Pages

Saturday, July 2, 2011

Search using Images is now a Reality

(This post was written some weeks earlier but somehow I never published it and it remained in the drafts. Some references might be a bit outdated)

Finally it has happened! We can now search for an image using another image. Yes, Content Based Image Retrieval (CBIR) has finally been rolled out by Google. You can try it out on Google Image Search.

I can't express how it felt when I saw it has been released. Indeed, it was like a dream come true. A new and long awaited dimension has been added to search. No longer are we constrained by words. Indeed, a picture is worth a thousand words.
I was (and still am) fascinated with this idea. So I thought of writing about it

Since my fourth semester (around Jan 2010) I had been fascinated with this idea. At that time it was something new to me, something I had never thought about. Basically it was this - "how would it be, if we could search using images?". Now what got me into thinking this? At that time I was taking a course named Research Methods. Our instructor (Dr. PK) had created a mailing list for discussing about interesting research. One day, he posted a link. 
Picture-driven computing: New research could enable computer programming based on screen shots, not just code.
This got me thinking. If they could program using images, then why not search using images? That's when I started getting excited about this topic. I replied enthusiastically to this post and I described CBIR in brief. At that time however I didn't know much. On searching more I found out about Google Goggles and Tineye. During the summer holidays, I devoted my spare time to find out more about this topic. The ACM Chapter at IIIT Delhi releases a half yearly newsletter called Kaleidoscope. I was part of the editorial board and wrote an article on CBIR. (You can download it here)

After going through some research papers, I felt like creating a working prototype. The Next Semester (5th sem Aug-Dec 2010) I took the Image Analysis course. That's when I thought of doing this. I and my friend Vibhas took up this as our course project.

Needless to say, we did not even aim for creating a web image search engine with CBIR. Our aim was only to create a working prototype with somewhat good accuracy. The image would be stored locally in the computer. We worked on it throughout the semester (and gradually realized that CBIR is not an easy thing to implement). Every research paper we read, gave us a new idea but at the same time pointed out the problem with another. We decided to focus only on colour, shape and texture out of which only the colour based algorithms worked with a relatively good accuracy. Even then there were problems - a search with red apple turned up images of a red Ferrari.

A Screenshot of our Image Search Engine using CBIR

Even though our implementation did not use any sophisticated algorithms, it was a wonder for many. When it was showcased in the Research Showcase, lots of people were excited to try it.

Now that I understand how difficult it is to implement on a large scale, I give me heartiest congratulations to the Google Team. You guys did what I could only dream of!

No comments:

Post a Comment