| With the advent of the internet and increasing infiltration of technology into the lives of billions worldwide, the need for automatic recognition and security systems is growing, and will continue to grow, far into the 21 st century. A large number of critical applications on the internet such as online banking, loan approval, credit card account management, stock market trading, and “e-commerce” require userids and passwords that can be forgotten, lost, or stolen. Not only is it difficult to remember so many passwords, but the passwords themselves are insecure. An alternative is biometrics; the measurement of unique personal characteristics, such as face recognition. Although much work has been done in this area, much work remains. One problem is the limited computing ability of digital computers. Although many of the functions of face recognition such as preprocessing, selective attention, and feature extraction are well suited for parallel processing, they are typically performed in serial on a digital processor. In most cases, the high degree of accuracy provided by the digital processor is not required and these functions can be performed more efficiently in terms of silicon area, speed, and power in analog very large-scale integrated (VLSI) hardware. In this dissertation, a number of circuits and systems loosely inspired by neurophysiology are presented for performing various “early vision” functions in standard complementary metal-oxide semiconductor (CMOS) technology, thereby offloading some of the processing onto dedicated hardware and freeing up the processor for more complex tasks. In order to keep the circuits as simple as possible, three-dimensional information about the face is neither required nor attempted to be recovered. The sub-systems presented include: diffusion, contrast enhancement, eye-tracking, face detection, morphological filtering, Gabor wavelet feature extraction, principal-component analysis, and translation-invariant feature extraction. New CMOS implementations of well-known algorithms are presented along with several new spatiotemporal paradigms such as dynamic linking and eigenwave networks for face detection. |
