Image Source / Aurora Open
Advances in computer technology and machine learning may change how we search for pictures. Mathieu Chuat from Xerox, pictures the changing world of picture search
Thursday at CEPIC (Coordination of European Picture Agencies Stock, Press and Heritage) sees the Metadata conference featuring some high-powered seminars on issues such as licensing and picture search. We asked, Mathieu Chuat one of the leading industry experts on Visual Search, to flag up some of the key developments in image searches,
What have been the biggest changes in searching for images over the last few years?
Mathieu Chuat: Over the last few years there has been an explosion in the number of images available on the web, in particular on social networks. Too often searching for images relies on the image file name or tags that have been labelled by anonymous users (sometimes in multiple languages). A somewhat extreme example is image tagging by crowds in off-shore locations, that could lead to inconsistent and inaccurate image tagging, and therefore to poor search results. Recent progress in computer vision and machine learning have enabled new, better and very fast technologies to automatically extract, identify and summarize the visual content of images.
What is Visual Search?
Mathieu Chuat: Visual search relies on the visual content of the image, not on any text keywords, tags, filenames or other text metadata associated with the image. A digital signature of the image is instantly extracted, as well as the proximity/distance with digital signatures of other images (or distributions of images). This allows intelligent search and navigation among very vast volumes of images.
How long have Xerox been developing it?
Mathieu Chuat: Since 2002.
How will it help photographers, libraries and users/purchasers of images?
Mathieu Chuat: Photographers are visual people, and they dislike having to type in keywords into the metadata of their image files, although they know they have to do it if they want their images to be hit by queries. A system that would automatically add keywords into the image file metadata would contribute to increase the accuracy of image search, and therefore the retrieval of images and the commerce of image copyrights. A system that would search for images based on similarity, or lack of similarity, without even the need for textual keywords/metadata, would also increase the convenience of image search and/or navigation through image collections. A system combining keyword/category-based image search and similarity-based image search with an intelligent user interface brings the best of both worlds, relying on words and visual content to find the right images.