This thesis describes two algorithms for face detection that rely on a generic feature
representation called spectral histogram representation. The sufficiency and generalization
of this representation is demonstrated through a statistical sampling technique. It is shown that the spectral histogram representation possesses desirable properties that commonly used representations are lacking. Presently there are a myriad of applications that critically rely on stringent parameters, alignment of templates or features of an image, and require large training sets. On the contrary the two algorithms proposed here based on the spectral histogram representation have relatively small training sets, do not require images to be aligned, and the spectral histogram representation is a non-parametric form so it does not rely on any stringent or underlying parameters. Both algorithms are tested on two widely used data sets and yield comparable results with the best existing algorithms.
