Learnable high-order MGRF models for contrast-invariant texture recognition

• A new learnable high-order ordinal Markov–Gibbs field model of textures.
• Fast analytic learning of up to 8th-order potentials and iteraction structures.
• Texture-specific local binary/ternary patterns for accurate retrieval.
• Learning shapes, sizes, and numbers of patterns for every query image.
• Learned descriptors outperform conventional hand-picked ones.

Frequent in practice spatially variant contrast/offset deviations that preserve image appearance hinder classification based on signal co-occurrence statistics. Contrast/offset-invariant descriptors of ordinal signal relations, such as local binary or ternary patterns (LBP/LTP), are popular means to overcome this drawback. This paper extends conventional LBP/LTP-based classifiers towards learning, rather than prescribing most characteristic shapes, sizes, and numbers of these patterns for semi-supervised texture classification and retrieval. The goal is to discriminate a particular texture represented by a single training or query sample from other types of textures. The proposed learning framework models images as samples from a high-order ordinal Markov–Gibbs random field (MGRF). Approximate analytical estimates of the model parameters guide selecting characteristic patterns of a given order, the higher order patterns being learned on the basis of the already found lower order ones. Comparative experiments on four texture databases confirmed that classifiers with the learned multiple LTPs from the 3rd to 8th order consistently outperform more conventional ones with the prescribed 9th-order fixed-shape LBP/LTPs or a few other filters.