Cover Colors

This is an example app, trying to mimic the iTunes cover colors algorithm.

How it works

It is a more or less direct port from Mathematica of the example code, provided by Seth Thompson.

Preparation of the image

• At first, the image is scaled down to max-width of 36px and max-height of 36 (to reduce the computation cost)
• A border of 1px around the image is removed, so that the new max. dimensions are 34x34.

The removal of the border is important, because some of the images have a fringy border (or just a normal border, too):

Found colors: , ,

Background Color

• A histogram of the 1px wide edge of the image is created and the most dominant color is selected

"Title" & "Song" Colors

• A histogram of the complete image is created, which groups all (in the YUV color space) similar colors into buckets
• These buckets are sorted downwars.
• Now, from the top down two colors are searched, which are different enough from the background color
• If no two colors are found, then the background color is analyzed, whether it is a dark color (luminance < 0.5) and white is used (black otherwise).

Possible further optimizations

This algorithm (and implementation!) is not optimal in many cases:

Simple AI additions to filter out highlights

Example: the bow tie of Aloe Blacc is not recognized:

Found colors: , ,

Looking for a better contrast between background color and text colors

Currently, the text colors are chosen, so that there is a large enough (to be defined) color difference between the colors. This can lead to potentialy low contrast, especially if you use the colors without any separation.

Example: Adele - 21

Found colors: , ,

This specific implementation

This implementation was created in just 3 hours, so there might be bugs or places for optimizations included. :-)

Sources & additional variations

• Original by Seth Thompson
• One of the many implementations of Seth Thompson's algorithm in JavaScript
• The YUV color space details (esp. translation RGB -> YUV)
• An implementation in Objective-C with some differences
• An other approach using R-Trees
• Yet another approach using k-means clustering

The rights of all used cover images lie with the appropriate copyright owners.