Face Recognition Using Transfer Learning on UFI Dataset

Abstract

Face Recognition is one of the most popular ways of biometric verification that are being applied worldwide, because of the convenience it offers. This encompasses large scale applications like corporate attendance systems to smaller ones such as unlocking hand-held devices and other various such kinds of applications. With the evolution of deep learning, face recognition systems have become increasingly accurate. One of the major reasons deep learning has become so popular for these types of tasks is because it does not require hand-crafted features. However, a major disadvantage of creating a deep learning model to recognize faces from scratch is the primary requirement of a huge amount of data. To counter this particular problem, another extremely popular technique called transfer learning is used to make the training process faster and with less data. In the current work, a novel real-world benchmark dataset is taken and the benchmark accuracy on it is increased by a large margin. The model in the current approach uses the concept of Siamese networks where triplets are generated for training. The triplets consist of an anchor image, a positive image of the same person and a negative image of a different person. This approach is particularly useful in this case because the amount of available data is less, along with the problem of class imbalance. The performance of the various models are compared with the previous results obtained using various features and classifiers and against one another.