Use this URL to cite or link to this record in EThOS:
Title: Graph spectral feature learning for shape representation
Author: Alwaely, Basheer
ISNI:       0000 0004 8501 454X
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Two decades ago, it was difficult to imagine a machine that can auto-detect patterns in images. The state-of-the-art studies have advanced shape recognition, which explores the representation of each modality. However, it remains a challenge for machines to learn shape representations accurately. One way to describe the topology of a shape is to use feature representations. Discriminative features have attracted attention because of their potential in recognising patterns with more accuracy at rapid execution speeds. In this thesis, we first introduce a novel method for in-air arbitrary shape recognition. Particularly, we propose a subset of graph spectral features, which are invariant to rotation, flip, and mirror changes, to classify different shape categories. A new dataset is also introduced for in-air hand-drawn that includes samples of shapes and numbers. This method has resulted in the highest performance with accuracies of 99.56% and 99.44%, for numbers and shapes, respectively, outperforming the existing methods for different datasets. The second contribution involves the development of an adaptive graph connectivity method based on the local details. Such graphs are created to fit the topology of the targeted shapes. A set of spectral graph features are then extracted to capture the local details, improving the state-of-the-art performance by 2% and 9% for two 2D datasets, and 2% and 6% for two 3D datasets. The third contribution focuses on simplifying the shapes by further exploring the local details. A spectral partitioning is applied to understand the geometric structure of each part. This is followed by extracting the spectral features to identify the shapes. The empirical evaluation shows that partitioning of the shapes provides deep geometric details that have demonstrated increasing accuracy levels in different datasets by 1.02%, 5.09%, 2.1%, and 7.89%.
Supervisor: Abhayaratne, charith ; Rockett, Peter Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available