In terms of human and machine vision it is often assumed that greyscale is simply the weighted sum of three colour responses. Thinking of greyscale (or luminance) in this way occasionally causes practical problems: details in pictures or graphics can be lost in greyscale reproductions if two different colours share the same weighted response. An alternative way to envisage the greyscale is as brightness encoding, whereby rather than using the luminance value at each pixel, the greyscale image represents the colour-contrast of the colour image as luminance-contrast. Socolinsky and Wolff (Socolinsky & Wolff 2002, IEEE Trans. Im. Proc. 11(8):923–931) have proposed an algorithm that achieves this goal. The algorithm consists of two stages: computing a gradient field using DiZenzo's structure tensor (DiZenzo 1986, Comp.Vis., Graph, and Im. Proc. 33(1):116 – 125), and then reintegrating the gradient field to produce a greyscale image whose local luminance-contrast reflects the colour-contrast of the original image. This approach has two problems: firstly, there is no guarantee that it is possible to preserve colour contrasts perfectly in a greyscale reproduction; secondly, the range of greyscale values required to maintain colour-contrasts may exceed the range displayable on a given display device. In this work we mitigate the second problem using a tone reproduction curve, which maps the grey-level histogram of the reintegrated image towards that of the luminance image. We will show images and numerical measures that will demonstrate the advantages of this new method and discuss the future implications for both image processing (e.g. digital photocopying) and displaying accurate greyscale reproductions of colour images.
This work is supported by joint EPSRC grants number EP/E012248/1 and EP/E12159/1.