Skin and non-skin probability approximation based on discriminative tree distribution

We investigate the probability tree models to approximate skin and non-skin distributions. These models have presented good results in solving the skin detection problem. However, there are two main disadvantages of the existing skin/non-skin tree distributions based models: (1) the structure of some tree distributions is predefined; and (2) the inter and the intra classes of skin/non-skin are not taken into account at the same time by the existing skin and/or non-skin tree models. To overcome these drawbacks, we propose a new classifier based on an image patch joint distribution approximation modelled by a discriminative skin/non-skin tree. On the Compaq database, we examine the performances of the proposed approach compared with the baseline model and two others based on dependency tree´s distributions. Experimental results show that the new approach is a significant improvement over the others.