Dissertations, Theses, and Capstone Projects

Author

Longlong Jing

Date of Degree

9-2021

Document Type

Dissertation

Degree Name

Ph.D.

Program

Computer Science

Advisor

Yingli Tian

Committee Members

Zhigang Zhu

Ioannis Stamos

Charles Qi

Keywords

computer vision, deep learning

Abstract

Large-scale labeled datasets are generally required to train deep neural networks in order to obtain better performance in visual feature learning for computer vision applications. To reduce the extensive cost of collecting and annotating large-scale labeled datasets, various machine learning methods are proposed to learn general visual features including semi-supervised methods which learn visual features from a small size of labeled data and a large amount of unlabeled data, weakly supervised methods which learn visual features from coarse-grained labeled data, and self-supervised methods which learn visual features from large-scale unlabeled data. In this thesis, we investigate a number of approaches to learn robust deep visual features from data with different level of supervisions including a weakly supervised method, a semi-supervised learning method, and several self-supervised learning methods. To demonstrate the generalization ability of the proposed methods to learn from limited supervisions, we validate the proposed methods on different tasks and demonstrate that the proposed methods indeed can learn robust visual features from limited labeled data.

Share

COinS