Chigozie Orji , Evan Hurwitz and Ali Hasan, University of Johannesburg, South Africa

This paper deals with the problem of errors in a biometric system that may arise from poor lighting and spoofing. To tackle this, images from the Terravic Facial Infrared Database have been used with Fast Wavelet Transform (FWT), an ensemble of classifiers and feature extractors, to reduce errors encountered in thermal facial recognition. By dividing the image set into a training set, comprising 1000 thermal images of 10 persons wearing glasses (X) and a test set comprising 100 image samples (y), of the same persons in glasses. A mean percentage error of 0.84% was achieved, when a Convolutional Neural Network (CNN) was used to classify the image set (y), after training with (X). However, when the images where pre-processed with Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and k-Nearest Neighbors (KNN) classifier, a mean percentage error of 0.68% was achieved with the CNN classifier.

Thermal imaging, ensemble of classifiers, Deep Convolutional Neural Networks, K-Nearest Neighbors, Eigen Vectors, Principal Component Analysis, Linear Discriminant Analysis, biometrics, sensing, imaging, security