Facial recognition, eigenfaces and synthetic discriminant functions

[摘要] ENGLISH ABSTRACT: Inthisthesiswe examinesome aspectsof automatic face recognition, withspecific reference totheeigenface technique.We providea thorough theoretical analysisof thistechniquewhichallows ustoexplainmanyof theresultsreportedin theliterature. Italso suggeststhatclusteringcan improve the performance of the system and we provide experimental evidence of this. From the analysis, we also derive an efficient algorithmfor updating the eigenfaces. We demonstrate the abilityof an eigenface-based systemto represent faces efficiently (using at most forty values in our experiments) and also demonstrate our updating algorithm. Since we are concerned with aspects of face recognition, one of the important practicalproblems is locating the face in a image, subject to distortions such as rotation.We review two well-known methods for locating faces based on theeigenface technique.eThesealgorithmsarecomputationally expensive, so we illustrate how theSyntheticDiscriminantFunctioncan be used to reduce thecost. For our purposes,we propose theconceptof a linearly interpolating SDF and we show how this can be used not only to locate the face, butalso to estimatethe extent of the distortion.We derive conditions which will ensure a SDF is linearly interpolating.We show how many of the more popular SDF-typefilters are related to the classic SDF and thus extend our analysis to a wide range of SDF-typefilters. Our analysis suggests that by carefully choosing the trainingset to satisfy our condition,we can significantlyreduce the size of the trainingset required. This is demonstrated by using the equidistributing principleto design a suitabletrainingset for the SDF.All this is illustrated with several examples. Our results with the SDF allow us to constructa two-stage algorithmfor locating faces.We use the SDF-typefilters to obtaininitial estimatesof the locationand extent of the distortion.This informationis thenused by one of the more accurateeigenface-based techniquesto obtainthe final location from a reducedsearchspace. Thissignificantlyreduces thecomputational cost of the process.

[发布机构] Stellenbosch University