High-order Markov-Gibbs Random Field Models for Texture Recognition

Abstract

Texture analysis plays a fundamental role in computer vision applications such as image understanding, object recognition, scene segmentation and so on. This thesis investigates novel types of high-order Markov-Gibbs random field (MGRF) models of images in application to texture recognition and retrieval. The goal is to discriminate a texture class represented by a single training (query) sample from other classes. Frequent in practice spatially variant perceptive (contrast/o set) deviations that preserve the image appearance hinder the recognition by signal co-occurrence statistics. Contrast/offset-invariant descriptors of ordinal signal relations, such as local binary or ternary patterns (LBP/LTPs), are efficient and popular means to overcome such drawback. The thesis explores the use of ordinal relations instead of more conventional signal co-occurrences or responses of filter banks used in today's MGRFs. Textured images are considered as samples from high-order ordinal models, which allow for learning, rather than prescribing characteristic shapes, sizes, and numbers of these patterns for texture recognition and retrieval. Approximate analytical estimates of the model parameters guide the selection of characteristic LBP/LTPs of a given order. The higher-order patterns are learned on the basis of the already found lower-order ones. The thesis introduces first the MGRFs based on complete ordinal relations of the signals. Because the cardinality of their set grows quickly, only up to the 4th{5th-order ordinal MGRFs remain feasible. Partial ordinal LBP/LTP-based relations have less limitations, so that higher-order models can be learned (up to 12th-order in this thesis) for representing textures. Comparative experiments on six databases confirmed that classi ers using multiple learned LTPs from the 8th{12th-order consistently outperform more conventional ones with prescribed fixed-shape LBP/LTPs or other local filters. The proposed learned models are mostly good enough to describe the textures. However, heuristic rules for feasible learning of the high-order models sometimes miss characteristic short-range dependencies. Adding the nearest-neighbour circular LTPs to the learned characteristic high-order LTPs overcomes this drawback and further improves the performance over the previous models. A learnable 5th-order LTP-based MGRF was also applied as a visual appearance prior for segmenting kidney images for automated medical diagnostics. Experiments confirmed that the proposed model outperforms two more conventional counterparts both qualitatively and quantitatively.