Classifying Dinoflagellates in Palynological Slides Using Convolutional Neural Networks

Abstract

The petroleum industry is still one of the largest contributors to the Norwegian economy. Experts estimates that of the total reserves on the Norwegian shelf only 52 percent have been discovered. During test drilling, core samples can be taken from the sedimentary rock and within these samples small fossils from micro-plankton known as dinoflagellates can be found. By evaluating the distribution and collection of different species and taxon of dinoflagellate the likelihood of finding petroleum in the area can be estimated.

Palynology is the study of such small objects, and have largely been done manually through a microscope. The Norwegian Petroleum Directorate have recently acquired a scanner to digitize their collection of over 200,000 palynological slides. In this thesis a solution is proposed to automatically detect and identify a number of different dinoflagellate species by using both traditional image processing and deep neural networks.

With the aid of traditional image processing a detection rate of 93 percent was obtained for detecting objects in the palynological slides. Using transfer learning, a deep convolutional neural network based on the VGG-16 network structure obtained a 99 percent accuracy on test data.