Color Coarse Segmentation and Regions Selection For Similar Images Retrieval
Jérôme Da Rugna, Hubert Konik · Conference on Colour in Graphics Imaging and Vision · 2002
The purpose of our visual information tool is to extract from a large database images that are similar to an image query. The most popular way to achieve object information contained in images is by segmenting the image. But, it is well-known that segmentation algorithms are not robust and more or less always adapted to a specific problem. Nevertheless, a simple and coarse segmentation associated with some statisticals features may be generally a worth solution for database filtering. Then, our work is first to purpose a new low cost and basic segmentation, based on the color gaussian pyramid and, secondly, to merge this approach with a region selection process for image indexation, illustrated with the Icobra System. The main principle of our segmentation is a bottom-up process linking the different pyramid level. We compute a selection process, which selects N regions (N fixed) most representative of the visual information. Then, on each region, local invariants are calculated and accumulated in order to make a global vector which describe the entire image. The efficiency of this new method will be demonstrated by a complete evaluation and a comparison with some classical methods, using the Icobra tool.