Enhancing Hierarchical Segmentation: Techniques and Evaluation using α-Trees

This thesis focuses on developing and improving hierarchical tree structures, particularly alpha-trees, for analyzing and segmenting complex images with narrow structures.
1. Detection of Narrow Objects: New techniques are introduced to improve narrow object detection in alpha-trees. By using orientation-based alpha-trees and alpha-omega trees with Gabor filters inspired by the human visual cortex, narrow structures are effectively protected from noise-induced fragmentation. Tests on synthetic and natural images show that this method significantly enhances segmentation quality.
2. Preventing Chaining: A method is proposed to prevent the undesirable “chaining” effect in alpha-trees. Based on odd 2D Gabor filters, this approach achieves a segmentation accuracy of 99.8%, effectively eliminating the chaining effect.
3. Alpha-Tree Quality Evaluation (Horizontal Cut): An evaluation algorithm for alpha-omega trees is developed, using horizontal cuts to automatically select optimal parameters. Core factors such as accuracy and efficiency are evaluated using satellite images, and this algorithm is also applicable to other tree structures.
4. Evaluation with Non-Horizontal Cut: The evaluation is extended to non-horizontal cuts, providing greater flexibility in segmentation. Satellite images are again used to test this quality measure, demonstrating the applicability of this approach to various hierarchical tree structures.