Design of machine learning based face recognition system

Abstract

Face recognition is a robust and reliable system which maps out the contour of the person’s face digitally into the computer storing it into the data as a faceprint. These faceprints of many different people in a dataset will be trained through machine learning and it will learn how to recognize faces in that dataset. In this report, a face recognition system called OpenFace which is an open source code based on a paper called FaceNet will be studied and implement. This system uses the faceprint and compact it into a Euclidean space where the distances between embeddings of the faces in the 128-dimensional Euclidean space indicates their face similarity. This approach uses a novel online triplet mining method to train on triplets that optimizes the embeddings of the face images through a deep convolutional network. The accuracy of this method will be tested on the author’s custom dataset consisting of 35 classes of people with 10 images each. The result of the face recognition will be presented in chapter 3.