Unsupervised detection of compound structures using image segmentation and graph-based texture analysis
MetadataShow full item record
Please cite this item using this persistent URLhttp://hdl.handle.net/11693/14913
The common goal of object-based image analysis techniques in the literature is to partition the images into homogeneous regions and classify these regions. However, such homogeneous regions often correspond to very small details in very high spatial resolution images obtained from the new generation sensors. One interesting way of enabling the high-level understanding of the image content is to identify the image regions that are intrinsically heterogeneous. These image regions are comprised of primitive objects of many diverse types, and can also be referred to as compound structures. The detection of compound structures can be posed as a generalized segmentation or generalized texture detection problem, where the elements of interest are primitive objects instead of traditional case of pixels. Traditional segmentation methods extract regions with similar spectral content and texture models assume specific scale and orientation. Hence, they cannot handle the complexity of compound structures that consist of multiple regions with different spectral content and arbitrary scale and orientation. In this thesis, we present an unsupervised method for discovering compound image structures that are comprised of simpler primitive objects. An initial segmentation step produces image regions with homogeneous spectral content. Then, the segmentation is translated into a relational graph structure whose nodes correspond to the regions and the edges represent the relationships between these regions. We assume that the region objects that appear together frequently can be considered as strongly related. This relation is modeled using the transition frequencies between neighboring regions, and the significant relations are found as the modes of a probability distribution estimated using the features of these transitions. Furthermore, we expect that subgraphs that consist of groups of strongly related regions correspond to compound structures. Therefore, we employ two different procedures to discover the subgraphs in the constructed graph. During the first procedure the graph is discretized and a graph-based knowledge discovery algorithm is applied to find the repeating subgraphs. Even though a single subgraph does not exclusively correspond to a particular compound structure, different subgraphs constitute parts of different compound structures. Hence, we discover compound structures by clustering the histograms of the subgraph instances with sliding image windows. The second procedure involves graph segmentation by using normalized cuts. Since the distribution of significant relations within resulting subgraphs gives an idea about the nature of corresponding compound structure, the subgraphs are further grouped by clustering the histograms of the most significant relations. The proposed method was tested using an Ikonos image. Experiments show that the discovered image areas correspond to different high-level structures with heterogeneous content such as dense residential areas with high buildings, dense and sparse residential areas with low height buildings and fields.