Reconhecimento de íris fake utilizando descritores de momentos de Zernike e classificadores de formas

Abstract

Biometric systems have attracted research interest due to the increasing improvements in fraud prevention for access authentication to these systems. As a suggested solution to the security holes in accessing these systems, this dissertation developed a synthetic iris detection system using Zernike moments as a descriptor for non-aligned or circular shapes. The contribution of this work is to investigate the performance of Zernike moments when combined with artificial neural network classifiers for detecting synthetic iris images. In the experiments, databases were created from combinations of other databases available in studies on the referenced literature. Zernike moments are applied locally as shape descriptors, providing good results when using more complex neural network classifiers. The performance evaluation of the proposed system was performed for Zernike moments combined with three different types of classifiers: Multilayer Perceptron, Convolutional Neural Network and Residual neural network. For the combination of Zernike with MLP, the accuracy was 79%, while for the combination of Zernike with CNN and Zernike with ResNet in the classification, the accuracy values were 90% and 93%, respectively.