Microfossil analysis is essential when identifying hydrocarbon resources for the
petroleum industry. Such analysis is conducted by paleontologists who interpret the depositional environment by examining microscope slides containing samples of organic microfossils known as palynofacies. The main goal of this thesis is to develop the components of an image analysis system for automatic segmentation of palynofacies.
Microscope images of palynofacies contain three types of material: kerogen, palynomorphs and amorphous matter. These types have very different appearances and significance for the image interpretation by the domain experts. Kerogen is a type of organic microfossil that yields oil upon heating. Two kerogen types are usually presented in images of palynofacies: inertinite and vitrinite. The prevalence and the appearance of the two types carry important information about the environment. Kerogen pieces are the darkest
objects in the image, highly irregular in shape, overlapping and touching. Distinguishing between the two kerogen types is not straightforward even for the trained paleontologists.
We propose a system for automatic classification of kerogen into vitrinite and inertinte using 5 image processing stages: image acquisition, background removal, microfossil segmentation, feature extraction and classification.
Background removal corrects for uneven lighting using multiple 1D parabolas.
A marker-based segmentation method is proposed, called Centre Supported Segmentation (CSS) for identifying touching and overlapping objects in a binary image. Its only parameter expresses the acceptable degree of overlap in segmenting the individual objects. A measure of the segmentation quality is proposed to compare marker-based segmentation results.
An expert palaeontologist labelled kerogen objects as either vitrinite or inertinite, which provided the ground truth labels for training a classifier. A study comparing ten state-of-the-art classifiers singled the logistic classifier out as the most accurate one for the task. We show that the classifier is stable with respect to the overlap parameter of CSS.
The palynomorph microfossils are deformed of folded around one another. They
are presented in the image as semi-transparent, partly or entirely elliptical objects. We propose a scheme for classification of complete elliptic palynomorphs using the logistic classifier. After segmentation, two classes of objects are formed - "complete palynomorphs", and "other", containing the kerogen, amorphous matter and other unspecified debris. ROC curve analysis is used to select a certainty threshold for the classification.
The methods and solutions proposed in this study offer a toolbox for developing
commercial systems for palynofacies classification.