2nd Portion of my ECE4553 Project. The scripts here won't work without any of the datasets described in the source. Trying to find a way to organize all 3 sets that are used in this program. The problem with pattern recognition and images is the large amount of features (pixels). With 4096 pixels, the feature space becomes less populated as the curse of dimentionality takes hold. Currently the best classification approach for the problem at hand is the use of Fisher Linear Discriminant analysis in order to signifigantly reduce the dimensionality of the dataset. Where the problem is a binary classification excercise, FDA reduces the data to a single feature. Classifiers trained based of the new feature, dubbed "Z", were found to give the greatest accuracy despite the 1 dimentional feature space.
Here's a link to the paper outlining our results: Open PDF.aodonnell/FDA-Classification-and-PCA-Analysis
Using Fisher LDA and Principal Components Analysis to evaluate the performance of a Generative Adversarial Network
Matlab