Determining effective colour components for skin detection using a clustered neural network

Object detection problems-skin detection here can be considered as object recognition problems with two classes. In this paper, each given class is clustered using the Kmeans algorithm into multiple subclasses and a Multilayer perceptron (MLP) neural network (NN) is trained for each clusters separately. In the testing phase, each point is compared with centers of clusters and the network related to closest center is selected for each new cluster. Besides the system performance improvement, it also can significantly reduce the testing time. Then the Utans algorithm as a trained NNs-based feature selection method is applied to 44 color components of 15 different color spaces. The obtained results show that the presented algorithm compare to other algorithms has higher performance and less execution time as well.