- Title
- Enhancing deep transfer learning for image classification
- Creator
- Shermin, Tasfia
- Date
- 2021
- Type
- Text; Thesis; PhD
- Identifier
- http://researchonline.federation.edu.au/vital/access/HandleResolver/1959.17/179551
- Identifier
- vital:15631
- Abstract
- Though deep learning models require a large amount of labelled training data for yielding high performance, they are applied to accomplish many computer vision tasks such as image classification. Current models also do not perform well across different domain settings such as illumination, camera angle and real-to-synthetic. Thus the models are more likely to misclassify unknown classes as known classes. These issues challenge the supervised learning paradigm of the models and encourage the study of transfer learning approaches. Transfer learning allows us to utilise the knowledge acquired from related domains to improve performance on a target domain. Existing transfer learning approaches lack proper high-level source domain feature analyses and are prone to negative transfers for not exploring proper discriminative information across domains. Current approaches also lack at discovering necessary visual-semantic linkage and has a bias towards the source domain. In this thesis, to address these issues and improve image classification performance, we make several contributions to three different deep transfer learning scenarios, i.e., the target domain has i) labelled data; no labelled data; and no visual data. Firstly, for improving inductive transfer learning for the first scenario, we analyse the importance of high-level deep features and propose utilising them in sequential transfer learning approaches and investigating the suitable conditions for optimal performance. Secondly, to improve image classification across different domains in an open set setting by reducing negative transfers (second scenario), we propose two novel architectures. The first model has an adaptive weighting module based on underlying domain distinctive information, and the second model has an information-theoretic weighting module to reduce negative transfers. Thirdly, to learn visual classifiers when no visual data is available (third scenario) and reduce source domain bias, we propose two novel models. One model has a new two-step dense attention mechanism to discover semantic attribute-guided local visual features and mutual learning loss. The other model utilises bidirectional mapping and adversarial supervision to learn the joint distribution of source-target domains simultaneously. We propose a new pointwise mutual information dependant loss in the first model and a distance-based loss in the second one for handling source domain bias. We perform extensive evaluations on benchmark datasets and demonstrate the proposed models outperform contemporary works.; Doctor of Philosophy
- Publisher
- Federation University Australia
- Rights
- All metadata describing materials held in, or linked to, the repository is freely available under a CC0 licence
- Rights
- Copyright Tasfia Shermin
- Rights
- Open Access
- Subject
- Deep Learning,; Transfer Learning; Domain Adaptation,; Zero-shot Learning,; Image Classification
- Full Text
- Thesis Supervisor
- Lu, Guojun
- Hits: 1326
- Visitors: 1243
- Downloads: 89
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | SOURCE2 | Australian Digital Thesis | 7 MB | Adobe Acrobat PDF | View Details Download |