Having spent a long time as a C++ image processing software engineer and SEO, I’ve always thought I had a good handle on what works best for image search. A background in the field provides insights into the possibilities for image classifiers, such as skeletonization, histograms and more complex image transformations.
So I thought I’d share an experiment, to put some theories to the test. Starting with an experiment with colour.
As with any test in SEO, there can be no conclusion as the rules change (Google keeps updating algorithms), and many SEO factors will interfere as they cannot be isolated adequately for statistical rigour. The following test is subject to competition from images already indexed and the inter-connectivity of the web through time. So this test is pseudo-scientific. That said, I hope it still proves to be interesting…
The SEO colour experiment involves a series of coloured images.
Pink SEO experiment images


The current results for ‘pink seo colour experiment’ are:

Green SEO experiment images
There are two green SEO experiment images


The current results for ‘green seo colour experiment’ searches are:

Yellow SEO experiment images


The current results for ‘yellow seo colour experiment’ searches are:

Purple SEO experiment images


The current results for ‘purple seo colour experiment’ searches are:

White SEO experiment images


The current results for ‘white seo colour experiment’ searches are:

Gray SEO experiment images


The current results for ‘gray seo colour experiment’ searches are:

So, the scene is set. A selection of big block colour images, with a ‘gray’ version of each image colour so we can see whether the filename and other descriptors can influence the decision for the classifier, or whether the information in the image histogram overrides this deliberate misdirection.
The number of indexed images is in the order of billions, so the capacity for complex image processing may be reduced. In other words, image classifiers may still be relatively simple to avoid huge processing overheads, hitting only the filter options offered in image search.
I’ll revisit this post when search results change.
A work in progress !