Högskolan Dalarnas logga och länk till högskolans webbplats

Background: Today, many offices are implementing smart solutions, including cloud services and the Internet of Things. One piece of technology many have found interesting is the use of facial recognition. Deciding which approach to choose to tackle this is unclear. Aim: The aim is to find what implementation and what configuration of that implementation yields the highest accuracy and speed. Method: Two sets of training data were generated. The first set were images of the subjects taken in a non-normalised manner and the second set were images taken in a normalised manner. After training the models, we had the subjects walk past the camera one by one and compared the number of true positive predictions compared to the total amount of frames the subject appeared in, as well as compared the delay between the frames. Results: On non-normalised training data, both Eigenface and Fisherface saw a drop in accuracy when the number of training images increased, LBPH instead saw an increase in accuracy with the same training. No humanly perceptible difference in delay was seen in any model. On normalised data, Eigenface again saw a drop in accuracy, but was now instead joined by LBPH. Fisherface initially performed well in one trained image, performed worse in two training images, and then reached its highest performance on five training images. Conclusions: When implementing face recognition in a limited environment, using models based on LBPH is preferable and will yield the highest prediction accuracy. Implications: The presented research is to be seen as a roadmap or subject to research further on when companies seek to implement face recognition.