Skellywag

Thursday, February 23, 2006

Enhancing Picasa

As a photographer, I have myriad photos. I need to be able to find photos quickly, by theme/tag/label, especially if I start submitting to stock photo sites. Crucially, I have to tag the photos in the first place. Every single one. *yawn*

So, what I want is a tool to suggest labels for me. Software which would inspect the images, and suggest labels.

Digression: Google Picasa, whilst sporting a dodgy installer, and over-the-top UI 'enhancements' (everything slides smoothly about, which is really patronising), is half-way to what I'd want. However, labels must be added manually. Boring!

How about using Google Images to pull out keywords from existing photos which could be applied, in reverse, to new photos? The complexity of the algorithms needed to suggest labels is thus reduced to simple image pattern matching, or feature extraction - non-trivial, but already done a thousand times. Then use the wealth of data already present in Google Images to suggest labels for new images, based on exising images and their associated search terms.

Given that Picassa runs on the desktop, we can use the high performance of the PC/Mac to do the image comparison (no huge server farm needed) - all we need is a way to hook into Google Images, pull down similar photos, and "reverse" the Google Image results: outputting text (labels) for a given image).

This might imply some standard means of quantifying elements/aspects of the image: colour, shapes, etc., and sending that data to a central server, in order to get the match. Alternately, Picasa could send a reduced-size image to the server, and try find matches based on that reduced size. I suspect for many images, you'd get helpful suggestions even based on thumbnails, but I have no way of justifying that ;-)

Shame there is no Images API yet.


0 Comments:

Post a Comment

<< Home