Research On Skin Color Segmentation Algorithm Models Based On Generalized Gaussian Distribution

Image processing and computer vision is widely used in aerospace, military, medical, production and living, et al. Image segmentation not is the prerequisite and foundation of image recognition and image understanding and but the basic key technique of image processing and computer vision. Recently, skin color segmentation based on color information has gained much attention. Skin detection plays an important role in a wide field on image processing such as face detection and tracking, gesture analysis, content-based image retrieval and human computer interaction. But still there are lots of problems need to be solved. Hence, skin detection and segmentation researches are essential in both theory aspect and practical aspect.In fact, the human skin colors of different individuals from all different races cluster in a small region in color space provided that the images are taken under illumination controlled environments. Naturally, many statistical models were designed to describe skin color cluster in color space.Skin color detection based on probability distribution is one of effective and efficient skin segmentation method. Many of the representative works on skin-color distribution modeling have used Gaussian distribution or Gaussian mixtures. The advantage of these parametric models is that they can generalize well with less training data and also have very less storage requirements. The advantage of these parametric models is that they can generalize well with less training data and also have very less storage requirements. Some researches, however, show that skin color cluster is not well enough approximated by the single Gaussian model because of asymmetry of the skin cluster with respect to its density peak. Usage of the symmetric Gaussian model can lead to high false positives rate. Furthermore, Gaussian mixtures assumption often leads to biased estimates of the statistical parameters and its accuracy is reduced when the region classes overlap significantly. In order to cope with these drawbacks from Gaussian or Gaussian mixtures assumption, in this paper, we use generalized Gaussian mixture distribution to estimate the probability of skin color character information. Generalized Gaussian distribution is a distribution family with wider range than Gaussian and will be more flexible and convenience. The generalized Gaussian model is built after defining and selecting a color space. The parameters of the generalized Gaussian distribution can be estimated by using moment estimation or maximum likelihood estimation in training dataset. Expectation and maximization (EM) algorithm is introduced to deal with the difficulty in solving the parameters of the mixture generalized Gaussian distribution. And some numerical computation methods such as Newton-Raphson algorithm are used to estimate the roots of the non-linear equation from EM algorithm.