Thursday, July 11, 2013

Language and Color

In a previous project I wondered if maybe a whipped up mix of the images that Google searches show me could be used as a cheap first-pass for testing color theory. Google search results are sort of a zeitgeist for any given term -and an image search is a portrait of that.

Because Google has language/culture variants for all over the world, I thought it would be a fun project to see how the zeitgeist-y image portraits might compare to each other across different cultures. This notion was in no small way inspired by the terrific and more rigorous work of David McCandless.

The result is an array of hues tracking various concepts (design, art, music, math, science, and philosophy) through five different languages.  You can pick a language and track across the six terms, or you can pick a term and compare it's coloration down the five languages.  Or whatever.

It's worth pointing out that almost all of these quilts ("quilts" are the snapshotted images of a Google Image search results page, coined by ET) have a multi-modal distribution of color.  That is, they may have a big bump around orange (due to the frequency of humans as the subject of so many images) and a big bump around blue -and the aggregation of those colors could result in green, which is sort-of nonsense.  So, temper your interpretation of the fully aggregated color palates with the little color histograms to their right.

Maybe a more meaningful illustration of cultural associations of colors to terms would be to aggregate the average of each term's colors across all countries then calculate where along the spectrum specific cultures tend to deviate from that "global" average. Yeah, that would rule and somebody else should totally do that.

Anyway, have fun looking at this and thanks for stopping by.  Below are various elements of this graphic if you'd rather look at it that way, with some wildly anecdotal commentary...

Warm red-shift tones abound. Fleshy portraits were common in the Chinese quilt, Hindi favored statuary or groups of people, Arabic images favored sculptures -frequently scenes of carved sand, Russian and English images were commonly bright abstract paintings.

Only the Chinese quilt shows a big departure from the crowd. Orange tones so conspicuous in just about all other tiles are way reduced for the Chinese term for design -frequently comprised of clean digital layouts (this could be an artifact of the translated term). Hindi is predominantly portraits.  Arabic is predominantly warm landscapes, often with poetry or word art. The Russian quilt is almost entirely interior design images (certainly a less-than-perfect translation choice on my part).  English is predominantly light and colorful desktop background image stock.

Another departure for the Chinese language quilt. The much greater proportion of blue in the images has pushed the overall average over to the cooler green.

All languages were fond of tagging chalkboard pictures but English image-taggers especially so.

Science scored the biggest counts for cool blue colors because of the common presence of dark, blue, illustrations of microscopic or cosmic scenes.  Except from Hindi and Arabic languages.  Hindi images tended to show teams of scientists at work, explaining the regression to flesh tones. The Arabic images tended to be more frequently clipart illustrations. Maybe because of the variability in the translated terms?

Overwhelmingly earthen in tones.  Chinese and Arabic images tended to be snapshots of texts on aged paper, Hindi images were almost entirely portraits of people or deities, Russian images tended to be diagrams and illustrations (in earth tones), while English images were full of statues.


  1. I found this post very exciting. I think you will have any other post on this topic? I am also sending it to my friend to enjoy your working style. Cheers!

  2. My only concern would be the tendency towards brown colours. Not sure how this works out with additive colours on the digital screen but traditionally mixing lots of colours together produces a muddy hue. I would be curious to see the result of compensating for the brown by reducing the red channel. Ultimately I feel that this is not a good realisation of the search term as Google inaccurately brings in off-concept results through unrelated imagery, but this is a really interesting concept.

    1. Thanks for your thoughts, Alex! I had the same concern about the tendency of averaging colors to merge into brown, especially multi-modal examples, which you can see above. I suppose a non-linear method of aggregating would be interesting. This is the reason I included the spectral histograms, in the event that a reader might pick out A) if two colors were dominant but merged to brown, or B) know how seriously to consider the resulting color given the histogram.
      Anyway, thanks again for the input!