Semester

Spring

Date of Graduation

2021

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

Nasser M. Nasrabadi

Committee Member

Matthew C. Valenti

Committee Member

Donald Adjeroh

Committee Member

Jeremy Dawson

Committee Member

Jacqueline A. Speir

Abstract

Face recognition has been one of the most challenging areas of research in biometrics and computer vision. Many face recognition algorithms are designed to address illumination and pose problems for visible face images. In recent years, there has been significant amount of research in Heterogeneous Face Recognition (HFR). The large modality gap between faces captured in different spectrum as well as lack of training data makes heterogeneous face recognition (HFR) quite a challenging problem. In this work, we present different deep learning frameworks to address the problem of matching non-visible face photos against a gallery of visible faces.

Algorithms for thermal-to-visible face recognition can be categorized as cross-spectrum feature-based methods, or cross-spectrum image synthesis methods. In cross-spectrum feature-based face recognition a thermal probe is matched against a gallery of visible faces corresponding to the real-world scenario, in a feature subspace. The second category synthesizes a visible-like image from a thermal image which can then be used by any commercial visible spectrum face recognition system. These methods also beneficial in the sense that the synthesized visible face image can be directly utilized by existing face recognition systems which operate only on the visible face imagery. Therefore, using this approach one can leverage the existing commercial-off-the-shelf (COTS) and government-off-the-shelf (GOTS) solutions. In addition, the synthesized images can be used by human examiners for different purposes.

There are some informative traits, such as age, gender, ethnicity, race, and hair color, which are not distinctive enough for the sake of recognition, but still can act as complementary information to other primary information, such as face and fingerprint. These traits, which are known as soft biometrics, can improve recognition algorithms while they are much cheaper and faster to acquire. They can be directly used in a unimodal system for some applications. Usually, soft biometric traits have been utilized jointly with hard biometrics (face photo) for different tasks in the sense that they are considered to be available both during the training and testing phases. In our approaches we look at this problem in a different way. We consider the case when soft biometric information does not exist during the testing phase, and our method can predict them directly in a multi-tasking paradigm.

There are situations in which training data might come equipped with additional information that can be modeled as an auxiliary view of the data, and that unfortunately is not available during testing. This is the LUPI scenario. We introduce a novel framework based on deep learning techniques that leverages the auxiliary view to improve the performance of recognition system. We do so by introducing a formulation that is general, in the sense that can be used with any visual classifier.

Every use of auxiliary information has been validated extensively using publicly available benchmark datasets, and several new state-of-the-art accuracy performance values have been set. Examples of application domains include visual object recognition from RGB images and from depth data, handwritten digit recognition, and gesture recognition from video.

We also design a novel aggregation framework which optimizes the landmark locations directly using only one image without requiring any extra prior which leads to robust alignment given arbitrary face deformations. Three different approaches are employed to generate the manipulated faces and two of them perform the manipulation via the adversarial attacks to fool a face recognizer. This step can decouple from our framework and potentially used to enhance other landmark detectors. Aggregation of the manipulated faces in different branches of proposed method leads to robust landmark detection.

Finally we focus on the generative adversarial networks which is a very powerful tool in synthesizing a visible-like images from the non-visible images. The main goal of a generative model is to approximate the true data distribution which is not known. In general, the choice for modeling the density function is challenging. Explicit models have the advantage of explicitly calculating the probability densities. There are two well-known implicit approaches, namely the Generative Adversarial Network (GAN) and Variational AutoEncoder (VAE) which try to model the data distribution implicitly. The VAEs try to maximize the data likelihood lower bound, while a GAN performs a minimax game between two players during its optimization. GANs overlook the explicit data density characteristics which leads to undesirable quantitative evaluations and mode collapse. This causes the generator to create similar looking images with poor diversity of samples. In the last chapter of thesis, we focus to address this issue in GANs framework.

Share

COinS