Feature Selection on Hyperspectral Data for Dismount Skin Analysis

Many security applications require the ability to accurately identify dismounts based on their distinctive identification properties. A dismount can be identified by many personal characteristics to include clothing, height, and gait. In particular, a dismount's skin can be used as an identifying feature because of the vast variability of skin pigmentation amongst individuals. Hyperspectral data, which is comprised of hundreds of spectral channels sampled from a nearly contiguous electromagnetic spectrum, is used to detect skin spectral variability amongst dismounts. However, hyperspectral data is often highly correlated and computationally expensive to process. Feature selection methods can be employed to reduce the data to a manageable size. This thesis presents the results of applying the fast correlation based filter (FCFB) [51] to a data set that contains hyperspectral data from the forearms of 62 subjects. The reduced data is used to train an artificial neural network (ANN) to discriminate a dismount of interest (DOI) amongst a group of 4 non-DOI's. The trained model is then tested to nd the same DOI amongst a group of 62 new non-DOI's. The FCBF selected four features (1014, 1024, 1033, and 1348nm) to discriminate amongst the dismounts. Using these four features, the ANN on average misclassified dismounts amongst four separate DOI validation tests. More specifically, the amount of possible DOI suspects was reduced from 62 to 4 dismounts. The FCBF outperformed three other feature selection methods with 4 times less misclassified instances. The original document contains color images.