Patch-based Finger Vein Verification Using Convolutional Variational Autoencoder

Finger vein recognition, a recent advancement in biometric technology, offers fast, contactless, and accurate identification; however, one of the current methods utilizing convolutional autoencoder (CAE) is sensitive to small translation errors, potentially compromising the verification process. This paper investigates a novel approach to improve the accuracy and robustness of patch-based finger vein verification using a convolutional variational autoencoder (CVAE) model with an additional loss term aimed at bringing the encodings of translated patch pairs closer to each other in the latent space, thereby mitigating the impact of minor misalignments on the system's performance. Furthermore, impostor pair embeddings are instead being distanced from each other to ensure that the false match rates do not increase, resulting in a more secure and reliable verification system. A comprehensive evaluation of the proposed CVAE model is provided through a series of experiments, comparing its performance with the CAE approach and assessing its effectiveness in managing spatial translations. Based on the results, the proposed CVAE model demonstrates better tolerance to translation errors and achieves an equal error rate (EER) of 0.278% on UTFVP dataset, improving upon the 0.556% EER of the CAE model.