Homogeneous and Heterogeneous Face Recognition: Enhancing, Encoding and Matching for Practical Applications.

摘要

Face Recognition is the automatic processing of face images with the purpose to recognize individuals. Recognition task becomes especially challenging in surveillance applications, where images are acquired from a long range in the presence of difficult environments. Short Wave Infrared (SWIR) is an emerging imaging modality that is able to produce clear long range images in difficult environments or during night time. Despite the benefits of the SWIR technology, matching SWIR images against a gallery of visible images presents a challenge, since the photometric properties of the images in the two spectral bands are highly distinct.;In this dissertation, we describe a cross spectral matching method that encodes magnitude and phase of multi-spectral face images filtered with a bank of Gabor filters. The magnitude of filtered images is encoded with Simplified Weber Local Descriptor (SWLD) and Local Binary Pattern (LBP) operators. The phase is encoded with Generalized Local Binary Pattern (GLBP) operator. Encoded multi-spectral images are mapped into a histogram representation and cross matched by applying symmetric Kullback-Leibler distance. Performance of the developed algorithm is demonstrated on TINDERS database that contains long range SWIR and color images acquired at a distance of 2, 50, and 106 meters.;Apart from long acquisition range, other variations and distortions such as pose variation, motion and out of focus blur, and uneven illumination may be observed in multispectral face images. Recognition performance of the face recognition matcher can be greatly affected by these distortions. It is important, therefore, to ensure that matching is performed on high quality images. Poor quality images have to be either enhanced or discarded. This dissertation addresses the problem of selecting good quality samples.;The last chapters of the dissertation suggest a number of modifications applied to the cross spectral matching algorithm for matching low resolution color images in near-real time. We show that the method that encodes the magnitude of Gabor filtered images with the SWLD operator guarantees high recognition rates. The modified method (Gabor-SWLD) is adopted in a camera network set up where cameras acquire several views of the same individual. The designed algorithm and software are fully automated and optimized to perform recognition in near-real time. We evaluate the recognition performance and the processing time of the method on a small dataset collected at WVU.