Automated Grain-Matrix Segmentation in Photomicrographs of Clastic Sedimentary Rocks using Neuro-visual Algorithms

Abstract

Sediments and sedimentary rocks, covering almost sixty-five percent of exposed earth’s surface, contain features that allow us to interpret ancient depositional environments, and their role in the evolution of life on earth. Study of sedimentary rocks is also relevant to exploration of geo-economic resources including petroleum, ore minerals, and groundwater as well as research in environmental geology, anthropological studies, and the bi-directional relation of man and environment. One of the important and crucial steps in this study of the detrital sedimentary rocks like sandstone, is to know the size and shape of the grains/clasts (fragments of different minerals or rocks), and the character of the fine-grained matrix (usually clay sized material or chemical precipitates), that surround the individual clasts. Geologists generally use a petrographic microscope to distinguish and study the clasts and the matrix. Traditionally the identified grains/clasts are demonstrated by the geologist on a photomicrograph of the thin section captured on a polarizing microscope. The images contain hundreds of grains/clasts of different sizes and characters. The manual delineation and identification of such grains is extremely time-consuming and tedious. Often the grain boundaries are hazy and almost overlapping, and their exact delineation is highly subjective at places. Therefore, the machine vision and image analysis techniques for grain segmentation from the matrix, and for the measurement of other properties of these grains, appear to be one of the potential tools. Yet, until now, to the best of our knowledge no robust general solution for that has been drawn. Also, for applying state-of-art algorithms in the domain to check their validity in grain-matrix separation is impossible as no generic dataset for this is available at present. In this regard, the present thesis builds a dataset of photomicrographs captured under a petrographic microscope of a thin section of rock samples, and their corresponding ground truths in collaboration with the geologists. This dataset consists of two views, the plane polarized (PPL) view observed under a single polarizer and the cross polarized (XPL) view when the analyzer is inserted as well. Three different approaches in computer vision have been adopted in this thesis to solve the above-mentioned grain-matrix segmentation problem relying upon the concept of mimicking the geologist’s eye on the petrographic microscope from various perspectives. The first of these, uses separate spatial filters for three commonly applied modes of vision viz. Vision applying Strong Contrast (VSC), Vision at a Glance (VG), and Vision with Scrutiny (VS). For this, a psychophysics based receptive field model is combined with a clustering algorithm for figure-ground segregation of the XPL images. The second approach employs a Laplacian of Gaussian filter based mid-level vision model for automated segmentation of the XPL images. It adaptively combines VG and VSC filters mentioned above and ends up with a filter termed as Vision with Adaptive Scrutiny (VAS). The third approach is a Linknet inspired Deep Semantic Grain Segmentation Network that takes cues not only from XPL, but also the PPL image, much like what is often actually done while using the microscope. This is compared with state-of-the-art deep neural networks (itself inspired by visual cortical processing) for image segmentation. The results have been quantitatively analyzed using suitable metrics. The thesis demonstrates that computer vision algorithms, especially the neuro-visually inspired ones, including deep neural nets, have a high potential in automating the difficult problem of grain segmentation from the photomicrographs of clastic sedimentary rocks.