Behold is a search engine for high-quality Flickr images. It aims to answer your queries based on what is inside the images — at the pixel level.

It offers a completely new way to search for images, using techniques of computer vision. It is different to standard image search engines, such as Flickr or Google, because those search through images using only image tags and filenames.

Behold looks for high quality images, so you don’t have to sift through hundreds of poorly taken pictures to find a good one. Behold uses both aesthetic and technical quality indicators to find some of the best images available online.

Behold draws computational power from Amazon Elastic Compute Cloud (EC2) to handle large volumes of images.

Behold is capable of recognizing a number of visual concepts in pictures. You can ask Behold to return images that look like one of these concepts. This new type of search can be flexibly combined with regular text-based search. For example you can ask Behold to return images tagged with the word ‘London’ that look like pictures of buildings (try it!). You can also filter text-based image search results based on what the images actually look like. Both of these features are demonstrated in these videos.

With a newly introduced feature, Behold goes one step further and automatically suggests visual filters after analyzing the words in your query. It shows you what your search results would look like if you apply one of these filters, so you save time on finding the right one. This feature is demonstrated in these two videos

Behold’s technology is in the early stages of its development. We expect significant improvements in Behold’s visual image search quality over the coming weeks and months. We are currently working on three fronts:

* teaching Behold to recognize more concepts in images
* improving the quality of concept recognition

* searching larger volumes of images

Behold’s visual concept search is based on the technique called “automated image annotation”. This technique calculates probabilities of concepts being relevant to images based on the pixel content of each image. The models for calculating concept probabilities are estimated using manually annotated training images.

Behold implements the automated annotation model proposed in my PhD thesis.

Alexei Yavlinsky

This entry was posted on Wednesday, September 30th, 2009 at 10:57 am and is filed under Guest Authors, Image, Verticals. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.