pixolu is a prototype Internet image search system using the pixolution technologies. pixolu demonstrates the potential of a modern visual image search system combining keyword search, visual sorting and similarity search, and semi-automatically learned semantic image relationships.

Try pixolu ...	Watch our video showing how to use pixolu.

Using conventional keyword image search systems, only a tiny fraction of the huge result set can be inspected, making the search for a particular image very time consuming. Most image search systems display their results on separate web pages containing sets of 20 images. This is due to the fact that human perception is limited to 20 - 50 unsorted pictures at a time. Other problems arise from homonyms, names that are similar to the query keyword, and incorrectly assigned keywords. If the goal is not to find an image that matches the keyword, but rather to find a special image with particular attributes, then the performance of most image search systems is rather poor, as the search systems neither know the intention of the searching user nor the semantic relationships of the images.

pixolu makes finding images easier
pixolu initially also uses a keyword search, however many more result images are displayed. This is possible because the images are sorted according to their visual similarity. Due to this several hundred images can easily be inspected simultaneously. In most cases the amount of images is large enough to get a good representation of the entire result set. Most likely the user will find good candidates for the desired search result. In the next step pixolu uses these candidate images as templates to search more visually similar images. In addition pixolu is able to learn semantic relationships between images from the candidates sets of different users and therefore also is able to find semantically similar images.

Click here to use pixolu to search images from Flickr © and Yahoo ©.



pixolu uses the pixolution core technologies. The web application embedding these core technologies was develeoped at HTW Berlin as a student project (Simon Burkard, Dominik Deichsel, Jörn Ehmann, Friedrich Maiwald and Ben Suksanguan) under the supervision of Prof. Dr. Kai Uwe Barthel.