Illuminant Invariant Chromaticity Distribution Coding Using the Fourier Power Spectrum

We have previously shown that log-opponent chromaticity distribution histograms provide an effective and perceptually relevant method of coding the colour content of an image. In this space, constant object or scene colours are separated from the variable illumination colour. Variation in illuminant colour, known to cause indexing to fail, is modelled as a translation within the log-opponent chromaticity space coordinate. In effect, the distribution shape remains constant, but translates within the coordinate system as illumination colour varies. We show that the Fourier Power Spectrum (FPS) captures the distribution shape in the frequency domain, and provides accurate illuminant invariant image indexing performance. Furthermore, the FPS can be effectively compressed down to a very small (8 value) match feature vector without affecting indexing accuracy. Results for the FPS method show some improvement over those from our earlier work on mean-centred log-opponent chromaticity distribution histograms.