Visualizing Important Areas for Facial Verification A Deep Learning Evaluation with ArtificialNoise Injection

Abstract: Facial verification is a growing area of research, where heavily trained neural network models obtain superhuman performance on recognizing individuals. However, because of network models' complex nature, it is hard to know what areas of the image are used. This thesis introduces three visualization techniques, to better understand the importance of different facial areas when performing verification. A method for measuring the visualizations' hierarchical correctness was established, evaluating the model's ability to preserve and damage the verification performance, when treated as artificial pixel noise. This is done for two types of networks, which are different in their architecture and their input dimensionality (one taking full color and the other taking grayscale images).