Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Natalia A Schmid
Matthew C Valenti
Cross-spectral face recognition remains a challenge in the area of biometrics. The problem arises from some real-world application scenarios such as surveillance at night time or in harsh environments, where traditional face recognition techniques are not suitable or limited due to usage of imagery obtained in the visible light spectrum. This motivates the study conducted in the dissertation which focuses on matching infrared facial images against visible light images. The study outspreads from aspects of face recognition such as preprocessing to feature extraction and to matching.;We address the problem of cross-spectral face recognition by proposing several new operators and algorithms based on advanced concepts such as composite operators, multi-level data fusion, image quality parity, and levels of measurement. To be specific, we experiment and fuse several popular individual operators to construct a higher-performed compound operator named GWLH which exhibits complementary advantages of involved individual operators. We also combine a Gaussian function with LBP, generalized LBP, WLD and/or HOG and modify them into multi-lobe operators with smoothed neighborhood to have a new type of operators named Composite Multi-Lobe Descriptors. We further design a novel operator termed Gabor Multi-Levels of Measurement based on the theory of levels of measurements, which benefits from taking into consideration the complementary edge and feature information at different levels of measurements.;The issue of image quality disparity is also studied in the dissertation due to its common occurrence in cross-spectral face recognition tasks. By bringing the quality of heterogeneous imagery closer to each other, we successfully achieve an improvement in the recognition performance. We further study the problem of cross-spectral recognition using partial face since it is also a common problem in practical usage. We begin with matching heterogeneous periocular regions and generalize the topic by considering all three facial regions defined in both a characteristic way and a mixture way.;In the experiments we employ datasets which include all the sub-bands within the infrared spectrum: near-infrared, short-wave infrared, mid-wave infrared, and long-wave infrared. Different standoff distances varying from short to intermediate and long are considered too. Our methods are compared with other popular or state-of-the-art methods and are proven to be advantageous.
Cao, Zhicheng, "Cross-Spectral Full and Partial Face Recognition: Preprocessing, Feature Extraction and Matching" (2016). Graduate Theses, Dissertations, and Problem Reports. 5311.