Learning generic third-order MGRF texture models

Descriptive abilities of translation-invariant Markov-Gibbs random fields (MGRF), common in texture modelling, are expected to increase if higher-order interactions, i.e. conditional dependencies between larger numbers of pixels, are taken into account. But the complexity of modelling grows as well, so that most of the recent high-order MGRFs are built to a large extent by hand. At the same time it is very difficult (if possible) to manually choose an efficient structure and strengths of pixel interactions for modelling a particular texture. This paper explores a possible extension of a computationally feasible framework for learning generic translation-invariant second-order MGRFs onto generic third-order models. Several information-theoretical heuristic methods for automatic learning of the latter are compared experimentally on a large and diverse database of realistic textures in application to a practically important problem of semi-supervised texture recognition and retrieval. However it is shown that better generative abilities of learnt models do not necessarily imply their higher discriminative power, and also the increased difficulty of learning third order order models may lead to worse generative performance.