An efficient finger vein indexing scheme based on unsupervised clustering

Finger vein recognition has emerged as the robust biometric modality because of their unique vein pattern that can be captured using near infrared spectrum. The large scale finger vein based biometric solutions demand the need of searching the probe finger vein sample against the large collection of gallery samples. In order to improve the reliability in searching for the suitable identity in the large-scale finger vein database, it is essential to introduce the finger vein indexing and retrieval scheme. In this work, we present a novel finger vein indexing and retrieval scheme based on unsupervised clustering. To this extent we investigated three different clustering schemes namely K-means, K-medoids and Self Organizing Maps (SOM) neural networks. In addition, we also present a new feature extraction scheme to extract both compact and discriminant features from the finger vein images that are more suitable to build the indexing space. Extensive experiments are carried out on a large-scale heterogeneous finger vein database comprised of 2850 unique identities constructed using seven different publicly available finger vein databases. The obtained results demonstrated the efficacy of the proposed scheme with a pre-selection rate of 7.58% (hit rate of 92.42%) with a penetration rate of 42.48%. Further, the multi-cluster search demonstrated the performance with pre-selection error rate of 0.98% (hit rate of 99.02%) with a penetration rate of 52.88%.