DEVELOPMENT OF ALGORITHMS FOR COLOUR IMAGE SEGMENTATION

Abstract

This thesis aims to advance research in image segmentation by developing robust techniques for evaluating image segmentation algorithms. The key contributions of this work are as follows. First, the characteristics of existing measures are investigated for supervised evaluation of automatic image segmentation algorithms. We show which of these measures is most effective at distinguishing perceptually accurate image segmentation from inaccurate segmentation. Then apply these measures to evaluating four state-of-the-art automatic image segmentation algorithms, and establish which best emulates human perceptual grouping. Second, a complete framework for evaluating interactive segmentation algorithms is developed by means of user experiments. This system comprises evaluation measures, ground truth data, and implementation software. Proposed measures are validated by showing their correlation with perceived accuracy. This framework is used to evaluate popular interactive segmentation algorithms, and demonstrate their performance. Finally, acknowledging that user experiments are sometimes prohibitive in practice, a new method is proposed based on evaluating interactive segmentation by algorithmically simulating the user interactions. Finally matting algorithm is used for segmentation and to compare it with existing segmentation methods.