Google made a visual search
Topic image search already raised in these pages, when the company considered Photodate , Developers who have And Idee Piximilar . Now into the act heavyweights.
Google is represented by Yushi Jing and Shumeet Baluja on the Conference WWW2008, held April 21-25, 2008 in Beijing has demonstrated its development (Link to PDF, 3,4 Mb), is able to analyze the pictures not only in name, as it was before, but the quality, relevance and most importantly - content.
Now search for images in Google images works on the basis of the text as it tries to file names. And as the name can be assigned arbitrarily, depressing picture of the issuance and funny. Try to enter a query . You will find on the first page of pictures of giraffes thick, fat women and fat children, which, of course, has no direct relation to the request.
Our approach relies on the analysis of the distribution of visual similarity among images.
The premise is simple: the author of a Web page is likely to choose the images that, with his or her point of view, are relevant topic. Instead of assuming that every user who has a web page that is relevant request, will provide a link to an image on the page of another user, our approach relies on the analysis of the combined preferences of many of the creators of the content network.
The new system image search on Google will work on the so-called "visual themes", which can be explained as meaning the center.
For example, to illustrate many of the images contain a familiar character «M». In some images, logo - the main center of the image, while in others it occupies only a small part. However, its recurrence in a large fraction of images - an important signal that can be used to derive a general "visual theme" throughout the set. Detection of multiple visual themes and their relative strength in a large set of images - the basis of image ranking system proposed in this study.
Like the Page Rank (PR), introduces the concept of Image Rank, IR. Formula Image Rank is defined as:
where S - the normalized matrix of similarity (normalized, symmetrical adjacency ma-
trix) of images, such as S u, v denote the visual similarity between image u and v.
Google has already demonstrated the operation of its search engine on the graphic example of 2000's most popular queries in Google images.
The company is close to the submission of the first beta version of the graphic search engine.
Related links: (English)
Blogobzor: And you know what a cloud of blogs ?
May 4, 2008 at 22:18
In fact, it was another hype that goes up as soon as Google sneezes. If you read the specified publication, as well as the company's own statement, it becomes clear that this is only research, and they are still very far from realization. By the way, in contrast to systems that were previously covered in this blog and who are acting version of the engine. Most likely, eventually Google or Microsoft will buy one of these companies (or all)), and that's when something will work.
May 4, 2008 at 23:45
Alexis, you're not from Picollator'a?
In my insightful look:) engines were illuminated before the company's blog at the moment are far from the ideal of human search.
That is, they look all right (with t.zr. robot) and the colors match, and shape, and is all wrong.
Just from Google you can expect something more serious.
All the prerequisites are. In addition, the ranking formulas of all prechislennyh systems are not observed, as far as I know, but it is for Google - a priority.
Well, let's wait and see.