Attribute combinations for image segmentation

Introduction

Image segmentation - an automated process of dividing an image into regions - offers a number of promising applications for seismic data. Among the most straightforward of these applications is to the task of picking salt bodies on seismic images, a process that can be ambiguous and time-consuming when undertaken manually, especially for large three-dimensional datasets with complex salt body geometries. The development of an algorithm for automatically tracking salt boundaries (Lomask, 2007; Lomask et al., 2007) in many cases allows for the quick, efficient and globally-optimized calculation of a salt interface location. Such information may then be used, for example, to quickly update a velocity model as part of an iterative migration system (Halpert et al., 2008).

The seismic image segmentation scheme is based on the Normalized Cut Image Segmentation (NCIS) algorithm (Shi and Malik, 2000), which calculates an eigenvector based on specific attributes gleaned from the image; the eigenvector is then used to trace a boundary across the image. Although the most straightforward attribute for delineating salt boundaries on seismic images is amplitude of the envelope, this attribute alone is not always sufficient to produce an accurate calculation of the boundary. In such cases, other attributes may be used for segmentation. For example, an estimate of dips in a seismic image is often used for interpretation purposes (Bednar, 1997), and strong variations in dominant dips within an image can be indicative of a salt interface. Halpert and Clapp (2008) provide details on using dip variability, as well as an instantaneous frequency attribute, for segmentation with a single attribute. Ideally, however, a segmentation algorithm will combine information from multiple attributes into a single result. In this paper, we discuss three strategies for combining attributes: a multiplication of attribute volumes, a combination of individually calculated boundaries, and a linear combination of individual eigenvectors. The latter method, when combined with an uncertainty measurement derived from the eigenvectors, produces results superior to those using only a single attribute. Since improvements in computing capabilities make increasingly complex segmentation problems tractable, it is important to extend this process to three dimensions. Initial results from a combined-attribute 3D segmentation scheme suggest that a more sophisticated, interpreter-guided segmentation process can be successful.

Attribute combinations for image segmentation