Face Recognition for Access Control Using Deep Neural Networks on Mobile Devices

Abstract

The Telenor-NTNU AI Lab would like to use facial detection as access control for their Lab. A proposed way of doing this is to mount a mobile device, like a tablet, outside the lab and run a facial recognition application on it. The goal of this project was to develop and test an Android application that performs facial recognition using artificial intelligence methods, as a proof of concept for a possible solution. The application was developed using the Android API and computer vision and deep learning libraries. To perform facial detection the app used a HAAR Cascade detector and an SSD Multibox detector. The recognition was achieved by using a deep neural net trained to make numerical representations of faces, called embeddings. When represented this way, it is possible to tell identities apart by calculating Euclidean distance between their embeddings. After development the application was tested in regards of performance, security and accuracy. Results of the tests showed that the application was successful in separating identities but was lacking in security. The performance depended on the device that the application was run on, and which detector was chosen. Even though the application as it is can not safely be used for access control, it is good at recognizing faces and demonstrates how neural networks can be used for facial recognition on Android.