Author ORCID Identifier

https://orcid.org/0000-0002-8845-5055

Semester

Spring

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Xin Li

Committee Member

Donald Adjeroh

Committee Member

Matthew Valenti

Committee Member

Yuxin Liu

Committee Member

Shuo Wang

Abstract

Computer Vision (CV) enables computers and systems to derive meaningful information from acquired visual inputs, such as images and videos, and make decisions based on the extracted information. Its goal is to acquire, process, analyze, and understand the information by developing a theoretical and algorithmic model. Biometrics are distinctive and measurable human characteristics used to label or describe individuals by combining computer vision with knowledge of human physiology (e.g., face, iris, fingerprint) and behavior (e.g., gait, gaze, voice). Face is one of the most informative biometric traits. Many studies have investigated the human face from the perspectives of various different disciplines, ranging from computer vision, deep learning, to neuroscience and biometrics. In this work, we analyze the face characteristics from digital images and videos in the areas of morphing attack and defense, and autism diagnosis. For face morphing attacks generation, we proposed a transformer based generative adversarial network to generate more visually realistic morphing attacks by combining different losses, such as face matching distance, facial landmark based loss, perceptual loss and pixel-wise mean square error. In face morphing attack detection study, we designed a fusion-based few-shot learning (FSL) method to learn discriminative features from face images for few-shot morphing attack detection (FS-MAD), and extend the current binary detection into multiclass classification, namely, few-shot morphing attack fingerprinting (FS-MAF). In the autism diagnosis study, we developed a discriminative few shot learning method to analyze hour-long video data and explored the fusion of facial dynamics for facial trait classification of autism spectrum disorder (ASD) in three severity levels. The results show outstanding performance of the proposed fusion-based few-shot framework on the dataset. Besides, we further explored the possibility of performing face micro- expression spotting and feature analysis on autism video data to classify ASD and control groups. The results indicate the effectiveness of subtle facial expression changes on autism diagnosis.

Download

Share

COinS