Biometric information analyses using computer vision techniques.

Hu, Bingzhang (2019) Biometric information analyses using computer vision techniques. Doctoral thesis, University of East Anglia.

[thumbnail of 2020HuBPhD.pdf]
Download (8MB) | Preview


Biometric information analysis is derived from the analysis of a series of physical and biological characteristics of a person. It is widely regarded as the most fundamental task in the realms of computer vision and machine learning. With the overwhelming power of computer vision techniques, biometric information analysis have received increasing attention in the past decades. Biometric information can be analyzed from many sources including iris, retina, voice, fingerprint, facial image or even the way one walks with. Facial image and gait, because of their easy availability, are two preferable sources of biometric information analysis.

In this thesis, we investigated the development of most recent computer vision techniques and proposed various state-of-the-art models to solve the four principle problems in biometric information analysis including the age estimation, age progression, face retrieval and gait recognition.

For age estimation, the modeling has always been a challenge. Existing works model the age estimation problem as either a classification or a regression problem. However, these two types of models are not able to reveal the intrinsic nature of human age. To this end, we proposed a novel hierarchical framework and a ordinal metric learning based method. In the hierarchical framework, a random forest based clustering method is introduced to find an optimal age grouping protocol. In the ordinal metric learning approach, the age estimation is solved by learning an subspace where the ordinal structure of the data is preserved. Both of them have achieved state-of-the-art performance.

For face retrieval, specifically under a cross-age setting, we first proposed a novel task, that is given two images, finding the target image which is supposed to have the same identity with the first input and the same age with the second input. To tackle this task, we proposed a joint manifold learning method that can disentangle the identity with the age information. Accompanied with two independent similarity measurements, the retrieval can be easily performed.
For aging progression, we also proposed a novel task that has never been considered. We devoted to fuse the identity of one image with the age of another image. By proposing a novel framework based on generative adversarial networks, our model is able to generate close-to-realistic images.

Lastly, although gait recognition is an ideal long-distance biometric information task that makes up the shortfall of facial image, existing works are not able to handle large scale data with various view angles. We proposed a generative model to solve this term and achieved promising results. Moreover, our model is able to generate evidences for forensic usage.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Chris White
Date Deposited: 07 Apr 2021 13:20
Last Modified: 07 Apr 2021 13:20


Downloads per month over past year

Actions (login required)

View Item View Item