DOI

https://doi.org/10.25772/RE06-AV14

Author ORCID Identifier

0000-0002-2527-6375

Defense Date

2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Krzysztof J. Cios

Abstract

One of the main problems of a supervised deep learning approach is that it requires large amounts of labeled training data, which are not always easily available. This PhD dissertation addresses the above-mentioned problem by using a novel unsupervised deep learning face verification system called UFace, that does not require labeled training data as it automatically, in an unsupervised way, generates training data from even a relatively small size of data. The method starts by selecting, in unsupervised way, k-most similar and k-most dissimilar images for a given face image. Moreover, this PhD dissertation proposes a new loss function to make it work with the proposed method. Specifically, the method computes loss function k times for both similar and dissimilar images for each input image in order to increase the discriminative power of feature vectors to learn the inter-class and intra-class face variability. The training is carried out based on the similar and dissimilar input face image vector rather than the same training input face image vector in order to extract face embeddings.

The UFace is evaluated on four benchmark face verification datasets: Labeled Faces in the Wild dataset (LFW), YouTube Faces dataset (YTF), Cross-age LFW (CALFW) and Celebrities in Frontal Profile in the Wild (CFP-FP) datasets. The results show that we gain an accuracy of 99.40\%, 96.04\%, 95.12\% and 97.89\% respectively. The achieved results, despite being unsupervised, is on par to a similar but fully supervised methods.

Another, related to face verification, area of research is on face anti-spoofing systems. State-of-the-art face anti-spoofing systems use either deep learning, or manually extracted image quality features. However, many of the existing image quality features used in face anti-spoofing systems are not well discriminating spoofed and genuine faces. Additionally, State-of-the-art face anti-spoofing systems that use deep learning approaches do not generalize well.

Thus, to address the above problem, this PhD dissertation proposes hybrid face anti-spoofing system that considers the best from image quality feature and deep learning approaches. This work selects and proposes a set of seven novel no-reference image quality features measurement, that discriminate well between spoofed and genuine faces, to complement the deep learning approach. It then, proposes two approaches: In the first approach, the scores from the image quality features are fused with the deep learning classifier scores in a weighted fashion. The combined scores are used to determine whether a given input face image is genuine or spoofed. In the second approach, the image quality features are concatenated with the deep learning features. Then, the concatenated features vector is fed to the classifier to improve the performance and generalization of anti-spoofing system.

Extensive evaluations are conducted to evaluate their performance on five benchmark face anti-spoofing datasets: Replay-Attack, CASIA-MFSD, MSU-MFSD, OULU-NPU and SiW. Experiments on these datasets show that it gives better results than several of the state-of-the-art anti-spoofing systems in many scenarios.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

8-11-2023

Download

Share

COinS