Bingjie Liu

Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Guodong Guo

Committee Co-Chair

Xin Li

Committee Member

Shuo Wang


Face is an important physical characteristic of human body, and is widely used in many crucial applications, such as video surveillance, criminal investigation, and security access system. Based on realistic demand, such as useful face images in dark environment and criminal profile, different modalities of face images appeared, e.g. three-dimensional (3D), near infrared (NIR), and thermal infrared (TIR) face images. Thus, researches with various face image modalities become a hot area. Most of them are set on knowing the modality of face images in advance, which contains a few limitations. In this thesis, we present approaches for face image modality recognition to extend the possibility of cross-modality researches as well as handle new modality-mixed face images. Furthermore, a large facial image database is assembled with five commonly used modalities such as 3D, NIR, TIR, sketch, and visible light spectrum (VIS). Based on the analysis of results, a feature descriptor based on convolutional neural network with linear kernel SVM did an optimal performance.;As we mentioned above, face images are widely used in crucial applications, and one of them is using the sketch of suspect's face, which based on the witness' description, to assist law enforcement. Since it is difficult to capture face photos of the suspect during a criminal activity, automatic retrieving photos based on the suspect's facial sketch is used for locating potential suspects. In this thesis, we perform photo-sketch matching by synthesizing the corresponding pseudo sketch from a given photo. There are three methods applied in this thesis, which are respectively based on style transfer, DualGAN, and cycle-consistent adversarial networks. Among the results of these methods, style transfer based method did a poor performance in photo-sketch matching, since it is an unsupervised one which is not purposeful in photo to sketch synthesis problem while the others need to train pointed models in synthesis stage.