, one for skin and one for non-skin. After dividing every slot by the total count of elements, we get an as- sociated probability on a [R, G, B] index. Statistics are derived from these histograms. Roughly, the con- ditional probability of a pixel with RGB values to be skin or non-skin is: P (rgb|skin)= Hist skin [r,g,b] T otal skin , and P (rgb|∼skin)= Hist non-skin [r,g,b] T otal non-skin . A new unseen pixel is labeled as skin if it satisfies a given threshold, P (rgb|skin) P (rgb|∼skin) ≥ θ, where θ is obtained empirically. The recognition ratio is a tradeoff between reducing false positives, and increasing skin classification. The idea behind SPMs seems reasonable given a large amount of training data. Nevertheless, due to the sparse distribution of skin points in RGB space (and memory requirements), one usually reduces the cube’s size. This step also helps to “generalise” and compact the histogram. Such generalisation in the his- togram model is guided by the number of bins. Thus, a SPM also introduces another parameter to fit. This extra parameter does not contribute to the idea of having a clear cluster where unseen points will fall. Of course, we can approximate any distribution with this pragmatic approach. Although it has been argued that skin points have a significant degree of dis- crimination in RGB space, and SPMs work reasonably