Segmentation of thin section images for grain size analysis using region competition and edge-weighted region merging

• Introduce additional gradient-based term for edge-weighted MDL-based region merging.
• Proposed term allows better balancing of the merging order regarding smaller and larger regions.
• Validity of our approach is tested on several thin section images.
• Validity of our approach is tested against a manually segmented image.

Microscopic thin section images are a major source of information on physical properties, crystallization processes, and the evolution of rocks. Extracting the boundaries of grains is of special interest for estimating the volumetric structure of sandstone. To deal with large datasets and to relieve the geologist from a manual analysis of images, automated methods are needed for the segmentation task. This paper evaluates the region competition framework, which also includes region merging. The procedure minimizes an energy functional based on the Minimum Description Length (MDL) principle. To overcome some known drawbacks of current algorithms, we present an extension of MDL-based region merging by integrating edge information between adjacent regions. In addition, we introduce a modified implementation for region competition for overcoming computational complexities when dealing with multiple competing regions. Commonly used methods are based on solving differential equations for describing the movement of boundaries, whereas our approach implements a simple updating scheme. Furthermore, we propose intensity features for reducing the amount of data. They are derived by comparing theoretical values obtained from a model function describing the intensity inside uniaxial crystals with measured data. Error, standard deviation, and phase shift between the model and intensity measurements preserve sufficient information for a proper segmentation. Additionally, identified objects are classified into quartz grains, anhydrite, and reaction fringes by these features. This grouping is, in turn, used to improve the segmentation process further. We illustrate the benefits of this approach by four samples of microscopic thin sections and quantify them in a comparison of a segmentation result and a manually obtained one.