next up previous contents
Next: Colour Segmentation Up: The Tina Colour Tool Previous: Colour Space Conversions   Contents

RGB Normalisation

The RGB normalisation button provides a simple colour histogram equalisation method. It scales each of the currently loaded red, green and blue fields so that their pixel values lie between 0 and 1, by dividing each pixel by the sum of the red, green and blue values. This can be useful for display purposes, as it will equalise the contrast in the colour fields and so bring out detail in dark regions. This function replaces the red, green and blue fields currently loaded into memory with their normalised versions.

It is an oft-repeated fallacy that colour normalisation (or histogram equalisation in general) should be used as a precursor to segmentation. In fact, since colour normalisation clearly cannot increase the information content of the original image, it should have no effect on the segmentation result if the segmentation algorithm is taking proper account of the noise on the data. If you see a paper or presentation which states that colour normalisation does affect the result of their segmentation, it means that their algorithm has an implicit assumption of scale (probably a blurring kernel of fixed size somewhere in the code). Stretching the colour space by normalisation therefore changes the scale of the space relative to the assumed scale, and so changes the segmentation result.

The Tina colour segmentation algorithm takes account of the noise on the data, by scaling all measurements in the colour space by the local noise. Therefore, applying colour normalisation prior to segmentation will have no effect on the segmentation result to first order. Some small second order effects may be observed, as the noise scaling is only performed locally, but for most natural images these should not change the number of labelled regions by more than about 10%.


next up previous contents
Next: Colour Segmentation Up: The Tina Colour Tool Previous: Colour Space Conversions   Contents
root 2017-11-23