Image Recognition Automation Fail

I had a friend who asked me to take a quick look at a business problem he was having.  He had a set of photos (of vinyl record albums) that he was looking at to identify the artist.  After finding the artist, he’d do some additional categorization work and then push the image and metadata to a ecommerce site he is running.

He wanted a way to more quickly identify the artist and album, preferably without his intervention.

The first thing I suggested was Mechanical Turk, as this seemed like a great example of a Human Interaction Task.  However, my friend tried this and found it to be more work (mostly proofing, I think) than it was help.

He also pointed out that Google Image Search does exactly what he wants.  You can post an image or URL to it (you have to use the camera in the search box), and it will give you back like images, best guess for the image matching, and links related to the image.  Pretty cool!

However, there is no API for the new Google Image Search.  Sure, you have the old, deprecated Images API, but it doesn’t have access to post an image or a URL, just keywords.  A bit of looking around revealed this StackOverflow post, which deconstructs the new Image Search parameters.  However, the deconstructed URL gives back javascript which needs to be executed by a browser.  I looked briefly at Selenium and webdriver to do that, but couldn’t figure it out.

I also looked at Kooaba’s API, but they didn’t get back to me when I signed up for a free developer account, and their API only covers books, CDs and DVDs.  Also from a StackOverflow post, I looked at MacroGlossa and IQEngines.  Neither of them seemed to work–MacroGlossa wanted a category (and, shocker, vinyl record albums was not a category) and IQEngines let me submit an image wasn’t successful in identifying it.

I had to admit I was defeated.