抄録
Detecting skin regions of an image is one of under interest arcas in computer vision and graphics. It can be primary step in several applications like advanced human-computer interaction, biometric authentication of contextual image retrieval. Several different researches and challenges have been done to classify the image regions into two groups of skin and non-skin areas. Most of them discussed well in theoretical aspects like effect of different color spaces or classifiers on large datasets. Here it is tried to have practical look to the problem. In a real final system we will have some limitations on speed, memory and even training dataset in different conditions. After a short overview to some common previous methods, a novel skin model is proposed based on LVQ neural networks. Consequently a study on homogenization methods gives more accurate choice for this step. It is shown with the experiments that this scheme can reserve advantages of several different methods at once. It can perform comparable accuracy of 87.92% with acceptable speed while covering wide variety of skin color with a limited size of training data.
