Efficient mining of informative descriptors from data with scarce annotations

Zhuang, Furen

Permalink

https://hdl.handle.net/2142/117596 Copy

Description

Title

Efficient mining of informative descriptors from data with scarce annotations

Author(s)

Zhuang, Furen

Issue Date

2022-12-02

Director of Research (if dissertation) or Advisor (if thesis)

Moulin, Pierre

Doctoral Committee Chair(s)

Moulin, Pierre

Committee Member(s)

• Veeravalli, Venugopal V
• Schwing, Alexander G
• Wang, Yuxiong

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Metric Learning
• Retrieval
• Hashing

Abstract

In this thesis informative descriptors are learnt where similar data are closer than dissimilar ones. Such descriptors are commonly used in Content-Based Information Retrieval applications, where the amount of data that needs to be handled is very large. To that end, it is ideal for the descriptors to be compact, the training method able to effectively utilize scarce annotations, and the representations able to generalize to unseen data. The second chapter shows how very short hash codes can be learnt which perform well in image retrieval. It is desired for such hash codes to be similarity-preserving, balanced and pairwise uncorrelated. Similarity-preserving means the hash codes of similar images have a shorter Hamming distance than dissimilar ones. Balanced bits means each bit coordinate has uniform probability. Balanced and uncorrelated bits encourage an equal number of data items to be mapped to each code. We utilize a variational autoencoder (VAE) and show how all three ideals can be seamlessly incorporated into the VAE to directly obtain hash bits from an intermediate layer. We also extend the framework to improve generalizability, allowing the hash codes to perform well even on classes that were unseen during training. This is achieved by explicitly tolerating variation found in the hash codes of similar data, while ensuring that label-consistency is maintained. The third chapter proposes a semi-supervised metric learning method which is computationally efficient due to the use of proxies and is able to effectively harness unlabeled data by identifying far-apart similar pairs and close dissimilar pairs. In the fourth chapter, we show how this identification can be improved using the provided labels. A new mixed label propagation method is also proposed to incorporate negative edge information into label propagation.

Graduation Semester

2022-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Furen Zhuang

Owning Collections