Fed-UIQA: Federated Learning for Unsupervised Finger Vein Image Quality Assessment

Xingli Liu, Jian Guo, Hengyu Mu, Lejun Gong, Chong Han

Abstract: Finger vein recognition, an emerging biometric technique, has been widely applied across various domains. Image quality significantly impacts vein recognition system performance, prompting considerable attention to finger vein image quality assessment. However, most existing methods for assessing finger vein image quality are limited to evaluating image quality levels and lack robustness and practicality. Furthermore, existing finger vein image datasets used for model training not only lack quality labels but also exhibit significant heterogeneity. Moreover, privacy concerns often constrain their creation and usage, further complicating image quality assessment. To address these challenges, this paper introduces a federated learning for unsupervised finger vein image quality assessment (Fed-UIQA). This method computes image spatial distances and relative classifiability to obtain image quality scores, enabling unsupervised quality assessment. Local incremental models are designed on the client-side to address heterogeneous vein datasets to enhance the global model aggregated on the server, thereby improving its adaptability to local data. Finally, extensive experiments conducted on five datasets validate the superiority of the proposed approach.

External IDs:dblp:conf/icic/LiuGMGH24