Improved texture retrieval by combining different variants of local binary patterns

This paper proposes the combination of different variants of local binary patterns for texture retrieval. Recent studies have shown that LBP features extracted at multiple resolutions give the best performance for a standard texture retrieval collection. Techniques have been proposed to create a multi-dimensional histogram of these features. In this study we use a simpler approach. We hypothesize that the different variants of LBP may actually extract slightly different but useful image information. If this were the case, then using them in combination will result in further improved performance, compared to using each one independently. We demonstrate this by using LBP, LBP^ri, and LBP^u2 features in combination. We take ideas from the GNU Image Finding Tool (GIFT), and use separate normalisation to ensure the different feature vector lengths do not bias the system towards one of the features. We performed experiments on the two standard texture collections. Our results demonstrate that combining the different LBP features do indeed result in improved performance. In fact, for one of the collections (Outex TR 00000) collections, using our method give better performance than the current state-of-the-art.