KLT and DCT Gender Classification of Face Images ------------------------------------------- A project proposal from: Junjie Liu Chad Netzer Pavithra Srinivasan Numerous methods have been proposed to automatically classify human faces by their gender. Many rely on the decomposition of images into "eigenfaces", by the use of a KLT transform. Once in this reduced space, the methods use a data set of known images to train a statistical model. Classification is done by projecting a new image into the eigenspace, after which the image is fit according to the statistical model. We propose to compare the method of "eigenfaces" using a KLT, with methods that use the DCT for dimension reduction. We will compare the performance of these two transforms, using the same statistical methods, and determine the tradeoffs in terms of accuracy and speed. The statistical methods we propose to use are a subset of the methods in the Diaco, DiCarlo, Santos classification project, namely the discriminant methods which performed best. We will also attempt to build a simple neural network model, using Matlab's Neural Net toolbox, and compare it with the discriminant models. Finally, we wish to determine if the results from the earlier classification project, trained and tested on databases composed solely of Stanford medical students, can be extended to a more general set of test databases. We will use exisiting face databases available on the web, and extend them with our own images, to assess classification accuracy against a wider range of ages, ethnicities, and backgrounds.