Voronoi Seeded Colour Image Segmentation

The goal of the segmentation scheme presented i s t o c ombine edge and region information to achieve a stable segmentation. The segmentation scheme presented is designed t o o p erate on general home and stock photographs, it returns a comprehensive region-based description of the visual content of an image (including a distinction between smooth and textured r egions and a description of the internal properties of the later). A colour edge detector is presented, where hue diierence i s w e i g h t e d m o r e h e avily than brightness diierence. Seed p oints for region growing are derived f r om the colour edge image as the peaks in the associated Voronoi image. R egions are g r own using gates on pixel colour relative to region central colour and region edge pixel colour. This permits regions to encompass shading gradients. Image edges act as hard b arriers during region growing. A discrete feature based texture model is derived and then used t o u n i f y g r oups of smaller regions into extended textured r egions. The segmentation scheme is designed to facilitate image retrieval and has been tested o n a corpus of over 40000 images and has been found to be r obust.