NOTE! This site uses cookies and similar technologies.

If you not change browser settings, you agree to it. Learn more

I understand

Learn more about cookies at : http://www.aboutcookies.org/Default.aspx?page=1

May 20, 2015 | Online image searches faster and more accurate

The latest CEA LIST technologies are making online image searches faster and more accurate.

Digital devices like cameras and smartphones are becoming permanent fixtures of our day-to-day lives, fueling exponential growth in the proportion of images among online data. Images are driving the increasing popularity of photo-sharing sites like Instagram, Picasa, and Flickr. But, because users often fail to tag their images sufficiently (or at all), it is difficult to perform content searches.

The latest technologies developed by CEA LIST, a member of the Instituts Carnot network, are helping overcome the three major hurdles currently hindering online image searches:

  • Gaps between the key points in an image and the way in which the images are interpreted by
  • The wide variety of images online (subject matter, resolution, technical quality)
  • The capacity of search algorithms to explore databases of hundreds of millions of images on a personal computer

Researchers at CEA LIST have overcome these challenges by using image content modelling based on a large battery of visual models—or classifiers—leveraging the “deep learning” descriptions of the images.

Researchers at CEA LIST have overcome these challenges by using image content modelling based on a large battery of visual models—or classifiers—leveraging the “deep learning” descriptions of the images. The researchers also simplified the way in which similarities between images, a key element of research techniques, are calculated. They used a semantic representation limited to a few dozen relevant classifiers. Finally, they gave their tool new capabilities to learn new visual models, creating new classifiers to respond to new and previously unidentified searches.

The next step will be to integrate these technologies into CEA LIST’s multimedia analysis tools, which could potentially be used in a very large number of fields, including for online search engines.

These advances were made under two EU-backed research projects in which CEA LIST participated: USEMP (FP7 611 596) and MUCKE (FP7 CHIST-ERA).