Celebrity Recognition in Visual Media with Incremental Learning in Convolutional Neural Networks and Vision Transformers
The modern era is characterized by the rapid increase in the volume of digital data (images, videos), which are available on the internet and in personal collections. This increase enables and also creates the need for the development of the field of artificial intelligence. More specifically, facial recognition techniques through deep learning have achieved particularly significant progress, reaching very high levels of accuracy.
In this thesis a bibliographic analysis of available methods for classifying faces and continuous learning methods can be found. In the first part of the experimental process, three different methods are used to classify a dataset of facial images, consisting of 8500 different classes, each with approximately 300 images. The resulting networks are reused for the classification of a dataset with 500 different identities, while the results obtained are compared with networks that have been pre-trained on a different dataset or have randomly initialized their weights
In the second part of the experimental procedure, the performances of five different continuous learning methods are examined in the classification of 500 different classes. For each method, a different number of initial classes is used, along with two different neural networks for extracting image features. Additionally, for methods that retain a number of samples from old classes, their performance is examined for different numbers of old samples. The methods are compared based on whether they retain old samples or not, as well as their overall performance.
The results derived from the experimental procedure lead to conclusions regarding the comparison of face classification methods in terms of their performance and computational cost. Furthermore, the influence of different initialization of free parameters on the performance and convergence of neural networks is investigated when the number of available samples is limited. For continuous learning methods, their general performance in different scenarios is examined, and the ideal scenario for training them in 500 classes is explored. Finally, the impact of using different methods on training time and computational cost is examined. Based on these conclusions, certain directions are proposed for research in these specific areas.