#52: Automatic object extraction from full differential morphological profile in urban imagery for efficient object indexing and retrievals


C. R. Shyu, G. Scott, M. Klaric, C. H. Davis, and K. Palaniappan

3rd Int. Symposium on Remote Sensing and Data Fusion Over Urban Areas (URBAN 2005), Volume 36, 2005

gis, data mining, features, machine learning, shape, remote sensing, cbir

PlainText, Bibtex, PDF, Google Scholar

Abstract

The differential morphological profile (DMP) can be used for automated extraction of multi-scale urban features, such as buildings, shadows, roads, and other man-made objects. However, characterization of urban features using the DMP is complicated by the fact that some objects will have response at multiple-scales within the DMP. This makes robust and efficient object indexing and retrieval difficult for large-scale remote sensing image databases utilized in the defense and intelligence communities. To address this issue, in this paper we present a novel approach called Multi-scale Extraction of Morphological Objects (MEMO),which is fully automatic and unsupervised. MEMO contains two processing modules for identifying urban objects: (1) Top-down object fusion: multi-scale objects from both morphological closing and opening profiles with certain topological relationships (TR), such as overlap, equal, and inside, will be selected for candidate objects and placed in the candidate pool. (2) Knowledge-based filtering: objects of the DMP are refined and filtered using information present in the original panchromatic image, spectral information of the scene, and the processed DMP. For example, areas of vegetative land cover are filtered out reducing the false labelling of tree clusters and fields candidate objects. Additionally, size and shape analysis of candidate objects can further eliminate possible false object extraction. The efficiency of our algorithm makes it applicable to large-scale object indexing and retrieval.