Date of Award
Doctor of Philosophy (PhD)
Modeling Simul & Visual Engineering
Modeling & Simulation
The selection and computation of meaningful features is critical for developing good deep learning methods. This dissertation demonstrates how focusing on this process can significantly improve the results of learning-based approaches. Specifically, this dissertation presents a series of different studies in which feature extraction and design was a significant factor for obtaining effective results. The first two studies are a content-based image retrieval system (CBIR) and a seagrass quantification study in which deep learning models were used to extract meaningful high-level features that significantly increased the performance of the approaches. Secondly, a method for change detection is proposed where the multispectral channels of satellite images are combined with different feature indices to improve the results. Then, two novel feature operators for mesh convolutional networks are presented that successfully extract invariant features from the faces and vertices of a mesh, respectively. The novel feature operators significantly outperform the previous state of the art for mesh classification and segmentation and provide two novel architectures for applying convolutional operations to the faces and vertices of geometric 3D meshes. Finally, a novel approach for automatic generation of 3D meshes is presented. The generative model efficiently uses the vertex-based feature operators proposed in the previous study and successfully learns to produce shapes from a mesh dataset with arbitrary topology.
"Feature Extraction and Design in Deep Learning Models"
(2021). Doctor of Philosophy (PhD), Dissertation, Modeling Simul & Visual Engineering, Old Dominion University, DOI: 10.25777/ss0q-gx75