Novel Orientation Representations for Deep Learning-Based Instance Segmentation

Abstract

This thesis is rooted in Computer Science and Deep Learning, with the primary application being in the interpretation of 3D seismic tomographic images, particularly the identification of geological faults therein. Deep Leaning methods can aid and automate parts of the seismic interpretation process. The outcome of the work in this thesis is a method that can predict the location of faults and separate them out into non-intersecting fault segments. These segments can then be selectively visualised in 3D. The method is powered by a convolutional neural network that predicts the location and orientation of faults, which are separated into segments using a post-processing algorithm that can run on the GPU and is scaleable to large seismic volumes. The means for the neural network to predict orientation is the biggest scientific contribution of this thesis, with four novel 3D representations. These representations are negation-invariant and continuous (f(v) = f(-v)). The work in this thesis also has relevance in other fields. A 2D counterpart to the task of fault separation is the separation of overlapping chromosomes for karyotyping. My work proposes a method for using orientation to separate chromosome instances and uncovers issues with existing approaches in the field, namely the use of semantic segmentation directly and issues with an existing dataset.