Preferred greyscale versions of coloured images: Human vs machine

Abstract

Each year millions of greyscale reproductions of colour images are made. The majority of these are produced by removing the chromatic information, which leaves a greyscale made by just the achromatic colour variable. One problem with this approach is how to make a greyscales for images that contain equilumiant edges and/or borders in a way that preserves the image content In our experiment we evaluated some of the more recent algorithmic attempts to tackle the colour-to-greyscale problem, and compared the performance of these methods against greyscale images created manually by human observers. We used an image preference experiment where 10 participants carried out pair-wise comparisons of greyscales produced by 5 different mathematical algorithms and 1 human. Overall, we employed 10 different test images with varying amounts of equiluminant detail, and included images with different levels of content complexity (ranging from outdoor scenes to pie charts). We find that two algorithms significantly outperform the other computational techniques (including luminance), and that these algorithms both attempt to preserve local colour contrast in the greyscale algorithms. Furthermore, we find that neither of these techniques is significantly preferred over human greyscales. Finally, the interactions between images and algorithm are strong, indicating that image content is important in deciding which is the best greyscale version. These results support the motivation for this research area: there are better ways to convert colour to greyscale than simply using luminance and that images created by human observers are comparable to those produced by algorithms.