Do web search giants and smaller ones neglect image retrieval?
by Sergey Egorov
Ignoring image retrieval, based on visual characteristics, is, at least, short-sighted search policy. Of course, direct search by image example in large databases is very time-expensive process. Most web engines offer different instruments for more precise (visual similar) search in hits, received after text query.
Google has it’s own Google Similar Images Labs. So we can search similar to our example images. Also we can filter results by image size, by content (news, faces, clip art, line drawings, photo content), by dominant colour (red, orange, yellow, green, teal, blue, purple, pink, white, gray, black, brown). All these opportunities are embedded to Google search machine. Unfortunately, we can search similar not for all images, only for some.
Yandex offers us the similar opportunities: filters by size, by dominant colour (red, orange, yellow, green, cyan, blue, violet, white, black), we can choose portraits and wallpapers. Yandex has interesting connexion between keywords and colour scale (for example, query “purple”).
Yahoo has embedded porno filter, filters by size, by content (wallpapers), by colour spectrum (”Black & White”, “Color”).
Ask is similar to Yahoo: filters by size, content and colour spectrum.
Rambler can filter porno (levels – temperate, strict, disable), filters by size and content (wallpapers).
Nigma, a dynamically evolutionary search system, offers only filters by size and content (wallpapers).
GoGo, a search engine from Mail company, asserts, that duplicates are removed from result hits. But this mechanism is not transparent and far from ideal.
Bing, a new web search engine from Microsoft Corporation, offers us some interesting opportunities. It has porno filter (also with 3 levels, as Rambler). Here we can filter by size, by model (square, wide, high), by colour spectrum, by style (photos, illustrations), by human beings presence (faces, head and shoulders). Also when you direct mouse to any image, you can click on “Show similar” link and gain more relevant hits. Of course, it is not ideal, but in most cases you can receive more precise results for any image (in contrast to Google).
Picsearch. Search by text and precise by size, colour spectrum, simple categories (animation, pictures). Picsearch has a family friendliness that allows children to surf in safety as all offensive material is filtered out by filtering systems.
Nevertheless, all this engines haven’t a satisfactory duplicates detection mechanism. Examples (query “mona lisa”): in Google, in Yandex, in Yahoo, in Ask, in Rambler, in Nigma, in GoGo, in Bing (and one more), in Picsearch. GoGo, you can notice, much better than other engines in duplicates detection, however the results are evidently less relevant. Opposite, Bing has the diversity of relevant images, but, at the same time, it has many duplicates.
Thus, as we could notice, any modern search engine simply can’t permit of neglecting multimedia retrieval. They should, at least, have embedded porno filter, filters by size and colour spectrum, by content and style, by dominant colour and, of course, opportunity of visual similarity searching. The nearest future – clusterization and classification of images collection (by colours, by content – portraits, monuments, sunsets, vehicles, landscapes and etc., by style – wallpapers, photos, illustrations, line drawings and etc.), search by outlines and rough sketches (for example, bright orange spot on less bright orange background – sunset), search simple geometric shapes (such as circles, ellipses, squares, rectangles, stars, crosses, triangles, hearts and etc.). Nowadays it is necessary to have good and visually convenient duplicates detection mechanism and opportunity of search by set of colours with their percentage (like QBIC).
Written by Sergey Egorov
No tags for this post.