Machine learning for adaptive parameter selection in image segmentation.

摘要

Applying an algorithm to a new domain often incurs much tuning to its many control parameters. A problem with such a tuning approach is that it is rarely possible for an algorithm to achieve the best segmentation result on a per image basis. This thesis investigates applying machine learning techniques to adaptively select parameters for image segmentation. We adopt Multi Resolution Adaptive Object Recognition (MR ADORE) as the overall framework of our system and extend it with several novel components into a system that is capable of adaptively selecting parameters for image segmentation algorithms. In particular, we implement a fragment based similarity scoring metric, a Generalized Gaussian Distribution based feature extraction method, and a new pruning strategy called the machine learned branch expansion. Experiments show that the new system achieves better accuracy than the best static algorithm.