Date of Award
Campus Access Thesis
Master of Science (MS)
Rejection of unknown faces is a convoluted problem in face recognition. A major problem that models face while trying to assign an identity to a face is that the model only classifies the face into one of the identities that it has been trained on. This gets harder if the dataset that the model was trained on is sparse. With fewer images, the model not only needs to learn to identify the person but also be confident in rejecting faces that are not a part of the sparse training data. For instance, a photo of an unknown person could get assigned to one of the identities that most closely resembles the input image. This is a problem that may arise while building a security system for, let us say, a building which requires access to be granted to only the employees who work there. If the system is unable to confidently deny access to the unknown person, it would be a great security risk. However, at the same time, it must also allow the authorized person entry without misidentifying the person as unknown. In this thesis, I will dive into multiple ways this problem can be circumvented and the accuracy of these models be improved in rejecting the unknown faces. One of the conclusions of this thesis is that the choice of method largely depends on the available computing power. Not all models run equally efficiently and therefore it is important to decide the method one would use based on the device the system must run on.
Pandita, Aseem, "Deep Face Recognition and Rejection of Unknown Faces" (2021). Graduate Masters Theses. 685.