FedFVIQA: Personalized Federated Learning for Two-Stage Finger Vein Image Quality Assessment

Xingli Liu, Jian Guo, Hengyu Mu, Chong Han, Lijuan Sun

Abstract: Finger vein recognition systems are widely used in several fields, and existing methods usually require many high-quality images for training to ensure accuracy. As a result, research on the finger vein image quality assessment (FVIQA) has received considerable attention. However, in reality, the available finger vein images are often distributed across multiple organizations or companies. Due to insufficient data and a shortage of quality labels, it is difficult for such organizations and companies to independently train accurate FVIQA models. At the same time, due to user privacy and ownership constraints, it is typically not practical to pool data from multiple organizations or companies for model training. To address this problem, this paper introduces federated learning into FVIQA for the first time and proposes a personalized federated learning method for two-stage FVIQA (FedFVIQA). In the first stage, each client labels the quality of unlabeled finger vein images based on their similarity distribution for personalized scoring. In the second stage, the clients collaborate with a server for the training of quality classification model, thereby producing optimal personalized models. Finally, this paper reports extensive experiments conducted on the SDUMLA-HMT, NJUPT-FVP, HKPU-FV datasets to verify the superiority of the proposed method.

External IDs:dblp:conf/icpr/LiuGMHS24