Using Local Binary Pattern operators for colour constant image indexing

Abstract

The Local Binary Pattern (LBP) operator computes a local texture measure that is invariant to monotonic transformations of the image grey-scale. As a result of this property, calculating the LBP value in each channel of a colour image results in a triplet of values that are invariant to changes in illumination colour. Previous research has shown that histograms of grey-scale LBP values, and histograms of LBP values calculated in the R, G and B channels independently, form useful feature vectors for image retrieval. In this paper we generate 3D histograms of LBP values for colour images. Our experiments demonstrate that image retrieval performance, on a database of objects viewed under different illuminants, is greatly improved when using the 3D histogram compared to when the histogram is calculated for each colour channel independently. Furthermore, we find that this method can give better performance than using colour-histogram based features.