Dominant local binary patterns for texture classification: Labelled or unlabelled?

This paper investigates the problem of learning sets of discriminative patterns from local binary patterns (LBP). Such patterns are usually referred to as 'dominant local binary patterns' (DLBP). The strategies to obtain the dominant patterns may either keep knowledge of the patterns labels or discard it. It is the aim of this work to determine which is the best option. To this end the paper studies the effectiveness of different strategies in terms of accuracy, data compression ratio and time complexity. The results show that DLBP provides a significant compression rate with only a slight accuracy decrease with respect to LBP, and that retaining information about the patterns' labels improves the discrimination capability of DLBP. Theoretical analysis of time complexity revealed that the gain/loss provided by DLBP vs. LBP depends on the classification strategy: we show that, asymptotically, there is in principle no advantage when classification is based on computationally-cheap methods (such as nearest neighbour and nearest mean classifiers), because in this case determining the dominant patterns is computationally more expensive than classifying using the whole feature vector; by contrast, pattern selection can be beneficial with more complex classifiers such as support vector machines.