Study of classification techniques for face detection in security practice

Abstract

Convolutional neural networks (CNN) are the foundation of current deep learning frameworks. The aim of this thesis is to present a comparative analysis of the performance of different CNN architectures in face recognition and classification tasks. The theoretical part deals with the description of the biometric features used, performance measurements of biometric recognition systems. The second part of the theoretical part deals with the application of CNNs to classification tasks with a focus on face recognition. The experimental part of this work focusses on comparing the performance and computational efficiency of different CNN architectures. A total of 18 CNNs have been included in the experimental part, and 5 (AlexNet, GoogLeNet, DarkNet-19, SqueezeNet, ShuffleNet) have been investigated in detail. The dataset Labeled Faces in the Wild with more than 13,000 face images was used for testing. The main experiment (B) was set with a learning rate of 0.001, a cross-entropy loss function, and 30 epochs (2400 iterations). DarkNet-19 achieved the best results (final validation accuracy).