Soft biometric thermal face recognition using FWT and LDA feature extraction method with RBM DBN and FFNN classifier algorithms

This paper deals with error reduction, in thermal face recognition, by using multi biometrics to improve accuracy. To tackle this, images from the Terravic Facial Infrared Database, have been used with Fast Wavelet Transform (FWT) image compression approach, Linear Discriminant Analysis (LDA) technique, for feature extraction, and Restricted Boltzmann Machines (RBM), Deep Belief Network (DBN) and Feed Forward Neural Network (FFNN) for pre-training, testing and classification. To learn four different sets of training data, categorized using semantics such as: plain faces, faces in glasses, faces in a head gear and faces with facial hairs. These where used to classify the images, and a classification error of 0.0268, 0.030310, 0.02381 and 0.024629 was achieved by the algorithm on the plain faces, glass faces, head gear faces and facial hair faces respectively. By comparing the classification errors across the 4 algorithms, using test images not in the training set, soft-biometric recognition, such as plain face, face in glasses, hairy face and face in a head gear was possible for the thermal images.