Google’s fuzzy logic for image search gets clearer

Google continues to make forays into the world of images, on several different fronts.

Image recognition in Google searches is taking shape. Although not highly publicized yet, there is a way to search images for certain elements such as faces. For example, if you do an image search for “USA”, you get a slew of maps. If you do the search “usa imgtype=face”, the search does indeed turn up pictures of people.

According to Google’s Sergey Brin, it will be possible for the computer to soon automatically search images for patterns, such as that of an elephant, in a picture.

While identification of physical objects in pictures will go a long way towards helping categorize the trillions of images on the web, there are aspects of classification which are abstract, and subjective. For example, the picture on the left can be categorized as 1. sky, 2. bird or 3. soaring.

A new Google project called Google Image Labeler has popped up for improving the relevance of image search. At any particular time, two random users who have signed up for the experiment, are paired up. Over a period of 2 minutes, both are shown a set of images for which each provides as many labels as they can think of. When the pair has a match, each of them get a certain number of points depending upon how specific the description is.

I tried out the application in guest mode (see the pictures below). Interestingly, each picture starts out with a list of “off-limit” terms which are actually the first ones that would come to mind. The purpose of these might be to look beyond the most obvious categorization. For example, in the map example below, the descriptions “map” or “Australia” were off limits.

While the Image labeler seems like a pretty entertaining project at this point, Google’s objective is to probably use the real human fuzzy network data to refine its automatic image filtering, rather than a simple grand scheme to use free cpus to classify all of the web’s pictures.

Blogged with Flock