Texture discrimination using local features and count data models

In this paper we consider the problem of textured images modeling and discrimination using local features. Local features are quantized to form a textons visual dictionary. This procedure allows the representation of each textured image by a vector of counts which represent the frequencies of the textons in the texture. Having the count vectors in hand, we introduce a new mixture model for the accurate modeling of these vectors. This mixture model is based on a composition of the Beta-Liouville distribution and the multinomial. The novel proposed model, that we call multinomial Beta-Liouville mixture, is optimized by expectation-maximization (EM) and minimum description length, and strives to achieve a high accuracy of textured image data discrimination. The developed approach is competitive with recent proposed count mixture models and its classification power is demonstrated through experimental results on various textured images data sets.