A complete unsupervised learning of mixture models for texture image segmentation

Mostly, in image segmentation, we do not know the prior knowledge of the number of classes, while many clustering approaches need this prior knowledge. This fact makes the segmentation more difficult. In this paper, we introduce a complete unsupervised approach based on Gaussian mixture models, namely complete unsupervised learning of mixture models (LMM) for image segmentation. Firstly, a new feature extraction method, combining the texture features from the gray-level co-occurrence matrix with the textural information yielded through the undecimated wavelet decomposition, is used to efficiently represent the textural information in images. Then LMM is introduced for image segmentation, which can determine the number of classes automatically. Segmentation results on synthetic texture images and real image demonstrate the effectiveness of the introduced method.