Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Katerina Goseva Popstojanova
Jacqueline A Speir
This dissertation investigates the advantages and limitations of the heterogeneous problem of matching infrared long-range, night time face images against their visible counterparts. The contributions of the thesis are three-fold. First, a multifeature scenario dependent fusion scheme is developed, where Gabor Wavelets, Histogram of gradients (HOG) and Local binary patterns (LBP) feature descriptors are empirically selected. Next, a set of fusion score level schemes are developed and applied before face matching. The developed fusion scheme results in 54 percent point (pp) higher recognition performance than the baseline established using a commercial and a set of academic face matchers. Second, to further improve the performance of baseline face recognition (FR) systems, a scenario dependent and sensor adaptable convolutional neural network (CNN) is developed that groups the data in terms of demographic information, including scenarios (situations) such as indoors or outdoors data, as well as distance and sensor based data. The automated grouping scheme developed is applied before the FR algorithms are used, improving baseline performance from 48% (all vs. all) to 70% (with data grouping). Third, an image quality restoration scheme is designed and developed. This scheme is beneficial to FR systems because the quality of face data captured under challenging conditions is affected by a variety of noise factors (including low illumination conditions, variable standoff distances, and uncooperative subjects). Thus, image quality is responsible for the performance degradation of conventional FR matchers. The developed scheme improves, first, the quality of distorted face images and, then, FR performance in terms of the rank-1 identification rate. Based on the experimental results the major conclusion from this research is that the aforementioned schemes discussed in this dissertation significantly improve cross-spectral face matching performance on diverse scenarios, when used either independently or in combination. The experimental results are further supported by statistical analysis tests, conducted to find the statistical significance of incorporating the developed score level fusion schemes, as well as the demographic filtering to FR systems.
Narang, Neeru, "Deep Features based Hierarchical Classification Scheme for Face Recognition in Heterogeneous Environments" (2017). Graduate Theses, Dissertations, and Problem Reports. 6294.