Weighted Linked Pyramids and Soft Segmentation of Colour Images

We describe the use of weighted linked pyramids for soft region segmentation of colour images. In many digital images definitive region boundaries can not be determined, casting doubt on the reasonableness of a conventional crisp segmentation. We propose that soft segmentation gives a more natural region segmentation of digital images than conventional crisp segmentation. Weighted linked pyramids provide a straightforward basis for a method of producing a soft image segmentation. This paper also describes the use of shift variance as a method for comparing different pyramid-based segmentation methods. We show that the soft region segmentation is closer to the original image and superior to the crisp segmentation in terms of shift invariance.