The Intelligent Mirror - A Personalized Smart Mirror Using Face Recognition

Abstract

Smart mirrors are a new addition to the smart product family that has been getting a lot of attention in recent years by both commercial manufacturers and hobbyists. This thesis explores enhancing such mirrors with intelligence. The goal is to develop a user recognition system that is able to differentiate between the users of the system by using a camera hidden behind a two-way mirror.

The recent resurgence of deep neural networks have pushed the fields of computer vision and face recognition to create powerful neural networks that by far outperforms the methods that was before. The Raspberry Pi(RPI) will be used as the embedded computational device for the intelligent mirror. The question is if it is able to utilize the power of the new face recognition methods or if it is too slow. This thesis will compare a various of face detection and face recognition methods to explore if it is possible to build a satisfactory user recognition system on the RPI. Therefore we propose two different architectures for the intelligent mirror. Either a local architecture that only uses the RPI, or a remote architecture where the RPI simply forwards the camera stream to a remote device that computes the results.

The results obtained makes it clear that a user recognition system employing deep learning is today way too computationally complex for the RPI to be a suitable device. Using older methods resulted in a barely acceptable user recognition system that can be used, but will often predict wrong user.

When keeping in mind that user recognition is just the first step in developing the intelligent mirror the conclusion is that the system should either use a remote architecture or look for more powerful embedded devices with hardware capable of feed-forwarding deep neural networks.