Colour-texture Indexing

It is often assumed that colours in an image correlate with intrinsic surface properties (e.g. reeectance). Indeed, this goes a long way to explain the predominance of colour as feature for image indexing. Unfortunately the assumption is unfounded since the image of a scene viewed under two diierent coloured lights contains two diierent sets of colours; that is, image colour is a measure of the intrinsic properties of light and surface combined. In our research we aim to provide alternative indices to colour which are invariant to illumination change, yet at the same time still correlate with intrinsic surface properties. The starting point for this work is the colour angle index developed in a previous studyy3]. This index comprises the 3 angles between the R, G, and B colour bands of an image. It is invariant to illuminant change and, importantly, captures low-order image statistics. However, because the colour angle index comprises just 3 numbers the number of images it can discriminate between must be limited. In this paper we extend the angular index by deriving a range of new angle invariants. The key observation which we make is that the colour angles of an image calculated post-linear ltering (e.g. convolution) are also illuminant invariant. Two new colour-texture angle-indices are considered in detail. The local-texture angles of an image are colour angles calculated post Laplacian of Gaussian ltering and global angles are calculated from the auto-and cross-correlation band images. Experiments demonstrate that both the local-texture global-texture angles deliver accurate image indexing. Moreover, performance is better than other colour based methods.