i202 Fall 2010 School of Information, UC Berkeley

 The sudden increase in Image proliferation while exciting comes with a lot of obvious problems. As mentioned in the previous blog and comments, New Mobile Photo Sharing Applications Flood the Market, how will we efficiently search for specific parts of an image, different versions of the same image and digitally altered images when there are just so many available everywhere.  After several debates and discussions about tagging and metadata, we can at least accept that while they are effective, as information in general and images in particular grow exponentially through the market driven social media popularity, we will have to find more automatic and efficient ways to organise and search for our images.

Content based image retrieval(CBIR) has been around for a while. Instead of the traditional metadata driven image search techniques a lot of research is being done on a more content based or “pixel” based search. One such company is Idee Labs and I found one of their products particularly interesting. http://www.tineye.com/ allows you to perform a reverse image search. So instead of entering your keywords for searching for images, you can upload an image. The search engine crawls through the web and returns all the occurrences of that image on the web, including digitally modified photos, it even allows you to sort the results based on how much the image has been changed! Try uploading a photo of Monalisa and you will be amazed by the results! Also, a great way to catch people who violate copyrights.

This does raise several points though. How accurate can this really be? How good is the recall and precision? Can this really substitute metadata or would it be a good compliment to it?