Hierarchical data representation structures for interactive image information mining

摘要

In this article an interactive image information mining protocol is presented aiming at a computationally efficient pattern interpretation. The method operates on very high resolution (VHR) remote-sensing optical imagery and follows a modular approach. Images are projected onto a hierarchical image representation structure, the Max-Tree, which interfaces multi-dimensional features of the image components. Positive and negative samples are selected interactively from the image space and are translated into features describing best the targeted and non-desired patterns. Sourcing the feature entries into a hierarchical clustering algorithm, the kd-Tree, yields a structured representation that ensures fast classification. A classification is computed directly from the kd-Tree and is applied on the Max-Tree for accepting or rejecting image components. The complete process cycle is demonstrated on gigapixel-sized VHR satellite images and requires 3 min for building the Max-Tree, 30 min for hierarchical clustering and less than 10 s for each example based query.