Objektivizační analýza vlivu degradačních faktorů na výkon klasifikace obličeje pro biometrickou identifikaci

Abstract

This thesis is focused on the objectification analysis of the influence of degradation factors, especially impulsive noise in the image, luminance transformation or blurring, on the performance of face classification for biometric identification. It is concerned with elucidating the impact on the performance of selected convolutional neural networks, namely GoogLeNet and ResNet-101, that are used for detecting and recognizing human faces from the static images. All testing is done in an interactive MATLAB programming interface. The beginning of the thesis explains the basis of biometrics, face extraction and general methods of localization of facial markers. Then it discusses the types of neural networks and their descriptions. The second part of the thesis deals with the creation of a dataset from the degraded data and the creation of combinations of network setup parameters that will be used in testing. The conclusion of the thesis focuses on evaluating the performance of the networks in terms of validation accuracy, time consumption and the difference between training and validation accuracy under different parameter combinations, from which the ResNet-101 network is considered to be more suitable. Results are tabulated and graphically represented.