Research

Image recognition of landmarks and scenes

Since automatic image recognition is an extremely complex problem, e.g., people use unconsciously up to fifty percent of their brain capacity for distinguishing a cow from a sheep, the BeeldCanon project focuses on the 100 most important Dutch / Flemish monuments and the 100 most important landscapes and scenes. We will develop detectors for automatic recognition of the castle of Horst, the Knight’s Hall in The Hague, the Grand Béguinage of Leuven, Alkmaar's cheese market and so on. We recognize these famous landmarks and scenes based on the image content and independent of the presence of any accompanying text or metadata information. Recognition consists of two steps. In step 1, the information in the image is reduced to a set of features that measure variations in color, shape and texture Step 2 will then learn the specific feature composition of a monument or scene by means of examples.

 

The Royal Palace on Amsterdam’s dam square as captured digitally throughout history. In the BeeldCanon project we aim for the automatic recognition of this landmark together with a hundred other characteristic Flemish / Dutch monuments and scenes based on the image content. By doing so we form an expandable visual lexicon (BeeldCanon) of Dutch - Flemish culture and the tools for retrieving the images and video clips containing these landmarks. (Photos courtesy of Spaarnestad Photo and Flickr).

Scientific challenge

The scientific challenge lies in the realization of a video search engine for landmarks and scenes that is at the same time robust, delivers results fast, and also possesses a high reliability. The BeeldCanon project aims to achieve this by applying state-of-the-art techniques from the fields of image processing, computer vision and video retrieval. Thanks to the enormous evolution in these fields over the last year, is has become possible to develop detectors for images containing the Atomium or the Zaanse Schans based on the visual content. That is different than just relying on analysis of relevant text information, which is often unavailable. The main scientific innovation lies in a strong improvement of the robustness, performance and effectiveness of the image search methods.