Google makes advances in image recognition



Google has just published an interesting research paper which suggests that they have made big advances in the image-recognition software field. At a computer vision and pattern recognition conference in Miami recently, Jay Yagnik, google’s head of vision research claimed that their new system has had an 80% accuracy in identifying untagged images of up to 50,000 famous landmarks.

This is exciting news that will, if and when the system is integrated into google’s search engine (to be clear, the research remains, at the moment, just that – research).

Exciting and challenging, because at the same time that google is increasingly adding both image and video results to their main search pages, the current s.e.o practices for images are almost exclusively text and code based. At the moment if you want your images to appear in the google index you need to focus on giving them appropriate file-names, filling in the alt tags with appropriate keywords, and generally optimising the context and page on which the image appears.

Sophisticated image recognition could potentially change all of that. Imagine a site that has lot’s of images on its pages but none that are actually labelled or set up in terms of the above s.e.o. The images are all simply called image1.jpg, image2.jpg etc and the alt tags are left blank. These images are not completely useless to google and other search engines currently – they can analyse the page context and hazard a guess that the images are of something in particular. The chances, though, of the images ranking well are poor though, and are almost completely in the hands of one’s competitors – if they’ve optimised well, they’ll outrank these images easily. It’s also highly unlikely that google will give prominence to these images on the main results page – as the possibility of them being spam is just too high.

If some computer vision system were able, however, to correctly identify these unlabelled images, then coupled with more general s.e.o. on the page you could start to see them crop up on that all important first page of results – spelling a lot of traffic for some, and perhaps a corresponding drop in traffic for others.

For the moment, though, this is all in the future – and related specifically to landmarks and readily definable visual content. The tip for the top is to make sure you have all your images presented on the page in a search engine friendly way (and in keeping with accessibility). So:

  1. Give image files descriptive filenames
  2. Make sure that images on your page have their alt tags properly filled in
  3. Put images in context – to perform well, make sure they’re not left hanging on their own on your site, but are placed in a proper textual context

Tags: , , , ,