Facial recognition using deep neural networks.

Abstract

Innovation or the introduction of the new. Innovation has always been the motor of civilization by introducing new ideas to solve the countless problems that such a complex system encounters. It has spurred economic growth and brought the comfort we live in today. In my opinion, in Belgium, particularly in the south, we have often rested on our laurels and forgot to innovate. This observation fostered my choice to undertake this thesis in the context of the RAGI project. This innovative project, relying on academic research to produce a commercial product, is developing an "intelligent system for recognition, welcoming, and guidance". Innovation in this project comes partly from the use of a machine learning technique called automatic face recognition to be able to identify people and guide them in a building. The goal of this thesis is to study this concept to help the team working on RAGI take appropriate decisions. To achieve this goal, I search for and analyze state-of-the-algorithms in both face detection and face recognition. The research for algorithms goes through the analysis of recent benchmarks, two of which (i.e. WIDER FACE and MegaFace) are also used for evaluating those algorithms. Simultaneously, this work points out the difficulties to perform such a research and testing process. The results on these benchmarks allows to determine which algorithms perform better, that is to say SSH for detection and both Dlib-R and ArcFace for recognition. For detection, the influence of facial attributes such as pose, size or blur is explored. Finally, to have more relatable results with regards to the RAGI project, we designed a specifi c dataset on which the same algorithms are tested. Composed of 494 frames with 3561 annotated faces from 13 different identities, it allowed us to study other parameters while confi rming the results obtained on the publicly available datasets. All those tests are performed with algorithm efficiency in mind and computation time measurements show that the best techniques tend to work slower but that they can achieve practical execution times.