Edge-preserving smoothing for segmentation of seismic images

Introduction

Because it can reduce the time and effort required for human-intensive image interpretation tasks, automatic image segmentation is a tool employed in a variety of disciplines - for example, medical imaging, photo editing/image processing, and seismic imaging and interpretation. A particular area of interest involves the pre-processing of images prior to automatic segmentation (Zahedi and Thomas, 1993). Here, I investigate the usefulness of an edge-preserving smoothing (EPS) technique for segmentation of seismic images.

A primary use for automatic seismic image segmentation is for identification and location of complex subsurface salt bodies, a task that is extremely time-consuming when undertaken manaully. In the examples here, the pairwise region comparison (PRC) segmentation algorithm (Felzenszwalb and Huttenlocher, 2004) is employed because it is designed to operate extremely efficiently, even when extended to three dimensions and adapted for seismic data (Halpert et al., 2010). However, like any segmentation algorithm, its accuracy can suffer, especially where boundaries are discontinuous or chaotic (for example, see Figure 10(a)). In such cases, smoothing the image before the segmentation procedure can reduce unwanted noise and improve performance of many image processing algorithms (Zahedi and Thomas, 1993).

For seismic images, naive box or Gaussian smoothing has clear disadvantages. When segmenting seismic images, clear and sharp boundaries are preferable for an accurate result; simple smoothing tends to blur these boundaries, or even render them uninterpretable if two reflectors are very close together (see Figure 3(a)). A variety of ``smarter'' smoothing or noise-reduction approaches for seismic data have been proposed, including inversion-based techniques like PEFs (Claerbout, 2005; Guitton, 2005), structure or dip-oriented filtering (Fehmers and Hocker, 2003), or bilateral filtering (Hale, 2011). Unfortunately, these algorithms require computationally-intensive inversions and/or solutions to differential equations like the diffusion equation, or prior information in the form of dip or structure interpretations. Therefore, a cheap, efficient smoothing algorithm that preserves sharp boundaries would represent a useful pre-processing step for seismic image segmentation.

masks
Figure 1. A set of 1D bar masks used to determine the most homogeneous orientation around a central pixel on a 2D image. For 3D images, additional masks would extend out of the page.

Edge-preserving smoothing for segmentation of seismic images