Face recognition on historical photographs

Sammanfattning: The Stockholm city museum contains a large collection of old photographs. Manually classifying and comparing the photos for each person is inefficient and consumes more time. The idea is to recognize if photographs belong to the same person and compare photos from the photos collection. This project investigates several Deep Neural Networks for face recognition and compares similar faces utilizing the relevant features extracted using the Convolution Neural Network. Training the models from scratch is inconvenient for the lesser dataset. Instead, the transfer learning approach has shown better results in the previous years. So, by utilizing the pre-trained networks, the results are more significant. Seven Different Deep learning architectures have been experimented with and evaluated under the same circumstances. The applied methods are evaluated, and the best accuracy is obtained from InceptionResnetV1. However, other networks have also shown interesting results, among which Alexnet and Squeezenet showed considerable performance. Nevertheless, the Siamese network is used to compare similar photos, which gave convincing results.  However, improvements can be made to improve the performance of the models; generating more datasets and increasing the photo quality will add better results.