Scale Space Exploration For Mining Image Information Content

Images are highly complex multidimensional signals, with rich and complicated information content. For this reason they are difficult to analyze with a specific automated approach. However, a hierarchical representation is helpful for understanding image content. In this paper, we describe an application of a scale-space clustering algorithm (melting) for exploration of image information content. Clustering by melting considers the feature space as a thermodynamical ensemble and groups the data by minimizing the free energy, having temperature as a scale parameter. We develop clustering by melting for multidimensional data, and propose and demonstrate a solution for the initialization of the algorithm. Due to the curse of dimensionality, for initialization of clusters we choose the initial clusters centers with an algorithm that performs fast cluster center estimation with low computation cost. We further analyze the information extracted by melting and propose a structure for information representation that enables exploration of image content. This structure is a tree in the scale space showing how the clusters merge. Implementation of the algorithm is through a multi-tree structure. With this structure, we can explore the image content as an information mining function, we obtain a more compact data structure, and we have maximum of information in scale space because we memorize the bifurcation points and the trajectories of the centers points in the scale space. The information encoded in the tree structure enables the fast reconstruction and exploration of the data cluster structure and the investigation of hierarchical sequences of image classifications. We demonstrate the effectiveness of the approach with examples using satellite multispectral image (SPOT 4) and Synthetic Aperture Radar (SAR) and Digital Elevation Models (DEM) derived from SAR interferometry (SRTM).