A Lightweight CNN Architecture for Face Recognition on Embedded Devices

In our daily life, faces are one of the most familiar and common biometric features that allow to identify a person through the use of artificial intelligence techniques. For this reason, in recent years the theme of Facial Recognition has had a huge impact on several fields such as forensic investigations but also such as access management to devices or entertainment applications. Although biometric facial information is not ideal compared to other biometric traits, due to its practicality and immediacy they are experiencing a very strong technological development. Current face recognition applications are based on deep neural networks that allow greater precision and adaptability to different uses, but require, at the same time, great resources for training and inference making their use inaccessible in those products that do not have large computing power. To compete in the video intercoms market by providing additional services to the final consumer, CAME S.p.a. commissioned the design and development of a lightweight neural network for facial recognition well suited for the embedded systems of their video intercoms. These embedded systems impose severe constraints on the available computational capabilities, further limited by privacy laws that forbid the use of delocalized infrastructures. In this thesis, we addressed the Facial Recognition problem by proposing a lightweight pipeline that will be executed online in CAME video intercoms. In particular, we will exploit a Multitask Cascaded Convolutional Network (MTCNN) for the Face Detection step to detect faces in images and to align the faces, and a Siamese Network based on the ResNet-18 architecture for the Face Identification step whose training is based on a custom-built dataset. The experiments showed reasonable results despite the lightness of the architecture.