Evaluation of Automatic Image Color Theme Extraction Methods

Determining key colors from a design perspective using dE‐means color clustering

“…In literature, various types of algorithms to extract a color palette from an image are described. In general, we can distinguish between three types of algorithms: Histogram-based methods, for example, 7,8 clustering-based methods, for example, [9][10][11][12] and supervised learning-based methods, for example. 13 The simplest type of method is histogram-based methods.…”

“…They are very simple and come at low computational costs but will typically result in color palettes that are too coarse and lack subtlety. 11 Slightly more complex methods are clustering-based methods, for example. [9][10][11][12] The mean-shift method 10 is a method that builds upon the concept of kernel density estimation.…”

“…11 Slightly more complex methods are clustering-based methods, for example. [9][10][11][12] The mean-shift method 10 is a method that builds upon the concept of kernel density estimation. As such, it searches for local maxima in the distribution of pixel values and returns the pixels that represent the centers of the densest regions.…”

“…Even though some of the palettes are quite similar, for example, runs 1, 3, To overcome this issue, we can eliminate the randomness from the initialization so that the same initial centroids are used for every run. Ciocca et al 11 propose a method to fix the initialization by sampling the initial centroids uniformly on a line running from the darkest pixel to the brightest pixel. This will ensure that the outcome of different runs on the same input image will always result in the same output palette.…”

“…However, because all initial centroids are shades of grey, it can be expected that the colors in the final palette will be slightly biased towards neutral as well and consequently will not result in more representative color palettes. 11 Finally, the most complex type of method is supervised learning-based methods such as the method proposed by Lin and Hanrahan. 13 This method uses training data to learn a linear regression model that can extract color themes from images.…”