Abstract
This thesis is a study of perception-driven automatic segmentation of colour images. Despite immediate practical interest for this task, there exist very few reliable algorithms suitable for unsupervised processing. Most of the results presented in this thesis are based on mathematical morphology. This is a relatively new field which explores topological and geometrical properties of images and which has proven to be useful for image processing. The overview of morphological techniques can be found in chapter 2. A brief overview of segmentation methods is presented in, chapter 3. Only a small proportion of the vast number of publications on the subject is reviewed, namely those that are papers directly relevant to the subject of the thesis. Two novel non-parametric algorithms have been developed by the author for processing colour images. The first one is for processing randomly textured images. It uses a bottom-up segmentation algorithm which takes into consideration both colour and texture properties of the image. An "LUV gradient" is introduced which provides both a colour similarity measure and a basis for applying the watershed transform. The patches of watershed mosaic are merged according to their colour contrast until a termination criterion is met. This criterion is based on the topology of a typical processed image. The resulting algorithm does not require any additional information, be it various thresholds, marker extraction rules and suchlike, thus being suitable for automatic processing. The second algorithm deals with non-textured images and takes into consideration the noise that is present during the image acquisition. The watershed algorithm is used to segment either the 2- or 3-dimensional colour histogram of an image. To comply with the way humans perceive colour, this segmentation has to take place in a perceptually uniform colour space such as the Luv space. To avoid over segmentation, the watershed algorithm has to be applied to a smoothed-out histogram. The noise, however, is inhomogeneous in the Luv space and noise analysis for this space based on experimentally justified assumptions is presented. Both algorithms have been extensively tested on real data and were found to give stable results that are in good accord with human perception.