Last Friday, 13 September 2024 Szymon Żmijewski successfully defended his MSc thesis titled “Uncontrolled IdentityPreserving Synthetic Hand Images for Hand Recognition“.
Abstract:
The thesis explores the generation of identity-preserving synthetic hand images using state-of-the-art generative models. Traditional biometric systems rely on vast datasets of real-world images, raising significant privacy concerns and creating a demand for artificial alternatives. This work applies latent diffusion modelling (LDM) and StyleGAN3 to generate realistic hand images while preserving the unique biometric characteristics essential for identity recognition. A novel conditional identity preservation mechanism that integrates angular margin loss and text prompt encoding with diffusion models is proposed. A hand recognition system against real-world data evaluates the utility of the synthetic images generated. The results demonstrate competitive performances in both verification and open- and closed-set identification tasks, validating the utility of synthetic data for biometric applications. This approach addresses privacy concerns while providing an alternative source of training data for biometric models, advancing the field of hand recognition with identity-preserving synthetic data.