Copy the page URI to the clipboard
Muftah, Asmail; Almurshed, Osama; Bennasar, Mohamed; Price, Blaine; Laurence, Sarah and Pike, Graham
(2024).
DOI: https://doi.org/10.1016/j.procs.2024.09.672
Abstract
In addressing the challenges of facial recognition in in the wild, our study initially investigated computationally efficient approaches for facial recognition in uncontrolled environments rather than conventional, computationally intensive techniques such as generative adversarial networks (GANs) and 3D reconstruction. We leveraged the capabilities of an off-the-shelf deep learning model, namely VGGNet, for efficiency and practical deployment in real-world scenarios. Our methodology included a dual phase training approach, starting with comprehensive training of the entire model, followed by selective fine-tuning of specific layers. This process was conducted using the CelebA dataset, known for its diversity and relevance to facial recognition research. The study demonstrates that this approach not only maintains robust generalisation across diverse conditions but also significantly reduces computational demands. Despite a slight trade-off in accuracy compared to more traditional methods, the benefits of the increased efficiency and the potential for real-time application deployment, such as in surveillance systems requiring quick processing, present a compelling case for further investigation and development within the field of facial recognition technology.