Face Recognition Based on Embedded Systems

Abstract: Machine learning in general, and artificial neural networks in particular, have gained a lot of attention in recent years. Using deep neural networks for classification tasks, such as face recognition, has proven more and more successful over time. The performance increase is partly due to more complex network architectures, and partly due to the use of larger datasets. The increased complexity of networks has lead to an increase in parameters, which in turn results in slower training and inference, making it hard to deploy such models on limited hardware. The main objective of this master's thesis is to train a convolutional neural network for face recognition, and deploy it on an embedded system, with the aim of real-time performance. By using transfer-learning as a means to adjust a pre-trained model to fit new data, the time needed for the training phase is reduced. The resulting model achieves an accuracy of 91.66%, while distinguishing between 2,904 identities. The model is then compressed by a method referred to as pruning, reducing the amount of parameters in the fully connected layers by a factor of 20, greatly reducing the memory footprint while remaining within 1% of the original accuracy. Finally, by combining the resulting neural network model with a custom built framework and a live video stream, real-time face recognition is achieved on an embedded device.