文摘
Semantic image segmentation aims to partition an image into separate regions, which ideally correspond to different real-world objects. It is one of the most critical steps towards content analysis and image understanding. Image segmentation can be viewed as a clustering problem attempting to determine which pixels belong together. Although many different clustering techniques have been applied in the attempt to achieve a better segmentation method, most of them use only simple similarity measures such as the distance between image features. We give a justification that segmentation methods using only simple similarity measures are inherently biased. In this dissertation, we aim to improve clustering-based segmentation using evaluation measures, which incorporate prior knowledge about what segmentations are more preferred. We propose two machine learning-based ensemble evaluation techniques: Co-Evaluation and Meta-Evaluation, to improve evaluation accuracy. And we extend ensemble evaluation to the collective supervised clustering (CSC) framework, which enables a segmentation algorithm to use a set of similarity measures or evaluation measures collaboratively in the clustering process. By assigning different weights to different similarity/evaluation measures according to the characteristics of the measures and the content of the image to be segmented, better segmentation results can be achieved using CSC. Clustering-based segmentation methods using prior knowledge about object(s), either by giving object templates or by learning, are also discussed in this dissertation.