Automatic Feature Construction and a Simple Rule Induction Algorithm for Skin Detection

Many vision systems use skin detection as a principal component. Skin detection algorithms, normally evaluate a single and thus limited color model, such as HSV, Y CrCb, YUV, RGB, normalized RGB, etc. Their limited performance, however, suggests that they are looking at the incorrect color models. This paper describes a new constructive induction algorithm that creates adequate attributes for skin detection. The algorithm starts with normalized RGB and is able to produce a single rule with a small number of easy to evaluate terms with superior performance than existing methods. The constructive induction algorithm uses a new restricted covering algorithm, called RCA, as its learning component, which produces a single rule with competitive results when compared against C4.5.